Unlocking de novo antibody design with generative artificial intelligence

Through the combination of generative deep learning model and high-throughput experiments, high-affinity and diversified antibody variants are designed, which solves the problems of low design efficiency and high cost in existing antibody design technologies, and achieves rapid and economical antibody development .

CN119096300APending Publication Date: 2024-12-06ABSCI CORPORATION
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202380029360.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2023-01-07
Filing Date
2023-02-09
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

Existing antibody design techniques lack efficient computer simulation methods, and cannot quickly and accurately design targeted antibodies against designated binding partners. Traditional laboratory methods are time-consuming and expensive, making it difficult to control the output sequence, resulting in binding of pilot candidates. Inadequate capability and developability.

Method used

Using generative deep learning models, we can design antibodies in a zero-sample way by training machine learning models, and use high-throughput wet laboratory screening and surface plasmon resonance technology to generate high-affinity antibody variants for specific antigens to avoid cumbersome libraries. Screening process optimizes the controllability and exploitability of antibodies.

Benefits of technology

Achieve rapid and economical design of high-affinity, diversified antibody variants, significantly reducing the time and resource requirements for therapeutic antibody development, and providing efficient de novo antibody design schemes that can be demonstrated in experimental verification comparable to traditional methods. Even better than the results of traditional methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119096300A_ABST
    Figure CN119096300A_ABST
Patent Text Reader

Abstract

A computing system for generating structural information of a biomolecule includes processors and one or more non-transitory computer-readable media having instructions stored thereon that, when executed by the one or more processors, cause the computing system to: receive a training input; processing the training input with a machine learning biomolecule prediction model; evaluating a loss function; and modifying parameters of the machine learning model. A computing system for generating structural information of a target biomolecule includes a processor and one or more non-transitory computer readable media having stored thereon a machine learning biomolecule prediction model and instructions, the instructions, when executed by the one or more processors, cause the computing system to receive a target input and predict the structural information of the target biomolecule.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] Incorporation by Reference of Electronically Submitted Materials

[0002] The sequence listing as part of the present disclosure is submitted in a computer-readable format at the same time as the specification. The file containing the sequence listing is named "57581_Seqlisting.xml", which was created on February 9, 2023 and is 489,494 bytes in size. The subject matter of the sequence listing is incorporated herein by reference in its entirety.

[0003] Cross-reference to related applications

[0004] This application claims priority to U.S. Provisional Patent Application No. 63 / 478,933, filed on January 7, 2023, entitled UNLOCKING DE NOVO ANTIBODY DESIGN WITH GENERATIVE ARTIFICIAL INTELLIGENCE, and U.S. Provisional Patent Application No. 63 / 308,495, filed on February 9, 2022, entitled structure-based design of binding partner-targeted biomolecules. The contents of each provisional patent application are incorporated herein by reference in their entirety. Background Art

[0005] Antibodies are multiple proteins that the immune system uses to naturally bind and neutralize foreign bodies (such as viruses, fungi and bacteria). Therefore, antibodies are likely to be drug candidates for many infectious diseases and cancers. For example, there is a lack of conventional techniques for automatic design and / or improvement of computer simulations for antibodies, because such protein design techniques based on deep learning unconditionally design proteins, only use open source data, and cannot be targeted for specified binding partners. Such techniques have not yet demonstrated the ability to adjust the design according to other criteria / purposes. Conventional laboratory-based techniques are time-consuming and expensive. For example, traditional de novo antibody discovery requires time and resource-intensive screening of large immune or synthetic libraries. These methods also have little control over output sequences, which may cause lead candidates to have suboptimal binding and poor developability attributes. Conventional generative antibody design techniques have not yet demonstrated de novo antibody design by experimental verification. Therefore, improved protein design techniques are needed. Summary of the invention

[0006] In one aspect, a computing system for training a machine learning model to generate structural information of a target biomolecule comprises one or more processors; and one or more non-transitory computer-readable media having instructions stored thereon, wherein the instructions, when executed by the one or more processors, cause the computing system to: (1) receive one or more training inputs, comprising one or more of: (i) input biomolecule structural information, (ii) input biomolecule binding partner structural information, or (iii) input biomolecule-input binding partner binding complex structural information; (2) process the one or more training inputs using a machine learning-based biomolecule prediction model to generate predicted biomolecule structural information; (3) evaluate a loss function that compares the predicted biomolecule structural information to a baseline true value; and (4) modify one or more values ​​of one or more parameters of the machine learning model based at least in part on the loss function.

[0007] On the other hand, a computing system for generating structural information of a target biological molecule includes one or more processors; and one or more non-transitory computer-readable media having stored thereon: a machine learning-based biomolecule prediction model trained to predict structural information of a biomolecule based on an input; and instructions that, when executed by the one or more processors, cause the computing system to: (1) receive a target input comprising one or more of the following: a target binding partner primary sequence, a three-dimensional coordinate of a target binding partner, a target binding partner epitope primary sequence, or a three-dimensional coordinate of a target binding partner epitope primary sequence, or a fragment or portion of any of the foregoing; and (2) predict the structural information of the target biological molecule by processing the target input with the machine learning-based biomolecule prediction model.

[0008] In yet another aspect, a computing system for predicting affinity of a target biomolecule comprises one or more processors; and one or more non-transitory computer-readable media having stored thereon: a machine learning-based affinity prediction artificial neural network comprising: (i) one or more biomolecule prediction layers trained to predict biomolecule structural information from a target input; (ii) one or more docking layers trained to generate docking complexes from two or more input three-dimensional biomolecules; and (iii) one or more affinity prediction layers trained to predict affinity of the input docking complex; wherein the one or more biomolecule prediction layers, the one or more docking layers, and the one or more affinity prediction layers are connected; and instructions, when executed by the one or more processors, cause the computing system to: (1) receive a target input comprising one or more of a target binding partner sequence, a target binding partner, or a target epitope; and (2) process the target input using the affinity prediction artificial neural network to generate a docking complex and a corresponding structural affinity value corresponding to the target input. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1A Depicted is zero-shot generative AI for de novo antibody design, in which deep learning models trained on antibody-antigen interactions are combined with ultra-high-throughput wet-lab experiments to enable the design of binders to never-before-seen antigens without the need for further affinity maturation or lead optimization.

[0010] Figure 1B An exemplary computing environment for performing the present techniques according to some aspects is depicted.

[0011] Figure 2A Depicted are exemplary computer-implemented methods for training a machine learning model to generate structural information of biomolecules according to some aspects.

[0012] Figure 2B Depicted is a data flow block diagram for training and operating one or more machine learning models according to some aspects.

[0013] Figure 2C Described are computer-implemented methods for generating structural information of a target biomolecule according to some aspects.

[0014] Figure 3A Depicted is a block flow diagram of an exemplary computer-implemented method for performing in silico de novo biomolecule discovery, according to some aspects.

[0015] Figure 3B Depicted is an exemplary block diagram according to some aspects.

[0016] Figure 3C Depicted is a block diagram of an exemplary CDR3 discovery input and output for a computer simulation according to some aspects, wherein block 310b includes SEQ ID NO:1, and wherein block 310c includes SEQ ID NO:2.

[0017] Figure 3D Depicted is an exemplary block diagram of machine learning based drug candidate design according to some aspects.

[0018] Figure 3E Depicted is an exemplary flow chart of antigen encoding according to some aspects.

[0019] Figure 3F An exemplary flow diagram for antibody generation training according to some aspects is depicted, wherein block 322a includes SEQ ID NO:3 and SEQ ID NO:4; and wherein block 322c includes SEQ ID NO:5.

[0020] Figure 3G An exemplary flow chart for antibody generation training according to some aspects is depicted, including SEQ ID NO:6, SEQ ID NO:7, and SEQ ID NO:8.

[0021] Figure 3H An exemplary flow chart for antibody generation training according to some aspects is depicted, including SEQ ID NO:8 and SEQ ID NO:9.

[0022] Fig. 3I An exemplary flow chart for antibody generation training according to some aspects is depicted, including SEQ ID NO:3, SEQ ID NO:4, and SEQ ID NO:5.

[0023] Figure 3J An exemplary flow chart for antibody generation training according to some aspects is depicted, including SEQ ID NO:3 and SEQ ID NO:4.

[0024] Figure 3K An exemplary flow chart for antibody generation training according to some aspects is depicted, including SEQ ID NO:3, SEQ ID NO:4, and SEQ ID NO:5.

[0025] Figure 3L Depicted is an exemplary block diagram of supervised machine learning model output according to some aspects.

[0026] Figure 4 Depicted is an exemplary graphical user interface that includes visualizing ground truth structural information and predicted structural information according to some aspects.

[0027] Figure 5A Depicted is an exemplary flow chart of fully in silico biomolecule design according to some aspects.

[0028] Figure 5B Depicted is an exemplary block diagram of a fully in silico de novo biomolecular design artificial intelligence model according to some aspects.

[0029] Figure 5C Depicted is an exemplary block diagram of a computer simulated biomolecule design pipeline according to some aspects.

[0030] Fig. 6A Depicted are exemplary computer simulated biomolecular engineering according to some aspects, including binder prediction based on optimization of CDRH3 while keeping the remaining CDRs unchanged, including SEQ ID NOs: 553 to 555.

[0031] Figure 6BDepicted are exemplary ranibizumab and trastuzumab bH1 visualizations according to some aspects.

[0032] Figure 6C Depicted are exemplary sensorgrams of surface plasmon response results corresponding to predicted computer simulated binding for trastuzumab-bH1 CDR3 variants binding to VEGF according to some aspects.

[0033] Fig.6D Described are some aspects corresponding to Figure 6C Exemplary tabular data showing the results of predicted in silico binding of trastuzumab-bH1 CDR3 variants to VEGF, including SEQ ID NOs: 10 to 22.

[0034] Fig. 6E Depicted are exemplary sensorgrams of surface plasmon response results corresponding to predicted computer simulated binding of trastuzumab CDR3 variants to Her2 according to some aspects.

[0035] Fig. 6F Described are some aspects corresponding to Fig. 6E Exemplary tabular data showing the surface plasmon response results generated by predicted computer simulation binding of trastuzumab CDR3 variants to Her2, including SEQ ID NOs: 23 to 38.

[0036] Fig. 7A Depicted is an exemplary block diagram of initial laboratory validation of modeling predictions according to some aspects.

[0037] Figure 7B Depicted is an exemplary block diagram of computer simulation generation of a de novo designed trastuzumab CDR3 having comparable or improved binding affinity to VEGF and without multispecificity according to some aspects.

[0038] Figure 7C Depicted is an exemplary block diagram of in silico generation of a de novo trastuzumab CDR3 that binds HER2 with high specificity according to some aspects.

[0039] Fig.7D Depicted are exemplary block diagrams of negative controls having scrambled CDR3 sequences that do not bind antigen, including SEQ ID NO:39 and SEQ ID NO:40, according to some aspects.

[0040] Fig. 7EDepicted are exemplary block diagrams of generated CDR3 sequences exhibiting high diversity according to some aspects, including SEQ ID NO:20.

[0041] Figure 7F Depicted are exemplary tree diagrams showing that, according to some aspects, in silico generation of de novo design variants is highly diverse.

[0042] Fig. 8A The landmark map of the HCDR3 of 421 trastuzumab binding variants is depicted, where the greater diversity observed in the center of the CDR (hypervariable domain) corresponds to the D germline genes, while the tails correspond to the J and V germline genes, which tend to be less diverse.

[0043] Figure 8B Binding affinities of AI-generated zero-shot binders according to some aspects are depicted, including 71 designs with comparable affinity to trastuzumab (<10 nM) and three designs with tighter binding.

[0044] Figure 8C The edit distances of the designed variant binding affinities relative to trastuzumab are depicted, where the edit distances range from 2 mutations (84.6% sequence identity) to 12 mutations (7.7% sequence identity), illustrating the novelty of the design and that while binding affinity tends to decrease with increasing edit distance, the inventive technique can still identify binders in each group with comparable (within 0.5 log) or higher binding affinity.

[0045] Fig.8D Pairwise edit distances between binders are depicted (eg, maximum is 15 mutations, median is 8, mean is 7.7, standard deviation is 2.1).

[0046] Fig. 8E Depicted are pairwise edit distance distributions (on a logarithmic scale) of pairwise HCDR3 edit distances of HER2 binders designed by an AI according to some aspects.

[0047] Fig.8F The length distribution of HCDR3 lengths of designed HER2 binders is depicted.

[0048] Fig. 9A The minimum edit distance of binders to the training data HCDR3 is depicted, where across the design, binders were observed to differ by at most eight mutations from any HCDR3 in the training data.

[0049] Fig. 9B Depicted are the binder minimum edit distances to any HCDR3 in a SAbDab for an AI-specified HER2 binder according to some aspects.

[0050] Fig. 9C Depicted are the binder minimum edit distances for OAS HCDR3 according to some aspects.

[0051] Fig.9D The natural metrics of binders designed according to the invention according to some aspects are depicted relative to their diversity towards trastuzumab.

[0052] Fig.9E Depicted are the naturalness of designed binders according to some aspects relative to a control.

[0053] Fig.9F Depicted are the distributions of the minimum sum of the edit distances of the AI-designed HER2 binders to the HCDR3 in OAS and the edit distances of the trastuzumab HCDR1 and HCDR2 to the OAS HCDR1 and HCDR2 sequences.

[0054] Figure 9G Depicted are the distributions of naturalness scores of de novo designed HER2 binders with minimal OAS ED relative to the minimum edit distance to HCDR3 in OAS.

[0055] Figure 9H The naturalness score is plotted versus the minimum edit distance to any HCDR123 in OAS. Fig.9I The distribution of minimum edit distances to HCDR3 in OAS for HER2 binders designed for multi-step multi-HCDR AI is depicted.

[0056] Fig.9I Depicted is the minimum edit distance to any HCDR3 in the OAS according to some aspects.

[0057] Figure 9J The distribution of minimum edit distances of HER2 binders to HCDR1, HCDR2 and HCDR3 in OAS for multi-step multi-HCDR AI designs is depicted.

[0058] Fig. 10A The conformational flexibility of de novo designed HCDR3 Her2 binders comprising SEQ ID NOs: 41 to 48 is depicted.

[0059] Fig. 10B Depiction of the HCDR3 loop and HCDR3 atoms Space-filling representation of the interactions of epitope residues within. Trastuzumab HCDR3 loops are red, and de novo Her2 binder HCDR3 loops are blue; two different epitope pockets interact differentially with residues of each HCDR3; and the interaction surface of the epitope is based on HCDR3 sequence and conformational changes.

[0060] Fig. 10C Depicted is a comparison of the trastuzumab-Her2 structure and the de novo designed HCDR3-Her2 complex, showing superposition of the trastuzumab-Her2 structure and the de novo designed HCDR3-Her2 complex, showing conformational differences in the HCDR3 backbone; the main chain backbone is depicted as ribbons and sterically conserved side chains are shown as sticks.

[0061] Fig. 10D Depiction of the HCDR3 loop and HCDR3 atoms Stick representation of the epitope residues interacting within . Trastuzumab HCDR3 loops are red and de novo Her2 binder HCDR3 loops are blue. Epitope residues are labeled according to the crystal structure 1N8Z. * indicates novel epitope residues in the de novo HCDR3 complex that are not observed in the trastuzumab-Her2 complex.

[0062] Fig.11A Depicted is an alignment of designed VEGF binders with ranibizumab sequences, including SEQ ID NO:49 and SEQ ID NO:50.

[0063] Fig. 11B Depicted is an alignment of SARS-CoV-2 Omicron binders with casirivimab. The designed HCDR3 is diverse and novel, as it differs from the HCDR3 of casirivimab by 6 mutations, from any HCDR3 in OAS by at least 2 mutations, and from any HCDR3 in CoV-AbDab (a database of antibodies capable of binding to coronaviruses) by at least 4 mutations, including SEQ ID NO:51 and SEQ ID NO:52.

[0064] Fig. 12A Depicted are designed VEGF-binding agent structures according to some aspects.

[0065] Fig. 12B Depicted are wild-type binder structures according to some aspects.

[0066] Fig.13A Depicted are sensorgrams according to some aspects, including SEQ ID NOs: 41 to 48.

[0067] Fig. 13B Sensorgrams according to some aspects are depicted.

[0068] Fig. 13C Sensorgrams according to some aspects are depicted.

[0069] Fig.13D Sensorgrams according to some aspects are depicted. DETAILED DESCRIPTION

[0070] The present disclosure addresses the need for artificial intelligence (AI) models and machine learning (ML) models that are trained to predict biomolecule sequences using known biomolecule / binding partner complexes. In particular, generative AI has the potential to greatly improve the speed, quality, and controllability of biomolecule (e.g., antibody) design. The present technology includes the use of a de novo generative deep learning model to design antibodies against three different targets from scratch in a zero-shot manner, where all designs are the result of a single round of model generation without subsequent optimization. In particular, the present technology can use high-throughput wet lab capabilities to screen a large number (e.g., 400,000 or more) of antibody variants designed to bind to human epidermal growth factor receptor 2 (human epidermal growth factor receptor; HER2).

[0071] From these screenings, the technology of the present invention was able to further biophysically characterize 421 binders using surface plasmon resonance (SPR), and found that three of them were more tightly bound than the therapeutic antibody trastuzumab. These binders are highly diverse and have low sequence consistency with known antibodies. In addition, these binders scored high in the previously introduced naturalness metrics (see Bachas S, Rakocevic G, Spencer D, Sastry AV, Haile R, Sutton JM et al. Antibody optimization enabled by artificial intelligence predictions of binding affinity and naturalness. bioRxiv. 2022; doi: 10.1101 / 2022.08.16.504181.), indicating that they may have an ideal developability profile and low immunogenicity. These results have opened up a path to accelerate drug development for novel therapeutic targets using generative AI combined with high-throughput experiments.

[0072] A particularly difficult aspect of antibody drug development is the initial step of lead candidate identification due to the labor-intensive and uncontrolled nature of traditional screening methods. AI-based generative de novo design has the potential to disrupt these shortcomings of the current drug discovery process. The zero-shot nature of AI design methods avoids the need for tedious library screening to identify binding molecules, saving significant time and costs. In addition, the controllable nature of model-based design allows the generation of proteins optimized for developability and immunogenicity characteristics, thereby mitigating downstream developability risks.

[0073] The technology of the present invention provides a step for completely de novo antibody design by showing the ability to generate novel antibody variants that confer binding and natural sequence characteristics comparable to parent antibodies and in some cases superior to parent antibodies in a zero-sample manner. The sequences generated by AI are different from any sequence observed in the model training set, and most are different from known sequences in the Observed Antibody Space (OAS) database, but maintain a high naturalness score, showing that the model can design antibody sequences along a biologically feasible manifold. In addition, the designed sequences are highly different from each other, indicating that different binding molecule solutions can be designed. In addition, by using a modified multi-step method to produce and verify a binding agent with up to 3 novel heavy chain CDRs, progress in de novo design of multiple CDRs is demonstrated. The technology of the present invention is generalizable, as demonstrated by the deployment of different antigens by the generative design method of the present invention. In addition, developing epitope specificity across multiple antigens for antibody design can allow precise interaction with biologically relevant target areas associated with disease mechanisms of action. In addition to advances in generative modeling, as the time and cost of DNA synthesis continue to decline, the speed and scale of wet laboratory verification of AI-generated designs will gradually increase.

[0074] Antibodies, in particular, are a growing class of therapeutic molecules because of their attractive drug-like properties, including high target selectivity and minimal immunogenic effects. Antibody drug development typically begins with initial lead molecule discovery. Existing methods for advance discovery typically consist of randomly searching large combinatorial sequence spaces by screening a large library of antibody variants against a target antigen. Techniques such as phage, yeast display, immunobinding hybridoma screening, or B-cell sequencing are typically used for initial discovery, followed by further molecular development. These methods are time and resource intensive, lack control over the properties of the designed antibodies, and often produce suboptimal leads. Applying generative artificial intelligence (AI) to design de novo antibodies in a zero-sample and controllable manner, rather than screening and developing lead molecules, can greatly reduce the time and resources required for therapeutic antibody development.

[0075] The application of AI methods to antibody design, and more generally to protein therapeutic design, is compelling given the availability of large protein sequence and structure databases that can facilitate model training. Indeed, recent work has shown that models trained on these data can be used for de novo design of certain classes of proteins. These works screened tens to thousands of protein designs, two to four orders of magnitude fewer than the proteins validated in our study. Furthermore, despite the enormous therapeutic relevance of antibody-based therapeutics (accounting for 30% of FDA-approved biologics in 2022), there are currently no methods that enable de novo design of antibodies with wet lab validation. De novo antibody design is of particular interest because key determinants of antibody function come from the complementarity determining regions (CDRs) of the sequence. These hypervariable regions interact directly with the antigen, and among these hypervariable regions, the heavy chain CDR3 (HCDR3) is often the most critical for binding, but also the most variable, making modeling particularly challenging. Several works with experimental validation have attempted to use supervised learning to optimize antibodies, but none have attempted zero-shot or de novo design.

[0076] Many groups have recognized the potential of zero-shot generative AI to impact antibody design. Several promising approaches have emerged recently that use ideas from language modeling to geometric learning to design antibodies. However, no such approach can demonstrate de novo antibody design in a zero-shot manner and validate it in the lab. The present technology combines generative modeling ideas with the high-throughput experimental capabilities of wet labs. Recent advances in DNA synthesis and sequencing, E. coli-based antibody expression, and fluorescence-activated cell sorting make it possible to quickly and in parallel experimentally evaluate hundreds of thousands of individual designs.

[0077] The technology of the present invention presents zero-sample antibody design through extensive wet lab experiments. As the first step of completely de novo antibody design, the technology of the present invention shows that HCDR3 can be designed by a generative AI method using a model system of trastuzumab and its target antigen human epidermal growth factor receptor 2 (HER2) as a model system. In some aspects, all antibodies binding to HER2 or HER2 homologs can be removed from the training set. The technology of the present invention can include many (e.g., about 440,000 or more) unique HCDR3 variants of trastuzumab designed from scratch, and screening for binding to HER2 using a proprietary active-specific cell enrichment ("ACE") assay. As used herein, the term "quantitative affinity active-specific cell enrichment or qaACE assay" refers to a high-throughput analysis of affinity and sequence data for obtaining biomolecule variants (US Provisional Application No. 63 / 371,474 filed on August 15, 2022 and PCT / US23 / 60167 filed on January 5, 2023, each of which is incorporated by reference in its entirety).

[0078] For example, quantitative affinity ACE ("qaACE") and ACE analysis further described herein and referred to as "de novo ACE" or "dnACE" are methods for sampling the binding of antibody variants at high throughput using flow cytometry and next generation sequencing. The main goal of the method is to generate high throughput binding information and / or training data for AI models to perform sequence-based binding predictions. The method can be applied to any antibody format, mab, fabs, scFv, scFAB, VHH, nanobodies, etc., and can also be envisioned to be applied to other binding drug formats.

[0079] In one embodiment, the first step of the qaACE process is to generate a mutationally diverse antibody library that uniformly samples the sequence space around the starting point antibody molecule. The library contains variants spanning a range of mutation distances from the original sequence.

[0080] In some embodiments including examples herein, the method provides the flow cytometry readings of antibodies combined with fluorescently labeled antigen probes expressed in SoluPro Escherichia coli. In qaACE determinations, the background expression of antibody molecules is standardized so that the change of fluorescent signal in the cell will be attributed to the different affinities of the antibody variants expressed in the cell combined with fluorescent antigen probes. This standardization is completed by a universal target molecule probe, which will be combined with all variants, and its signal will be located in a fluorescent channel orthogonal to the antigen probe. In this context, the fluorescent signal of the variant is proportional to the measured KD of the antibody variant within a certain range. Considering this proportionality, using FACS, cells containing antibody variants can be sorted, which span the range (e.g., distribution) of affinity.

[0081] After sorting a range of affinity values ​​by gating across library population distribution, the cell material is sequenced and quantified to determine the prevalence of variants observed on the affinity gate (bin, tube). Using quantification, an enrichment score is calculated for each variant. Due to accuracy and throughput, the enrichment scores generated via qaACE are an ideal data type for AI modeling purposes.

[0082] In one exemplary workflow, the present disclosure provides a qaACE assay comprising some or all of the following general steps:

[0083] 1) Generate a library of antibodies or other drug molecules for screening via qaACE expressed in host cells (e.g., SoluPro E. coli).

[0084] 2) Identify antigen or binding partner probes that are fluorescently labeled for use in, for example, FACS, through an initial cytometry development process.

[0085] 3) Use universal probes for target molecule variants that will allow detection of expression levels within cells. This expression signal is used to gate uniform expression populations to eliminate affinity and expression signals associated with epitope binding signals.

[0086] 4) Sorting cells across the affinity distribution.

[0087] 5) Sequencing of cells sorted across the affinity distribution.

[0088] 6) During sequencing, DNA barcodes or UMIs can be added via PCR amplification of regions of interest. These UMIs will enable absolute quantification of variants retrieved from the gate.

[0089] 7) Generate affinity-related enrichment scores for each observed variant.

[0090] 8) AI model training using enrichment scores and antibody variant sequences.

[0091] Additional ACE assay analyses are discussed further below.

[0092] Based on these designs, the present technology can be used to functionally validate many (e.g., 421 or more) binders using SPR, and estimate the existence of many more (e.g., about 4,000) binders in total. The designed binder sequences are not only different from those present in the training dataset, but are also highly diverse and different from anything previously observed in structural antibody databases or large datasets of known antibodies. In addition, according to a previously described naturalness metric (Bachas S et al.), the designed binders are likely to be developable and have favorable immunogenicity characteristics. We demonstrated the scalability of the present method by designing and validating binding molecules to two additional antigens, namely human vascular endothelial growth factor A (VEGF-A) and SARS-CoV-2 spike protein (COVID-19 Omicron variant).

[0093] While the primary focus of this work was on the in silico design of HCDR3, fully de novo antibody design will require the generation of multiple antibody CDR regions. We demonstrate initial progress toward this goal by using a multi-step generative AI approach to design antibodies with any antibody domain / sequence that differs from the parent antibody (e.g., all three heavy chain CDRs (HCDR1, HCDR2, HCDR3), any light chain, etc.). In summary, this work paves the way for rapid progress in fully de novo antibody design using generative AI, which has the potential to revolutionize the availability of therapeutics for patients.

[0094] The technology of the present invention represents an important advance in computer simulation antibody design, which has the potential to revolutionize the availability of effective treatments for patients. Generative AI-designed antibodies will significantly reduce development timelines by generating molecules with the desired quality without the need for further optimization. In addition, the controllability of AI-designed antibodies will enable the creation of customized molecules for specific disease targets, resulting in potentially safer and more effective treatments than traditional patient development methods. The core platform of generative AI design methods and ultra-high throughput wet lab screening capabilities will continue to drive progress in this area, unlocking new capabilities in the rapidly accelerating field of protein therapeutic design.

[0095] Screening hundreds of thousands of model-generated sequences for binding

[0096] The present technology uses our previously described ACE assay to screen a large library of antibody variants containing hundreds of thousands of members expressed in fragment antigen binding (Fab) form. The present technology verifies the ACE assay of the de novo discovery workflow by sampling the sequence for subsequent analysis by SPR, which is the gold standard in binding affinity measurement and detection. Empirical evidence found that the ACE assay can correctly classify binders with nearly 60% accuracy and >95% recall (Table S3).

[0097] Table S3. Precision and recall of the binary ACE (bACE) assay. The bACE assay predicted 353 variants to bind in SPR, and 209 were confirmed binders (precision 59.2%). An additional four binders were predicted not to bind via bACE (recall 98.1%).

[0098] Predicted binders Validated binders Missing binder Accuracy Recall 353 209 4 0.592 0.981

[0099] This enables a powerful workflow where a large number of predictions can be first screened by the ACE assay, and the expected binding population can then be screened via SPR to remove false positives and collect high-quality binding affinity measurements ( Figure 1A ).

[0100] Exemplary Computer-Implemented Machine Learning Training and Operations

[0101] 1 depicts an exemplary computing environment 100 for training and / or operating one or more machine learning (ML) models according to some aspects. The environment 100 includes a client computing device 102, a structure prediction server 104, an assay device 106, and an electronic network 108. Some aspects may include multiple client devices 102, multiple molecular modeling servers 104, and / or multiple assay devices 106. Typically, one or more molecular modeling servers 104 are used to perform training and operation of full or partial in silico molecular modeling as described herein.

[0102] The client computing device 102 may be a single server, a group of multiple servers (e.g., a cluster), or another suitable type of computing device or system (e.g., a collection of computing resources). For example, the client computing device 102 may be any suitable computing device (e.g., a server, a mobile computing device, a smartphone, a tablet, a laptop, a wearable device, etc.). In some aspects, one or more components of the client device 102 may be embodied by one or more virtual instances (e.g., a cloud-based virtualization service) and / or may be included in a corresponding remote data center (e.g., a cloud computing environment, a public cloud, a private cloud, a hybrid cloud, etc.). The client computing device 102 includes a processor and a network interface controller (NIC). The processor may include any suitable number of processors and / or processor types, such as a CPU and one or more graphics processing units (GPUs). Typically, the processor is configured to execute software instructions stored in a memory. The memory may include one or more persistent memories (e.g., a hard drive / solid-state memory) and store one or more sets of computer executable instructions / modules. For example, the executable instructions may receive and / or display results generated by the server 104.

[0103] The client computing device 102 may include a corresponding input device and a corresponding output device. The corresponding input device may include any suitable one or more devices for receiving input, such as one or more microphones, one or more cameras, a hardware keyboard, a hardware mouse, a capacitive touch screen, etc. The corresponding output device may include any suitable device for transmitting output, such as a hardware speaker, a computer monitor, a touch screen, etc. In some cases, the input device and the output device may be integrated into a single device, such as a touch screen device that accepts user input and displays output. The NIC of the client computing device may include any suitable (multiple) network interface controllers, such as a wired / wireless controller (e.g., an Ethernet controller), and facilitates bidirectional / multiplexed networking on a network between the client computing device 102 and other components of the environment 100.

[0104] The structure prediction server 104 includes a processor 150, a network interface controller (NIC) 152, and a memory 154. The structure prediction server 104 may also include a data repository 180. The data repository 180 may be a structured query language (SQL) database (e.g., a MySQL database, an Oracle database, etc.) or another type of database (e.g., not only a SQL (NoSQL) database). In some aspects, the data repository 180 may include a file system (e.g., an EXT file system, an Apple file system (APFS), a networked file system (NFS), a local file system, etc.), an object storage area (e.g., Amazon WebServices S3), a data pool, etc. The data repository 180 may include a variety of data types, such as pre-trained data (e.g., SAbDab data, OAS data) derived from a public data source, pre-trained data, and fine-tuning data. The fine-tuning data may be proprietary affinity data derived from quantitative determination ACE, Carterra, or any other suitable source. The data repository 180 may include machine learning model training data represented in any suitable data format, such as the Protein Data Bank (PDB) format, the JavaScript Object Notation (JSON) format, the Extensible Markup Language (XML) format, and the like.

[0105] The server 104 may include a client binding library for accessing the data repository 180. In some aspects, the data repository 180 is located remotely from the structure prediction server 104. For example, in some aspects, the data repository 180 may be implemented using a RESTdb.IO database, an Amazon relational database service (RDS), etc. In some aspects, the structure prediction server 104 may include a client-server platform technology that is responsive to receiving and responding to electronic requests, such as Python, PHP, ASP.NET, Java J2EE, Ruby on Rails, Node.js, web services 77, or online APIs. In addition, the structure prediction server 104 may include an instruction set for performing machine learning operations, as described below, which may be integrated with the client-server platform technology.

[0106] Determination device 106 can be a surface plasmon resonance (SPR) machine, such as Carterra SPR machine. Device 106 can be physically connected to structure prediction server 104 or data repository 180, as shown. Device 106 can be located in a laboratory, and can be accessed from one or more computers in a laboratory (not depicted) and / or from structure prediction server 104. Device 106 can generate data, and directly and / or via (multiple) laboratory computers upload the data to data repository 180. Determination device 106 can include instructions for receiving one or more sequences (e.g., mutant sequences) and synthesizing those sequences. Synthesis can sometimes be performed via another technology (e.g., via different devices or via people). In some aspects, device 106 may not be configured as a device, but is configured to measure the alternative determination of protein-protein interactions listed in other parts of the application. For example, device 106 can be configured as a kit of device / workflow alternatively, including plate and liquid handling. In general, device 106 may be replaced by suitable hardware and / or software, optionally including a human operator to generate affinity data.

[0107] The network 108 may be a single communication network, or may include multiple communication networks of one or more types (e.g., one or more wired and / or wireless local area networks (LANs), and / or one or more wired and / or wireless wide area networks (WANs), such as the Internet). For example, the network 108 may enable bidirectional communication between the client computing device 102 and the structure prediction server 104.

[0108] Processor 150 may include any suitable number of processors and / or processor types, such as one or more graphics processing units (GPUs), one or more central processing units (CPUs), etc. In general, processor 150 is configured to execute software instructions stored in memory 154. Memory 154 may include one or more persistent memories (e.g., hard drives / solid-state memories) and store one or more sets of computer-executable instructions / modules 160, including input / output (I / O) modules 162, machine learning training modules 164, and machine learning operation modules 166. In some aspects, more or fewer modules may be included, and in some aspects, one or more models may be combined or aggregated into a fewer number of modules.

[0109] As will be further described below, each of the modules 160 implements specific functionality related to the present invention. Module 160 may store machine-readable instructions, including one or more applications, one or more software components and / or one or more APIs, which may be implemented to facilitate or perform the features, functions or other disclosures described herein, such as any method, process, element or limitation illustrated, depicted or described for various flowcharts, diagrams, illustrations, drawings and / or other disclosures herein. In some aspects, multiple modules 160 may work together to implement a particular technology. For example, the machine learning operation module 166 may load information from one or more other models before, during and / or after initiating an inference operation. Thus, the module 160 may exchange data within a single computing device (such as the structure prediction server 104) via suitable technologies (e.g., via inter-process communication (IPC), representational state transfer (REST) ​​API, etc.). In some aspects, one or more modules 160 may be implemented in multiple computing devices (e.g., multiple servers 104). Module 160 may exchange data between multiple computing devices via a network (such as network 108). Now, the following will be described in more detail. Figure 1B Module 160.

[0110] Typically, the I / O module 162 includes instructions that enable a user (e.g., an employee of a company) to access and operate the structure prediction server 104 (e.g., via the client computing device 102). For example, the employee may be a software developer who trains one or more ML models using the ML training module 164 in preparation for using the one or more trained ML models to generate output for an antibody prediction project, a docking complex prediction project, and / or an affinity value prediction project. Once the one or more ML models are trained, the same user (or another user) may access the structure prediction server 104 via the I / O module in order to initiate a molecular modeling process. The I / O module 162 may include instructions for generating one or more graphical user interfaces (GUIs) (not depicted) that collect and store parameters related to biomolecular modeling, such as a user selecting a particular reference protein, biomolecule, binding partner, etc. from a list stored in the data repository 180.

[0111] In general, a computer program or computer-based product, application or code (e.g., a model, such as a machine learning model or other computing instructions described herein) may be stored on a computer-usable storage medium, or a tangible, non-transitory computer-readable medium (e.g., a standard random access memory (RAM), a compact disc, a universal serial bus (USB) drive, etc.) in which such computer-readable program code or computer instructions are embodied, wherein the computer-readable program code or computer instructions may be installed on the processor 150 or otherwise adapted to be executed by the processor 150 (e.g., working in conjunction with a corresponding operating system in the memory 154) to facilitate, implement or perform machine-readable instructions, methods, processes, elements or limitations, as illustrated, depicted or described with respect to the various flowcharts, diagrams, illustrations, drawings and / or other disclosures herein. In this regard, the program code may be implemented in any desired programming language and may be implemented as machine code, assembly code, byte code, interpretable source code, etc. (e.g., via Golang, Python, C, C++, C#, Objective-C, Java, Scala, ActionScript, JavaScript, HTML, CSS, XML, etc.).

[0112] The technology of the present invention may include a generative model to automatically design or improve antibodies for a specific target of interest (e.g., a target biomolecule). The model may be trained for an antibody-antigen structure or a large number of existing antibody sequences. The model may also be trained using proprietary high-throughput binding interaction data. In some aspects, the model may be validated by designing antibodies that bind to receptor proteins involved in cancer. The technology of the present invention may achieve zero-sample generation of antibodies customized to bind any target. The technology of the present invention may redesign the complementary determining region (CDR) of an antibody, which in some aspects imparts binding specificity to a target. The generative model of the present invention may automatically design or improve antibodies for a specific target of interest, and may pre-train a collection of antibody-antigen complexes mined from a public data source (e.g., a protein database).

[0113] In some aspects, a model trained according to the technology of the present invention can be trained to receive an antigen structure and predict the structure, sequence and / or affinity of the antibody based on the received antigen. In some aspects, the prediction can be performed by iteratively refining the model to predict the structure while autoregressively unraveling the sequence. In such cases, general autoregressive or bidirectional training can be used to capture long-range patterns in the sequence. The antibody and antigen modules can capture independent SE(3) equivalences, thereby achieving data-efficient protein-protein rigid body docking.

[0114] In some aspects, the high GPU memory of 80GB A100s can be used to realize the deconstruction of long structure graphs. In some aspects, the scalable implementation scheme of paired SE (3) equivalence can help improve training efficiency. After pre-training, the technology of the present invention can fine-tune the model to predict the binding affinity of antibodies to the antigen of interest. The affinity prediction module can use the structural and sequence features inferred by the model to predict the binding affinity of antibodies to antigens. In particular, the embedding at each residue can be inferred and propagated to the affinity prediction module with attention. The affinity prediction module can combine high-throughput binding interaction data (for example, data generated on the Integrated Drug CreationTM platform of Absci). Therefore, advantageously, the technology of the present invention realizes the novel coupling of the structural model of antibody binding with real-world binding affinity data.

[0115] The technology of the present invention can verify the trained model by designing antibodies to neutralize Her2, an important target in breast cancer. Starting from trastuzumab (a well-known Her2 binding agent), the technology of the present invention can include designing a library containing trastuzumab variants with single-order, double-order and higher-order mutations. The binding affinity of the variants in the library can be measured by high-throughput screening, and the library is scored using a generative model to prove the high correlation between model prediction and high-throughput data. In addition to pre-training structural data, or alternatively, the technology of the present invention can pre-train language models on large databases of existing antibody sequences. Language models and prediction models can be effectively integrated to further improve the prediction of binding affinity data. After training, the model can generate various new antibodies for unseen antigens by iteratively autoregressively decoding sequences. Compared with previous protein design work based on deep learning (wherein proteins are unconditionally designed), our model can be targeted for specific antigens.

[0116] In some aspects, the ML training module 164 may include a set of computer executable instructions that implement machine learning training, configuration, parameterization, and / or storage functions. The ML training module 164 may initialize, train, and / or store one or more ML models, as discussed herein. The trained ML models and their weights / parameters may be stored in a data repository 180 and / or in a memory 154, which may be accessible or otherwise communicatively coupled to the structure prediction server 104.

[0117] For example, the ML training module 164 can train one or more ML models (e.g., an artificial neural network (ANN)). One or more training data sets can be used for model training in the present technology, as discussed herein. The input data can have a specific shape that may affect the ANN network architecture. The elements of the training data set can contain tensors scaled to small values ​​(e.g., in the range of (-1.0, 1.0)). In some aspects, a preprocessing layer can be included in the training (and operation) that applies principal component analysis (PCA) or another technique to the input data. PCA or another dimensionality reduction technique can be applied during training to reduce the dimensionality from a high number to a relatively small number. Reducing the dimensionality can result in a significant reduction in the computing resources (e.g., memory and CPU cycles) required to train and / or analyze the input data.

[0118] In general, training an ANN may include establishing a network architecture or topology, adding layers, including activation functions for each layer (e.g., "leaky" rectified linear units (ReLU), softmax, hyperbolic tangent, etc.), loss functions, and optimizers. In one aspect, the ANN may use different activation functions at each layer or between hidden layers and output layers. Suitable optimizers may include Adam and Nadam optimizers. In one aspect, different neural network types may be selected (e.g., graph convolutional neural networks, message passing neural networks, geometric vector perceptron networks, recurrent neural networks, deep learning neural networks, etc.). The training data may be divided into training, validation, and test data. For example, 20% of the training data set may be retained for later validation and / or testing. In the example, 80% of the training data set may be used for training. In the example, the training data set data may be shuffled before being so divided. Dividing the data set may also be performed in a cross-validation setting, for example, when the data set is small. The data input to the artificial neural network may be encoded in N-dimensional tensors, arrays, matrices, and / or other suitable data structures. In some aspects, training may be performed using labeled training samples through continuous evaluation (e.g., loops) of the network. The process of training an ANN can cause the weights or parameters of the ANN to change. The weights can be initialized to random values. When the network is continuously trained by using one or more gradient descent algorithms, the weights can be adjusted to reduce losses and converge the values ​​of the network output to the expected or "learned" values. In one aspect, regression without an activation function can be used. Among them, the input data can be normalized by mean centering, and in addition to the mean absolute error, a mean squared error loss function can be used to determine the appropriate loss and quantify the accuracy of the output.

[0119] In some aspects, the ML training module 164 may include computer executable instructions for performing ML model pre-training, ML model fine-tuning and / or ML model self-supervised training. Model pre-training may be referred to as transfer learning, and can implement the training of a general base model, because it can be used, for example, as a universal grammar for all antibody sequences. In some instances, pre-training can be used to train multiple models of independent artificial neural networks, and / or multiple corresponding layers of a single artificial neural network (for example, an artificial neural network for predicting affinity values ​​according to biological structure information). The term "pre-training" can be used to describe the scenario where the second training may occur (that is, when the model can be "fine-tuned"). Transfer learning refers to the ability of the model to better initialize the second training with the result (weight) of the first pre-training, otherwise random initialization may be required. The second training, i.e., fine-tuning, can be performed using affinity data as discussed herein. The technology combined with pre-training and fine-tuning advantageously improves performance, because after pre-training (for example, using the natural antibody structure from SAbDab) the training results of affinity data perform better than when not pre-training. In some aspects, model fine-tuning can be performed for a given antibody-antigen pair. In some aspects, the ML model can be performed self-supervised learning to impart to the model an understanding of antibody grammar during pre-training.

[0120] Typically, supervised, semi-supervised, or unsupervised machine learning programs or algorithms can be used to train ML models as described herein. The machine learning program or algorithm can employ a neural network, which can include one or more of the following: a graph convolutional neural network, a message passing neural network, a geometric vector perceptron network, a recurrent neural network, a deep learning neural network, a convolutional neural network, a deep learning neural network, a converter, an autoencoder, and / or a combined learning module or program that learns in two or more features or feature data sets (e.g., structured data, unstructured data, etc.). The machine learning program or algorithm can also include natural language processing, semantic analysis, automated reasoning, regression analysis, support vector machine (SVM) analysis, decision tree analysis, random forest analysis, K nearest neighbor analysis, naive Bayes analysis ( Bayes analysis), clustering, reinforcement learning and / or other machine learning algorithms and / or techniques (e.g., generative algorithms, genetic algorithms, etc.).

[0121] In some aspects, an ML algorithm or technique may be selected for a particular input based on the problem set size of the input. In some aspects, algorithms based on artificial intelligence and / or machine learning may be based on or otherwise incorporated into aspects of one or more machine learning algorithms included as libraries or packages executed on server 104. For example, libraries may include libraries based on TensorFlow, Pytorch libraries (e.g., PyTorch Lightning), Keras libraries, Jax libraries, HuggingFace ecosystems (e.g., converters, data sets, and / or word segmentation libraries therein), and / or scikit-learn Python libraries. However, these popular open source libraries are good, but not necessary. The technology of the present invention may be implemented using other frameworks / languages.

[0122] Machine learning may involve identifying and recognizing patterns in existing data (e.g., structural information, docking ability, binding affinity, etc.) to facilitate prediction, classification, and / or identification of subsequent data (such as using a trained model to generate antibodies based on input antibodies, predicting the ability of the generated antibodies to dock with the input antigen (or other antigens) and / or predicting the binding affinity of the generated antibodies. Machine learning models can be created and trained based on exemplary data (e.g., "training data") inputs or data (which can be referred to as "features" and "labels") to make effective and reliable predictions for new inputs. In supervised machine learning, a machine learning model can be created and trained based on exemplary data (e.g., "training data") inputs or data (which can be referred to as "features" and "labels") to make effective and reliable predictions for new inputs. In supervised machine learning, a machine learning model can be created and trained based on a server, computing device, or other A machine learning program running on its processor provides exemplary inputs (e.g., "features") and their associated or observed outputs (e.g., "labels") so that the machine learning program or algorithm determines or discovers rules, relationships, patterns, or other machine learning "models" that map such inputs (e.g., "features") to outputs (e.g., labels), such as by determining and / or assigning weights or other metrics to the model across various feature categories of the model. Subsequent inputs of such rules, relationships, or other models can then be provided so that the model executing on the server, computing device, or other processor predicts the expected output based on the discovered rules, relationships, or models.

[0123] For example, the ML training module 164 can analyze the labeled data at the input layer of a model with a networked layer architecture (e.g., an artificial neural network, a convolutional neural network, a deep neural network, etc.) to generate an ML model. The training data can be, for example, structural information of an antibody. In some aspects, the output can be the sequence of a new protein. During training, the labeled data can be propagated through one or more connected depth layers of the ML model to establish the weights of one or more nodes or neurons of the corresponding layer. Initially, the weights can be initialized to random values, and one or more suitable activation functions can be selected for the training process, as will be understood by a person of ordinary skill in the art. The ML training module 164 can include a corresponding output layer for training one or more machine learning models. The output layer can be trained to output predictions. For example, the ML model trained herein can predict the structural information of an antibody by analyzing the labeled examples provided during training. In some aspects, the binding affinity can be expressed as a real number (e.g., in a regression analysis). In some aspects, the binding affinity can be expressed as a Boolean value (e.g., in a classification). In some aspects, multiple ANNs can be trained and / or operated separately. For example, a single model can be fine-tuned (i.e., trained) for multiple different antibody-antigen pairs using transfer learning based on a pre-trained model.

[0124] In unsupervised or semi-supervised machine learning, a server, computing device or other processor may need to find its own structure in an unlabeled exemplary input, wherein, for example, multiple training iterations are performed by a server, computing device or other processor to train multiple generations of models until a satisfactory model is generated. In the technology of the present invention, semi-supervised learning can be used in particular for natural language processing purposes, and for learning the grammar of antibody sequences using targets (such as masked language model targets). Supervised learning and / or unsupervised machine learning can also include retraining, relearning or otherwise updating the model with new or different information, which can include information received, ingested, generated or otherwise used over time. In various aspects, training the ML model herein can include generating an integrated model comprising multiple models or sub-models, including models trained by the same and / or different AI algorithms, as described herein, and configured to operate together.

[0125] Once the model machine learning training module 164 has initialized one or more ML models (e.g., which may be ANNs or regression networks), the model machine learning training module 164 trains the ML models by inputting labeled data (e.g., antibody / antigen complexes; docked biomolecules, etc.) into the models. The trained ML models can provide accurate predictions based on inputs that the models have not seen before (i.e., inputs that were not used during training).

[0126] The model machine learning training module 164 can divide the training data into corresponding training data sets and test data sets. The model machine learning training module 164 can use the labeled data to train the ANN. The model machine learning training module 164 can use the test data to calculate the accuracy / error metric (e.g., cross entropy) and test the corresponding label set. The model machine learning training module 164 can serialize the trained model and store the trained model in a database (e.g., data repository 180). Of course, those of ordinary skill in the art will understand that the model machine learning training module 164 can train and store more than one model. For example, the model machine learning training module 164 can train a single model for performing antibody generation, another model for performing antibody docking, and another model for performing affinity prediction. It should be understood that the structure of the described network can be different according to the embodiment. For example, in some aspects, antibody generation, antibody docking, and affinity prediction can each correspond to the corresponding layer of the overall machine learning model. For example, in some aspects, the trained ML model predicts many data fragments: the sequence of the antibody, the affinity of the antibody, and the docking configuration. The present invention technology enables supervision of any one of these components, and the remaining components can be improved through training. This ability is a beneficial result of the end-to-end training of the present invention technology.

[0127] The machine learning operation module 166 may include a set of computer executable instructions that implement machine learning loading, configuration, initialization and / or operation functions. The ML operation module 166 may include instructions for storing the trained model (e.g., in an electronic data repository 180, in a picked-up binary form, etc.). After training, the trained ML model can be operated in an inference mode, and when a de novo input not previously provided to the model is provided, the model can output one or more predictions, classifications, etc. as described herein. In some aspects, the loss minimization function can be used, for example, to teach the ML model to generate outputs that are similar to known outputs (i.e., ground truth samples)

[0128] Once the model is trained by the model machine learning training module 164, the model operation module 166 can load one or more trained models (e.g., from the data repository 180). The model operation module 166 generally applies new data that has not been previously analyzed by the trained model to the trained model. For example, the model operation module 166 can load a serialized model, deserialize the model, and load the model into the memory 154. The model operation module 166 can load new data (e.g., binding partner structural information) that is not used to train the trained model. For example, the new data can include data encoded as input tensors as described herein (e.g., antigen sequence data, etc.). The model operation module 166 can apply one or more input tensors to the trained ML model. The model operation module 166 can receive outputs (e.g., tensors, feature maps, etc.) from the trained ML model. The output of the ML model can include structural information of one or more biomolecules (e.g., antibodies) that are predicted to have good docking and high binding affinity with the input binding partner (e.g., antigen). The output of the ML model can include docking complexes and binding affinity values. In this way, the technology of the present invention advantageously provides a means, such as quantitatively designing a target antibody corresponding to an input target antigen, while also quantitatively docking the input antigen and the designed target antibody, and quantifying the binding affinity of the two. These techniques are more accurate and data-rich than traditional industry practices. By using ML, the technology of the present invention avoids the expense and time-consuming laboratory experimental process caused by computer simulation performance without the need to continue to use a wet lab.

[0129] The model operation module 166 can be accessed by another element of the structure prediction server 104 (e.g., a web service). The ML operation module 166 can pass its output to another module for further processing / analysis. In some aspects, a user can interact with the ML model during training and / or operation using a command line tool, an application programming interface (API), a software development kit (SDK), a Jupyter notebook, etc.

[0130] With respect to modules 160, one of ordinary skill in the art will appreciate that in some aspects, the software instructions comprising modules 160 may be organized differently and may include more / fewer modules. For example, one or more of modules 160 may be omitted or combined. In some aspects, additional modules (e.g., a positioning module) may be added. In some aspects, a software library implementing one or more modules (e.g., Python code) may be combined so that, for example, ML training module 164 and ML operation module 166 are a single set of executable instructions for training and making predictions.

[0131] Exemplary ACE Assay Analysis

[0132] In operation, the assay device 106 can be used to generate binding scores from the reads. For example, in one embodiment, the following procedure can be used:

[0133] 1. Paired-end reads can be merged using FLASH2 and the maximum allowed overlap is set based on amplicon size and sequencing read length (e.g., 150 bases for all libraries described herein).

[0134] 2. The cutadapt tool can be used to remove primers from both ends of the merged reads and discard reads if no primers are detected.

[0135] 3. Reads can be aggregated on all FACS sorting gates and aligned to the reference sequence (parental form of the amplicon) in amino acid space. The Needleman-Wunsch algorithm implemented in Biopython can be used to perform the alignment. In one example, the following parameters can be used: PairwiseAligner, mode = global, match_score = 5, mismatch_score = -4, open_gap_score = -20, extend_gap_score = -1. In some aspects, the parameters can be selected by manual inspection of multiple processed libraries.

[0136] 4. A read can be discarded if (1) the average base quality is below a number (e.g., 20), or (2) the sequence is found in fewer than a number (e.g., 10) of the reads (in DNA space).

[0137] 5. The present technology can also include marking: (1) sequences that align with a reference with a low score (e.g., defined as less than 0.6 of the score obtained by aligning the reference with itself); (2) sequences containing stop codons outside the region of interest and (3) sequences containing frameshift insertions or deletions. In some aspects, the marked sequences may not be included in any mutation-related statistics, but can be used for count normalization in conjunction with score calculations. FastQC and MultiQC can be used to generate sequencing quality control metrics.

[0138] 6. For each gate, the prevalence of each sequence (read count relative to the total number of reads from all sequences in that gate) can be normalized to the number of counts (e.g., 1 million).

[0139] 7. A binding score (e.g., ACE score) can be assigned to each unique DNA sequence by taking a weighted average of the normalized counts across the sorting gates. Across all experiments, weights can be assigned linearly using an integer scale: the gate that captures the lowest fluorescence signal is assigned a weight of 1, the next lowest gate is assigned a weight of 2, etc.

[0140] 8. Any detected sequences not present in the originally designed and synthesized library can be discarded.

[0141] 9.ACE scores can be averaged across independent FACS sorts, discarding sequences with a standard deviation of greater than 1.25 across repeated measurements. Amino acid variants can be retained only if the present technology collects multiple (e.g., at least three) independent QC passes between repeated FACS sorts.

[0142] Exemplary Dual Assay (dnACE) Analysis

[0143] In some embodiments, the enrichment scores of individual variants screened by the binary version of the ACE assay can be calculated. For example, the following program can be used:

[0144] 1. You can use Fastp to merge paired-end reads and enable quality filtering and base correction in the merged region.

[0145] 2. The cutadapt tool can be used to remove primers from both ends of the merged reads and discard reads if no primers are detected.

[0146] 3. Unique sequences can be counted to provide a raw count of each variant observed in each sample. Sequences that do not match the designed sequence in the library can be discarded.

[0147] 4. For each sample, the proportional abundance of each variant can be calculated. The enrichment score can be calculated by dividing the proportional abundance of each variant in the gate by its proportional abundance in the unsorted library sample.

[0148] Calculation of expected binding rate for ACE assay

[0149] The ACE score was discretized into a set of 6 bins, and for each bin, a set of variants belonging to the bin was screened via SPR. The percentage of bound bin variants was used to estimate the expected binding rate for the bin. These values ​​can be found in Table S2.

[0150] Table S2. SPR binding rates corresponding to ACE scores of sequences in HDL2. The binding rates in SPR increase as the sequences fall into higher ACE bins. This validates the ACE assay as a useful method for selecting binders (i.e., selecting sequences from the highest gate). The data can also be used to infer the total number of binders in each bin.

[0151] ACE Score Range Sequences screened in SPR Number of binders Binding rate in SPR [0,1) 6 134 0.045 [1,2) 13 67 0.194 [2,3) 24 46 0.522 [3,4) 41 44 0.932 [4,5) 117 128 0.914 [5,6) 180 191 0.942

[0152] Exemplary Surface Plasmon Resonance (SPR) Aspects

[0153] As discussed above, SPR can be used for a variety of techniques, including functional validation of binding agents.

[0154] Surface preparation

[0155] For example, in some aspects, the sample can be transferred to a plate (e.g., a 96-well plate (e.g., Greiner Bio-One)) after induction, precipitated and lysed in 50 μL lysis buffer (e.g., 1X BugBuster protein extraction reagent containing 0.01X KU Benzonase nuclease and 1X protease inhibitor mixture). The plate can be incubated at 30°C for 15 to 20 minutes and then centrifuged to remove insoluble debris. After lysis, the sample can be adjusted to a final volume of 260 μL and filtered into a 96-well plate with 200 μL SPR running buffer (e.g., 10 mM HEPES, 150 mM NaCl, 3 mM EDTA, 0.01% w / v Tween-20, 0.5 mg / mL BSA). The lysed sample can then be transferred from a 96-well plate to a larger plate (e.g., 384-well plate) for high-throughput SPR, for example, using a Hamilton STAR automated liquid handler. Prior to lysis, colonies can be prepared in two independent sets of replicates and each replicate measured in two separate experimental runs. In some cases, a single replicate can be used as indicated.

[0156] SPR

[0157] High-throughput SPR experiments can be performed on a microfluidic Carterra LSA SPR instrument using SPR running buffer (e.g., 10 mM HEPES, 150 mM NaCl, 3 mM EDTA, 0.01% w / v Tween-20, 0.5 mg / mL BSA) and SPR wash buffer (e.g., 10 mM HEPES, 150 mM NaCl, 3 mM EDTA, 0.01% w / v Tween-20). Prior to performing the experiment, the Carterra LSA SAD200M chip can be pre-functionalized with 20 μg / mL biotinylated antibody capture reagent for 600 seconds. Lysed samples in a 384-well block can be immobilized on the chip surface for 600 seconds, followed by a 60 second elution step for baseline stabilization.

[0158] Antigen binding can be performed using a non-regenerative kinetic method with a 300 second association phase, followed by a 900 second dissociation phase. For analyte injection, six leading blanks can be introduced to create a consistent baseline before monitoring antigen binding kinetics. After the leading blank, five concentrations of HER2 extracellular domain antigen (e.g., ACRO Biosystems, prepared from a starting concentration of 500 nM with three-fold serial dilution) can be injected into the instrument and the time series response is recorded. In some experiments, the measurement of a single DNA variant can be repeated four times. Each experimental run can be composed of two complete measurement cycles (ligand fixation, leading blank injection, analyte injection, chip regeneration), and each run can provide two repeated measurement attempts for each clone. In some experiments, the technical repeats measured in a separate run can further double the number of measurement attempts for each clone to four times.

[0159] Demonstration Sequencing Technology

[0160] Low diversity library sequencing

[0161] In order to identify the DNA sequence of a single antibody variant evaluated by SPR, a repeat plate can be provided for sequencing. A portion of the precipitated material can be transferred to a 96-well PCR (e.g., Thermo-Fisher) plate by a pinner (e.g., Fisher Scientific), which can contain reagents for performing the initial phase PCR of a two-phase PCR, for adding Illumina adapters and sequencing. For example, the reaction volume used can be 12.5 μl. During the initial PCR phase, part of the Illumina adapters can be added to the amplicon by 4 PCR cycles. The second phase PCR can add the remainder of the Illumina sequencing adapter and Illumina i5 and i7 sample indexes. The initial PCR reaction can be used for, for example, 0.45 μM UMI primer concentration, 6.25 μl Q5 2x master mix (NEB). Reactions can be denatured initially at 98°C for 3 min, followed by 4 cycles of denaturation at 98°C for 10 sec, 59°C for 30 sec, 72°C for 30 sec, and a final extension of denaturation at 72°C for 2 min. After the initial PCR, 0.5 μM of the secondary sample index primer can be added to each reaction tube.

[0162] The reaction can then be denatured at 98°C for 3min, followed by 29 cycles of denaturation at 98°C for 10 seconds; denaturation at 62°C for 30 seconds; denaturation at 72°C for 15 seconds; and finally extended to denaturation at 72°C for 2min. The reactions are then combined into 1.5mL tubes (Eppendorf). The size of the combined samples can be selected using the 1xAMPure XP (Beckman Coulter) bead program, and the resulting DNA samples can be quantified by a Qubit fluorometer. In some aspects, the pool size can be verified via a Tapeststation 1000HS and sequenced on an Illumina MiSeq Micro (2×150nt) for a HCDR3 library or on an Illumina MiSeq kit v3 (2×300nt) for a HCDR1-HCDR3 library with 20% PhiX.

[0163] After sequencing, amplicon reads can be merged using Fastp, trimmed with cutadapt

[44] , and each unique sequence enumerated. Next, a custom R script can be applied to calculate the sequence frequency ratio between the most abundant sequence and the second most abundant sequence in each sample. The Levenshtein distance between the two sequences can be calculated. These distance values ​​can be used for downstream filtering to ensure that clonal populations are measured by SPR. The dominant sequence within each sample can be compared to the design sequence and discarded if it does not match any expected sequence. The dominant sequence can then be combined with its companion Carterra SPR measurement.

[0164] Exemplary computer-implemented method

[0165] In some aspects, the present technology may include training one or more machine learning models to generate target biomolecular structures and / or sequences.

[0166] Figure 2A An exemplary computer-implemented method 200 for training a machine learning model to generate structural information of a biomolecule according to some aspects is depicted. The method 200 can include receiving one or more training inputs, including one or more of: (i) input biomolecule structural information, (ii) input biomolecule binding partner structural information, or (iii) input biomolecule-input binding partner binding complex structural information (block 202).

[0167] Typically, the one or more training inputs may include data representing one or more of: (i) an amino acid; (ii) a peptide sequence; (iii) a polypeptide sequence; (iv) a primary sequence; (v) one or more secondary structures; (vi) one or more tertiary structures; (vii) one or more quaternary structures; or (viii) three-dimensional coordinates of a primary sequence corresponding to an input biomolecule, an input biomolecule binding partner, or an input biomolecule-input binding partner binding complex.

[0168] The method 200 may include processing one or more training inputs with a machine learning biomolecule prediction model to generate predicted biomolecule structure information (block 204). The method 200 may include evaluating a loss function that compares the predicted biomolecule structure information to a ground truth (block 206). The method 200 may include modifying one or more values ​​of one or more parameters of the machine learning model based at least in part on the loss function (block 208). In some aspects, the method 200 may be performed by the ML training module 164 of FIG. 1 .

[0169] In some aspects, evaluating the loss function for comparing the predicted biomolecule structure information to the ground truth to train the machine learning model at block 206 may include configuring a second model that acts as a critic to train the machine learning model to make predictions based on the ground truth data. Such a technique is an adversarial architecture (e.g., a generative adversarial network) that may be used in this case.

[0170] In some aspects, the biomolecule can be an antibody or a protein. In some aspects, the biomolecule structural information comprises the three-dimensional coordinates of the primary sequence. In some aspects, the biomolecule binding partner is an antigen, a receptor, a ligand or a cell membrane. In some aspects, the training input is represented as at least one of the following: (i) protein database (PDB) data format, (ii) JSON data format or (iii) XML data format. In some aspects, other suitable formats can be used. In some aspects, the biomolecule binding partner is a protein, and the biomolecule binding partner structural information comprises a three-dimensional structure. In some aspects, method 200 may also include receiving training inputs from a database comprising antibody structures and / or antigen structures. For example, training inputs may be received / retrieved from the structural antibody database (SAbDab). Those of ordinary skill in the art will appreciate that data processing steps may be helpful in performing the technology of the present invention. For example, in order to assess whether the ML model can be generalized to a specific binding partner (e.g., an antigen) that is very different from that used for model training, all biomolecules in the training set that are within 40% of any protein sequence identity in the corresponding validation and test sets may be removed.

[0171] Figure 2BA data flow block diagram 220 for training and operating one or more machine learning models according to some aspects of the present technology is depicted. In diagram 220, the antibody generation step occurs at box 222a. During training, as shown, training information flows from the antibody generation box 222a to the docking box 222b, where the docking step occurs. Training information also flows from the docking box 222b to the affinity prediction box 222c, where affinity prediction occurs. During training, training signals also flow in the reverse direction, as shown. For example, Figure 2B Can be depicted using Figure 2A Method 200 for data flow during training. In some aspects, the training and / or operation of one or more antibody generation machine learning models can be performed as discussed in "Iterative Refinement Graph Neural Network for Antibody Sequence-Structure Co-design" Jin et al., arXiv:2110.04624[q-bio.BM]. However, as will be understood by those of ordinary skill in the art, the technology lacks the use of input binding partners (e.g., antigens). In some aspects, docking can be performed as discussed in "Independent SE (3)-Equivariant Models for End-to-End Rigid Protein Docking" Ganea et al., arXiv:2111.07786[cs.AI].

[0172] Figure 2C Depicted is a computer-implemented method 230 for generating structural information of a target biomolecule according to some aspects. The method 230 may include receiving a target input including one or more of a target binding partner sequence, a target binding partner, or a target epitope (block 232). The method 230 may also include predicting the structural information of the target biomolecule by processing the target input with a biomolecule prediction model (block 234). In some aspects, the structural information is represented in a computer-readable medium as at least one of: (i) a protein data bank (PDB) data format, (ii) a JSON data format, or (iii) an XML data format. In some aspects, the machine learning biomolecule prediction model of the method 200 and / or the method 230 is an artificial neural network. The artificial neural network may include at least one of a geometric neural network, a transformer network, or a geometric vector perceptron network.

[0173] Typically, the structural information of the target biomolecule includes data representing one or more of the following: (i) the amino acids of the target biomolecule; (ii) the peptide sequence of the target biomolecule; (iii) the polypeptide sequence of the target biomolecule; (iv) the primary sequence of the target biomolecule; (v) one or more secondary structures of the target biomolecule; (vi) one or more tertiary structures of the target biomolecule; (vii) one or more quaternary structures of the target biomolecule; or (viii) the three-dimensional coordinates of the primary sequence of the target biomolecule.

[0174] Method 230 may include storing (e.g., in memory 154, data repository 180, etc.) and / or outputting data representing a portion or all of the structural information of a biomolecule. For example, the data may represent one or more of: (i) amino acids of a target biomolecule; (ii) a peptide sequence of a target biomolecule; (iii) a polypeptide sequence of a target biomolecule; (iv) a primary sequence of a target biomolecule; (v) one or more secondary structures of a target biomolecule; (vi) one or more tertiary structures of a target biomolecule; (vii) one or more quaternary structures of a target biomolecule; or (viii) three-dimensional coordinates of the primary sequence of a target biomolecule. In some aspects, the polypeptide sequence of the target biomolecule corresponds to a complementarity determining region of an antibody. In some aspects, the structural information of the target biomolecule is selected from the group consisting of: an alpha helix, a beta pleated sheet, and a coil. In some aspects, the target biomolecule is an antibody and the target binding partner is an antigen. Method 230 may include repeating block 234 until the structural information of the target biomolecule is complete. In some aspects, method 230 may be performed iteratively.

[0175] In some aspects, method 230 may include receiving an input parameter specifying an expected length of amino acids of a target biomolecule and generating a target biomolecule having the expected length. In some aspects, a machine learning biomolecule prediction model may determine the length by predicting a sequence end token. In some aspects, method 230 may include updating structural information of the target biomolecule by folding one or more proteins. In some aspects, the target binding partner is at least one of (i) a primary amino acid sequence, (ii) an antigenic epitope, and / or (iii) a known three-dimensional structure of an antigen. In some aspects, the structural information of the target biomolecule is predicted in at least one of (i) an N-terminal to C-terminal order, (ii) a C-terminal to N-terminal order, or (iii) by random sampling. Advantageously, the present technology is not limited to predictions in any fixed order.

[0176] In some aspects, method 230 may further include operating and / or training a machine learning docking model, the model being trained to generate a docking complex from two or more input three-dimensional biomolecules. In some aspects, method 230 may further include processing the predicted structural information of the target input and the target biomolecule using the docking model to generate a docking complex comprising the target input and the target biomolecule. In some aspects, method 230 may include providing the docking complex as a docking complex output. In some aspects, method 230 may further include operating and / or training a machine learning binding affinity prediction model, the model being trained to predict the binding affinity from the input docking complex. Method 230 may include processing the docking complex using the trained affinity prediction model to generate a binding affinity value. In some aspects, the machine learning model is an artificial neural network. In some aspects, the artificial neural network is at least one of a graph convolutional neural network, a message passing neural network, or a geometric vector perceptron network. In some aspects, method 230 may include using a pooling operation to summarize the node information contained within the artificial neural network to generate a binding affinity value. In some aspects, method 230 may include providing the docking complex as a docking complex output. In some aspects, a machine learning model can be trained to predict binding specificity in addition to or instead of binding affinity.

[0177] In some aspects, method 230 may include: forcing teaching (teacherforce) the biomolecule prediction model by providing sequence data as input, thereby training the machine learning biomolecule prediction model; forcing teaching the docking model by providing antibody structure data as input, thereby training the docking model; and / or forcing teaching the affinity prediction model by providing sequence data as input, thereby training the affinity prediction model. Specifically, the biomolecule prediction model can design one or more new antibodies. In the early stages of training, the design capability may perform poorly (for example, the model may never reach a level where the training signal is meaningful). Under forced teaching, the problem becomes simpler. At each stage, the model is trained to view the last benchmark truth position, so that the model focuses on the most important information during the teaching training stage.

[0178] The model can be trained using antibody / antigen structures (e.g., antibody / antigen complexes in a protein database such as SAbDab). The model can be configured to issue an initial guess about a drug molecule or antibody. The guess can be generated by an instruction to iteratively output one residue or structure at a time based on an input binding partner (e.g., an antigen). Initially, the trained model can predict the first structure of the target biomolecule without knowing the corresponding sequence but knowing the target antigen (e.g., provided as an input to the trained model). The ML model can iteratively predict the first amino acid and update the structure of the target biomolecule, then predict the second amino acid, and so on, until the structure of the target biomolecule is complete, or until another stopping condition occurs. Stopping conditions can include a trained model that has made predictions for each corresponding input, convergence criteria, filling a portion of a CDR, etc. The predictions of the trained ML model can be made in any direction, including (i) in the order of N-terminus to C-terminus, (ii) in the order of C-terminus to N-terminus, or (iii) by random sampling. The antigen provided as input to the trained ML model can be a primary amino acid sequence, an antigenic epitope, and / or a known three-dimensional structure of an antigen, optionally including both the antigen and the epitope. The model is capable of designing the entire structure or a portion thereof (e.g., CDR3). However, empirical studies may show that better results indicate that coarsening or averaging residues may perform poorly, that is, not comparable to aspects of the present technology that analyze the entire structure (e.g., the entire antibody).

[0179] In some aspects, method 230 may include training a machine learning affinity prediction artificial neural network, including: (i) one or more biomolecule prediction layers, which are trained to predict biomolecule structural information based on a target input; (ii) one or more docking layers, which are trained to generate docking complexes from two or more input three-dimensional biomolecules; (iii) one or more affinity prediction layers, which are trained to predict affinity from the input docking complex. The one or more biomolecule prediction layers, the one or more docking layers, and the one or more affinity prediction layers may be connected (fully connected or otherwise connected). Method 230 may include receiving a target input comprising one or more of a target binding partner sequence, a target binding partner, or a target epitope; and processing the target input using the affinity prediction artificial neural network to generate a docking complex corresponding to the target input and a corresponding structural affinity value. In some aspects, method 230 may include receiving a sequence affinity prediction value from an affinity sequencing prediction model; and averaging the structural affinity value with the sequence affinity prediction value to generate an overall affinity prediction value.

[0180] In some aspects, method 230 may include training a machine-learned affinity prediction artificial neural network (or its layers) to generate different generations. For example, given an antigen, it is predicted that the ANN may generate millions of antibodies. In order to sort / filter those antibodies and select the antibodies that we think are most likely to work, the technology of the present invention may include using other techniques to sort antibodies, including those described in U.S. Provisional Patent Application No. 63 / 297,679; the provisional patent application is incorporated herein by reference in its entirety.

[0181] In some aspects, method 230 may include pre-training one or more biomolecule prediction layers using bound or unbound structures: pre-training one or more biomolecule prediction layers and one or more docking layers using bound structures; and / or pre-training a machine-learned affinity prediction artificial neural network using affinity training data. In some aspects, method 230 may include receiving an external docking complex corresponding to a target input; and comparing the external docking complex with the generated docking complex. Therefore, the technology of the present invention can be advantageously used for validation purposes. In some aspects, method 230 may include deriving affinity training data from an assay (e.g., an ACE assay). The affinity training data may include an affinity score proportional to activity. Method 230 may include controlling the gradient flow of a machine-learned affinity prediction artificial neural network by applying a stop gradient function in at least one of the one or more biomolecule prediction layers, one or more docking layers, or one or more affinity prediction layers. For example, in PyTorch, stop gradients can be implemented using `.detach()`, which separates the associated variables from the computational graph, thereby creating a new graph. In aspects where the same network weights are applied in each iteration, stopping gradient can be employed so that each iteration is a separate graph. The gradients can then be averaged. In this context, stopping gradient can be conceptually viewed as a strategy for enhancing limited data.

[0182] In general, method 230 can be used to perform de novo discovery of biomolecules (eg, drug candidates such as antibodies) in silico. Figure 3AA flow chart of an exemplary computer-implemented method 300 for performing de novo computer-simulated biomolecule discovery according to some aspects is depicted. Method 300 may include receiving a scaffold and an input target (box 302a). Method 300 may include analyzing the scaffold and the input target (e.g., via method 200) to generate one or more biomolecules (e.g., one or more antibodies) using a trained machine learning model (e.g., via method 230). Method 230 may also include analyzing the output of the machine learning model using an advance optimization machine learning model at box 302b (box 302c). The advance optimization machine learning model may generate one or more candidates that method 230 may validate (box 302d). Method 230 may also include generating one or more candidate drugs and / or one or more corresponding manufacturing cell lines (box 302e). In Figure 3A The ML model for designing biomolecules (e.g., one or more antibodies) against a selected target (e.g., a target antigen) at block 302(b) may include binders. These binders may be input to the machine learning model at block 302c to further optimize potential candidates.

[0183] Figure 3B An exemplary block diagram 304 is depicted according to some aspects. For example, the blocks of the block diagram 304 may correspond to various elements of the computing environment 100 of FIG. 1 . The block diagram 304 includes an antibody generation block 306a. At block 306a, representing antibody generation, a structure-modulated antibody design may be performed. For example, block 306 may correspond to Figure 2B At block 222a of FIG. 304 , one or more biomolecule generation machine learning models are trained / operated. Block diagram 304 includes an end-to-end docking block 306b, which in some aspects may include geometric learning (i.e., graph-based learning). For example, block 306b may correspond to Figure 2B At block 222b of the present invention, one or more docking (e.g., antibody-antigen docking) machine learning models are trained / operated. Block 304 may include affinity (e.g., binding affinity) prediction block 306c, which may perform graph-based affinity prediction. In some aspects, block 306c may correspond to Figure 2B The affinity prediction machine learning model training / operation at block 222c of FIG. In some aspects, block diagram 304 may include a data source block 306d, which may include public data sources (e.g., public databases) and proprietary assay data for performing high-throughput experiments. Block 306e depicts a computer architecture that may be used to perform the computer-implemented techniques of the present invention. For example, block 306e depicts a GPU that may correspond to GPU 150 of FIG. 1 .

[0184] Figure 3CA block diagram 308 of an exemplary CDR3 in a computer-simulated discovery input and output is depicted. Block diagram 308 includes a target (objective) (box 310a). In the depicted example, the target is to design a CDR3 targeting epidermal growth factor receptor (EGFR) (see, for example, https: / / www.rcsb.org / structure / 4krl). However, it will be appreciated by those of ordinary skill in the art that other targets are envisioned, including other binding partners (e.g., other antigens, epitopes, etc.) and other scaffolds. Block diagram 308 also includes a target input, which includes a target input, which includes one or more of a target binding partner sequence, a target binding partner, or a target epitope (box 310b). At box 310b, one or more trained ML models can analyze the target input to predict structural information. Starting from an EGFR antigen and a given scaffold (e.g., VHH), the model can be trained to generate a CDR3 of interest. At this point, the model may have never seen (i.e., provided with representations) similar CDR3 or training data for similar antigens. In some aspects, even those sequences that share more than a certain percentage (e.g., 40% sequence homology) with the target binding partner can be removed from the training data. In some aspects, antigen structures of the same class, architecture, topology, or homology superfamily as the target binding partner can be removed from the training data. Block 310c can include an antibody scaffold. Given an input antigen and an antibody scaffold sequence, the ML model can output a desired sequence (e.g., a CDR3 sequence of a VHH scaffold targeting EGFR).

[0185] Figure 3D An exemplary block diagram 312 of drug candidate design based on machine learning according to an embodiment is depicted. Figure 312 includes a block 314a, which depicts an isovariant representation of a biomolecular structure. In particular, block 314a depicts structural prediction, rotational isovariance, and a local coordinate system for biomolecular structural information. In addition, block 314b depicts that the attention layer can enable nonlinear modeling of potential epitope-paratope interactions, as well as various modeling architectures that can be used. Further, block 34c depicts iterative design of candidate drugs.

[0186] Figure 3EAn exemplary flow diagram 316 of antigen encoding according to some aspects is depicted. Diagram 316 includes an ML model trained by receiving one or more training inputs, which may be one or more of: (i) input biomolecule structure information, (ii) input biomolecule binding partner structure information, or (iii) input biomolecule-input binding partner binding complex structure information (EGFR antigen, to continue the above example) (block 318a). Diagram 316 may inject the antigen onto the graph (block 318b). The ML model may be trained by passing the training input around the graph (e.g., trained by the ML training module 164 of FIG. 1 ). Thus, for example, each "node" or residual of the graph at block 318b may be summarized by its neighbors. The ML model may retain this information, and during the training process will learn how to extract meaningful information, such as point cloud relationships (block 318d).

[0187] Figure 3F An exemplary flowchart 320 of antibody generation training according to some aspects is depicted. In Figure 320, the ML model starts with an initial guess of the antibody structure based on the input sequence (e.g., SAbDab sequence) (box 322a). The ML model can predict the target biomolecule structure information (e.g., the next sequence of the antibody structure) based on the antigen encoding information learned at flowchart 316 (box 322b). The model can output a prediction that can be compared with the benchmark truth data to adjust the weight of the model, depending on whether the prediction is correct (box 322c). For example, the ML model can evaluate a loss function at box 322c and modify one or more parameters of the ML model based on the loss function. As discussed herein, the antibody training in flowchart 320 can continue until the structural information of the target biomolecule is complete.

[0188] Figure 3G Depicted is an exemplary flow chart 324 of antibody generation training / operation according to some aspects. Diagram 324 can include analyzing an input antibody scaffold including an unknown region (e.g., a CDR3 region), as depicted by a string including the letter "X". Block diagram 324 can include making a first prediction using a trained ML model, wherein the output is a probability distribution about the next sequence in the unknown region. In some aspects, the attention function can be used with a softmax activation function. The first prediction can be the next sequence of CDR3. Figure 3H Depicted is an exemplary flow diagram of antibody generation training / operation according to some aspects. Figure 3HA second prediction (i.e., prediction t+1) is depicted, in which the second sequence of CDR3 is filled by the model. In some aspects, the ML model can make all predictions non-iteratively in a single time step. For example, in some aspects, the guessed structure is a random structure. In other aspects, the same starting structure is used each time the model is run. In some aspects, other machine learning architectures can work, including those used in other disciplines / arts (e.g., visual artificial intelligence / machine learning fields). In particular, those of ordinary skill in the art will appreciate that diffusion models and other generative modeling techniques that work well on continuous systems (e.g., image data) can be applicable to the present technology.

[0189] Fig. 3I Depicted is an exemplary flowchart of antibody generation training according to some aspects. In particular, Fig. 3I Depicts the predicted instances that match the known ground truth. Therefore, Fig. 3I is an example of negative reinforcement. Figure 3J Depicted is an exemplary flowchart of antibody generation training according to some aspects. Specifically, Figure 3J The generated antibodies are depicted as examples of three-dimensional structures / frameworks. For example, the depicted three-dimensional structure can be presented using three-dimensional coordinates corresponding to the primary sequence of the target biomolecule. Figure 3K Depicted is an exemplary flowchart of antibody generation training according to some aspects. In particular, Figure 3K Depicts, for example, Figure 3F and Fig. 3I The third correct value of the reference truth value. Figure 3L An exemplary block diagram depicting supervised machine learning model outputs according to some aspects is depicted. Block diagram 334 includes structural information of a target biomolecule and prediction of one or more sequences thereof. Block diagram 334 also depicts that additional loss functions can be applied with respect to physical properties, naturalness, confidence, docking, affinity, etc. of the structural information of the target biomolecule.

[0190] Demonstration results

[0191] Figure 4Depicted is an exemplary graphical user interface 400, which includes visual predicted structural information 402a and visual benchmark true value structural information 402b. Visual predicted structural information 402a can correspond to the output of affinity prediction block 222a. Visual benchmark true value structural information 402b can correspond to data received from a database or other source (e.g., protein database, wet lab crystallography, etc.). Visual predicted structural information 402a can be generated by inputting a target input, and the target input includes one or more of the following: target binding partner primary sequence, target binding partner three-dimensional coordinates, target binding partner epitope primary sequence, or target binding partner epitope primary sequence three-dimensional coordinates, or a fragment or part of any of the foregoing as described herein. For example, in some instances, the input to the ML model can be one or more protein sequences. Visual predicted structural information 402a and visual benchmark true value structural information 402b can be generated by calculating the pairwise distance algorithm between the individual points of the visual predicted structural information 402a and the visual benchmark true value structural information 402b. Brighter values ​​may indicate values ​​that are further apart, while darker values ​​may indicate values ​​that are closer together. Note that the diagonal lines (coordinates i,i) represent the same residue.

[0192] With respect to the results of the inventive technique, the fact that the visualized predicted structural information 402a and the visualized ground truth structural information 402b are visually similar indicates that the trained ML model / ANN is working as expected. In particular, the protein folding component of the network is working and outputting the design along the way.

[0193] Demonstration of computer-simulated biomolecular design

[0194] Figure 5A An exemplary flow chart 500 for fully computer-simulated biomolecule design according to some aspects is depicted. Specifically, a set of patient samples can be collected, where each patient has one or more diagnosed diseases (block 502). One or more AI models can be applied to the patient samples to discover new biomolecular targets (blocks 504, 506). For example, the AI ​​model can correspond to the model trained by the ML training module 164 of FIG. 1.

[0195] FIG. 5 depicts an exemplary block 500 of a fully in silico de novo biomolecular design artificial intelligence model according to some aspects. Figure 3A As discussed, an antibody scaffold may be processed by an ML model to generate one or more designed CDRs 516 (eg, corresponding to one or more respective drug candidates) (block 512 ). Figure 5BThe lead optimization model is further depicted, including in silico selection of amino acid identities within the trastuzumab CDR3 to increase the affinity of the antibody for the antigen (Her2) (block 514).

[0196] Figure 5C An exemplary block diagram 520 of a computer-simulated biomolecule design process according to some aspects is depicted. A set of patient samples can be collected, wherein the patient suffers from one or more diagnosed diseases (box 522). Given a target input, de novo discovery can be performed in silico using one or more trained ML models as discussed herein to predict structural information of the target biomolecule (box 526-A). In some aspects, the ML model can receive additional information (e.g., information from an ACE assay) (box 524). In some aspects, a lead optimization machine learning model as discussed herein can be used to perform advance optimization (box 526-B). The ML model can output novel biomolecules that target specific epitopes with desired scaffolds and / or desired properties (box 528). These outputs can be used to manufacture advanced production cell lines and generate drug candidates (box 530). One of ordinary skill in the art will appreciate that current computer-simulated design techniques may benefit from the use of supercomputing resources (e.g., accelerated neural network training environments such as the GNN SE(3) converter, hosted solutions for AI development (e.g., DGX Manufacturing), and graphics processing unit (GPU) supercomputers (e.g., AX100), etc.).

[0197] Demonstration Machine Learning Modeling Results

[0198] As discussed herein, the present technology includes ML modeling to automatically design or improve biomolecules for specific targets of interest. For example, in some cases, models can be pre-trained (e.g., using biomolecular complexes mined from protein databases) to predict the structure, sequence and / or affinity of antibodies given the antigen structure. Recent tests have quantified that the present technology is very effective even without the added benefit of pre-training ML models to predict the binding affinity of target binding partners and predicted target biomolecular structure information.

[0199] Fig. 6A Depicted are exemplary computer-simulated biomolecular engineering according to some aspects, including prediction of binders based on optimization of CDRH3 while keeping the remaining CDRs unchanged. Specifically, Fig. 6A Depicted are the wild-type heavy biomolecules, namely, trastuzumab wild-type light chain HER2, bH1 light chain HER2, and bH1 light chain VEGF-A. Each of these biomolecules has a different K d values, 2nM, 26nM and 300nM respectively. Figure 6BDepicted are exemplary Ranibizumab and Trastuzumab bH1 visualizations according to some aspects.

[0200] Figure 6C Depicted are exemplary sensorgrams generated corresponding to SPR results for predicted in silico binding of trastuzumab-bH1 CDR3 variants to VEGF, according to some aspects. Fig.6D Described according to some aspects, corresponding to Figure 6C Exemplary tabular data of SPR results generated from predicted computer-simulated binding of trastuzumab-bH1 CDR3 variants to VEGF. Specifically, Figure 6C and Fig.6D Depicted is the output corresponding to inputting the CDRH1 and CDRH2 of trastuzumab-Bh1 and the structure of VEGF into the trained ML model to design CDR3 of trastuzumab-Bh1. The heavy chain CDRs can be identical for trastuzumab and trastuzumab-Bh1, but the light chain CDRs may be different. The results show that the predicted variants specifically bind to VEGF, as shown in Fig.6D , where the Name column depicts control biomolecules (e.g., ABS26144 = Trastuzumab and ABS26148 = trast-bH1). Trastuzumab bH1 is a known variant of trastuzumab that is designed to bind both HER2 and VEGF (see Bostrom J, Haber L, Koenig P, Kelley RF, Fuh G (2011) High Affinity Antigen Recognition of the Dual Specific Variants of Herceptin Is Entropy-Driven in Spite of Structural Plasticity. PLoS ONE 6 (4): e17887. doi: 10.1371 / journal.pone.0017887)). Using the structure of bH1 bound to VEGF, CDR3 variants that will bind to VEGF are predicted. The Name column also depicts the predicted biomolecules, starting with the prefix "Trast_".

[0201] Fig. 6E Depicted are exemplary sensorgrams generated corresponding to SPR results for predicted in silico binding of trastuzumab CDR3 variants to Her2 binding, according to some aspects. Fig. 6F Depicted are exemplary tabular data corresponding to SPR results generated for predicted computer-simulated binding of the trastuzumab CDR3 variants of FIG. 6 to Her2, according to some aspects. Specifically, Fig. 6E and Fig. 6F Depicted is the output corresponding to the input of the CDRH1 and CDRH2 of trastuzumab and the structure of Her2 into the trained ML model to design the CDR3 of trastuzumab. In particular, the structure of trastuzumab bound to HER2 can be used to predict the CDR3 of the heavy chain (the other five CDRs are from trastuzumab) to predict the CDR3 variants that will bind to HER2. Antibodies can be produced and affinity measured using SPR to verify the predicted binding. The results show that the predicted variants specifically bind to Her2.

[0202] Generally speaking, Figures 6A-6D The data presented in demonstrate that the inventive technique for in silico design of target biomolecules is capable of predicting biomolecules that bind to intended targets, in some cases with significantly higher affinity than known controls, while providing the advantages discussed herein, such as completely avoiding time-consuming and resource-intensive traditional wet lab methods. Furthermore, the results identified to date do not necessarily require fine-tuning of the docking complexes and affinity prediction techniques described herein. That is, Figures 6A-6D The results in can be achieved by one or more biomolecule prediction layers trained to predict biomolecule structural information based on target input. Adding fine-tuning (e.g., one or more docking layers trained to generate docking complexes from two or more input three-dimensional biomolecules, and one or more affinity prediction layers trained to predict affinity from the input docking complexes) is expected to produce higher quality results.

[0203] Screening hundreds of thousands of model-generated sequences for binding

[0204] In some aspects, the technology of the present invention utilizes ACE determination (e.g., Bachas S et al.; Liu J.Activity-specific cell enrichment; Patent Publication No. WO2021 / 146626, 22.07.2021, etc.) to screen a large-scale antibody variant library containing hundreds of thousands of members (e.g., expressed in Fab 64 format). After screening by ACE, the technology of the present invention can verify the determination of the de novo discovery technology described herein by sampling the sequence for subsequent analysis by SPR, which is the gold standard in binding affinity measurement and detection. Empirical results show that the ACE determination has a high recall rate, correctly classifying> 95% of SPR binders. In addition, the determination has high accuracy, with 60% of ACE binders verified in SPR. This can achieve a powerful workflow, in which a large number of predictions can be first screened by ACE determination, and then the expected binding population can be screened via SPR to remove false positives and collect high-quality binding affinity measurements, as discussed herein.

[0205] The technology of the present invention demonstrates the ability of generative AI to design de novo antibodies and antigens of interest and to generate new HCDR3 sequences for known antibody zero samples. The technology of the present invention focuses on the design of the HCDR3 region, which is a key determinant of antibody function due to its high sequence diversity and high density of paratope residues in the immune repertoire. For example, the technology of the present invention can select trastuzumab bound to HER2 as a scaffold antibody to test the sequence of HCDR3 design. The HCDR3 design is generated by a model conditioned on the HER2 3-D structure (PDB: 1N8Z (chain C)) and the trastuzumab scaffold sequence (excluding HCDR3) modified only by the antigen. For the model used herein, the technology of the present invention can remove any antibody known to bind the target or any homolog (> 40% sequence identity or part of the same homologous superfamily) to the target. In some environments, all antibodies with> 40% sequence identity to the wild-type antibody are removed from the training set instead. In all cases, binders were observed. In general, 440,354 antibody variants were generated and screened using the ACE assay to identify binding variants. The inventive technology found a total of approximately 4000 estimated binders based on expected ACE assay binding rates and advanced a subset for further characterization, with a total of 421 AI designs confirmed for HER2 binding by SPR.

[0206] Table S5. Mean naturalness scores for different groups and p-values ​​for scaffolds using trastuzumab.

[0207] Group Average naturalness p-value (compared to de novo) From the beginning 0.175 N / A Trastuzumab 0.235 <![CDATA[3.66·10 -95 ]]> OAS 0.194 <![CDATA[3.63·10 -16 ]]> Frequency Baseline 0.143 <![CDATA[4.67·10 -53 ]]> Phage display baseline 0.136 <![CDATA[1.11·10 -87 ]]> Out-of-order OAS 0.101 <![CDATA[8.05·10 -300 ]]>

[0208] The identified binders showed a range of affinities for HER2, with 71 designs exhibiting affinities <10 nM ( Fig. 8A ). Excitingly, three of the zero-shot designs showed tighter binding than trastuzumab, with one binding in the subnanomolar affinity range. These high-affinity design examples were generated from model zero-shots without any additional affinity maturation, thus skipping a typically critical step in the development process of therapeutic antibodies. The ability to generate desired antibodies without the need for additional optimization could significantly reduce development timelines.

[0209] Fig. 8A The binding affinities of the AI-generated zero-shot binders were plotted, and 71 designs were found to have comparable affinity to trastuzumab (<10 nM), with 3 designs having tighter binding. Figure 8B The edit distances of the designed variant binding affinities relative to trastuzumab are depicted. The edit distances range from 2 mutations (84.6% sequence identity) to 12 mutations (7.7% sequence identity), illustrating the novelty of the designs. Figure 8CHCDR3 signatures of 421 binding affinity trastuzumab variants are depicted. Greater diversity is observed in the center of the designed CDRs. The sequence logo below is the trastuzumab HCDR3 sequence. Fig.8D Pairwise edit distances between binders are depicted (minimum 1, maximum 15, median 8, mean 7.7, ±2.1 SD).

[0210] In addition to favorable affinity, the AI ​​model designs had high sequence diversity in both amino acid length and consistency. The validated binders had HCDR3s ranging from 11 to 15 amino acids (compared to trastuzumab, which is 13). Fig.8F The designed sequences are also different from the trastuzumab antibody, where the edit distance of the trastuzumab sequence is between 2 and 12 ( Figure 8B ). As the edit distance of the trastuzumab sequence increased, the average affinity decreased, but interestingly, the present technology found designs that still exhibited an affinity of less than 10nM at all edit distances. The present technology observed a design with an edit distance of 9 that exhibited higher affinity than the trastuzumab antibody. In addition, the present technology found that the center of the HCDR corresponding to the more diverse D germline genes had higher diversity compared to the less diverse flanking J and V germline genes ( Fig. 8A The designs were also different from each other, with an average edit distance of 7.7 ± 2.1 SD ( Fig.8D , Fig. 8E The inter-design diversity is noteworthy because it indicates that the binders generated by the model do not converge to shared sequence motifs, as is typically seen in traditional antibody screening methods such as phage display.

[0211] Designed binders display sequence novelty

[0212] Although the 421 designed binders have high sequence diversity, to ensure that the model generates novel sequences rather than simply reproducing training examples, the present technology compares the model output to the training set. This phenomenon has been observed in machine learning models, and past methods have been criticized for generating molecules that are similar to previously known molecules. The present technology calculates the minimum distance between the designed binders and all HCDR3s in the model training and validation sets and finds that the designed binders are different from those observed during training ( Fig. 9A The present technology then calculates the distance to all HCDR3 in the Structural Antibody Database (SAbDab), a database of antibody-antigen complexes, and finds that the binder is far away from all antibodies in the database in sequence space ( Fig. 9B ).

[0213] Fig. 9AThe minimum edit distance of the binders to the training data HCDR3 is depicted (minimum 2, maximum 8, median 5, mean 4.68 ± 1.34 SD). Fig. 9C The minimum edit distance of the binders to the OAS HCDR3 is depicted. 9.3% (38 out of 421) of the HCDR3 designs were included in the OAS. (Minimum 0, maximum 5, median 2, mean 1.91 ± 1.08 SD). Fig.9D Depicts the naturalness score of the designed binders relative to the control. De novo binders are those identified in our study. OAS refers to a randomly selected HCDR3 from OAS. The frequency baseline samples amino acids at each position based on the position frequency observed in the OAS of HCDR3. The phage display baseline is a HCDR3 set sampled from binding and non-binding antibodies from Liu et al. The scrambled OAS is a randomly arranged version of the OAS sequence set. The average naturalness score of the AI ​​design is significantly higher than that of the control population (p<10e-50), but on average, the naturalness score is lower than that of trastuzumab and sequences randomly sampled from OAS (p<10e-15). The red dotted line is the naturalness score of trastuzumab. Fig.9E The naturalness scores of the designed binders are plotted versus the edit distance to trastuzumab. The red dashed line is the naturalness score of trastuzumab.

[0214] The present technology examined the sequence similarity of the model output with sequences in the Observed Antibody Space (OAS), a database for immune repertoire sequencing studies. The present technology found that some model designs already existed in OAS, while other designs were unique, with a minimum edit distance between 1 and 5 ( Fig. 9B ). This indicates that the model is able to generate sequences that achieve high naturalness scores as well as sequences that are dissimilar to antibodies in the observed immune repertoire.

[0215] Zero-shot designs have high naturalness scores

[0216] Successful therapeutic lead antibody candidates in the drug development process typically have high affinity and are developable to have low immunogenicity. In previous work, a language model was described that assigns a score to antibody sequences indicating the likelihood of the sequence being found in a typical immune repertoire. This metric is referred to as naturalness. High naturalness scores are associated with favorable developability and immunogenicity development outcomes. Using the naturalness scoring model in our design, we find that our model can generate sequences with high naturalness scores and high affinity in a zero-shot manner, despite not being trained or sampled based on these qualities ( Fig.9E ). Many designs show naturalness scores

[0217] Fig.9DDepicted are naturalness measures of the binders designed in the present invention relative to their diversity with trastuzumab. Notably, Fig.9D Many of the designs depicted in are more natural than the wild type (depicted by the horizontal line), even though the wild type corresponds to an FDA-approved molecule. As with binding affinity, the naturalness of the binder decreases as the edit distance to trastuzumab increases. However, empirical testing using the technology of the present invention still identifies binders with high edit distances (5 to 9 mutations) that have a higher naturalness than trastuzumab than trastuzumab. Table S5 shows the p-values ​​for the average naturalness scores of different groups and the relevant statistical comparisons with de novo binders. These results highlight the possibility of zero-sample designs bypassing part of the traditional lead optimization process, thereby potentially saving time and resources in drug development.

[0218] 3D structural representation of the de novo designed HCDR3 and its comparison with the trastuzumab-HER2 complex

[0219] Table 1. Properties of different HCDR3 candidates selected for 3D structural modeling. HCDR3 candidates were selected based on affinity, length and edit distance to trastuzumab. RMSD values ​​on all main chain and side chain atoms were calculated from the alignment of HCDR3 residues. All other atoms were excluded from the calculation. The total mean of the hydrophilicity values ​​of the HCDR3 residues was calculated by averaging the hydrophilicity values ​​of each residue and dividing by the sequence length. 58

[0220]

[0221]

[0222] The technology of the present invention then predicts the binding mechanism of the HCDR3 variants designed from scratch by the present invention to better understand the structural basis of highly sequence-diverse variants. To this end, the technology of the present invention constructs a structural model of eight different HCDR3 candidates in Fab format for binding to HER2. These eight variants are selected based on their edit distance to trastuzumab HCDR3, length diversity (ranging from 12 to 15 amino acids) and affinity range, spanning three orders of magnitude (Table 1). The technology of the present invention uses trastuzumab Fab-HER2 complex (PDB: 1N8Z) as the starting template for structural modeling. The technology of the present invention runs locally restricted backbone geometry and side chain rotational isomer optimization, and then relaxes the Fab-HER2 complex to correct global conformational ambiguity, spatial collisions and suboptimal ring geometry. As a control, the experimental trastuzumab-Fab complex is optimized using the same scheme and used as a reference for comparing the final HCDR3 structural model. The technology of the present invention uses the lowest free energy posture of the HCDR3 model from scratch for structural analysis and comparison.

[0223] Despite sequence dissimilarity, the eight de novo Fab-HER2 structural models were generally similar to trastuzumab, with all-atom HCDR3 RMSDs ranging from 1.9 Å to Minimal structural rearrangements were observed in the unmodified regions of the heavy chain, the light chain and the epitope residues of the antigen ( Fig. 10A In selected cases, side chains making contacts with HCDR3 show slight rotational differences to account for the presence of longer loops or steric clashes from residues with larger side chains ( Fig. 10B In terms of the designed HCDR3 regions, alignment with the trastuzumab Fab-HER2 complex revealed a dynamic set of conformations adopted by each HCDR3 ( Fig. 10C ). The HCDR3 loop structures vary widely, with RMSDs ranging from 1.1 to 2.5 when aligned across all backbone and side chain atoms. (Table 2). Fig. 10C In spite of the sequence and length diversity, there are key residues that are conserved in space, corresponding to trastuzumab residues W107, G109, and Y113 (IMGT numbering scheme).

[0224] Table 2. Multi-step AI designed trastuzumab variant binders against HER and designed all three HCDRs. ED indicates the edit distance from HCDR to trastuzumab value. Note that the model occasionally recovers native trastuzumab HCDR1 and HCDR2. We present nine variants here and open-source the full set of 23 designs with accompanying sequence data.

[0225]

[0226]

[0227] Despite the de novo HCDR3 adopting different conformations, important positional similarities exist between all structures. A closer analysis of the spatial orientation of the side chain conformers revealed conservation of the same side chains at five discrete spatial positions, two of which correspond to IMGT residue positions R103 and Y117 in trastuzumab, which are highly conserved in most antibodies. However, physicochemical conservation exists in all structures corresponding to the spatial positions of trastuzumab IMGT residue numbers 109, 113, 117, which contribute to the paratope of the trastuzumab-Her2 complex. Although spatially conserved, these side chains originate from a variety of residue positions, which highlights the observed conformational flexibility that may be required to orient key paratope side chains to generate the same important protein-protein interactions with the HER2 epitope.

[0228] Although the overall binding region was identical, each designed HCDR3 exhibited a different binding pattern between the HCDR3 and the epitope. In most cases, novel interactions not observed in the trastuzumab-HER2 complex were formed between the designed HCDR3 and domain IV HER2 ( Fig. 10B , Fig. 10D ). These interactions are different and consist of novel hydrogen bonding interactions, non-polar interactions, aromatic interactions, and electrostatic interactions formed with each HCDR3 and Her2 epitope. In order to further decipher the determinants of binding, the present invention technology calculated the surface area buried by each HCDR3 variant Fab when binding to HER2. The surface area buried when the antibody Fab binds to Her2 is defined as the binding interface area between the paratope and the epitope (Table 2).

[0229] In several cases, the de novo HCDR3 variants showed a larger binding interface area than trastuzumab, which may indicate a new interaction with the HER2 epitope. Interestingly, no correlation was observed between the binding interface area and the binding affinity. The findings will indicate that hydrophobic contribution and surface area burial are not key determinants of binding. In addition, the specific contact formed between each designed HCDR3 and the epitope is crucial for the binding stability of the Fab-Her2 complex. In addition, the technology of the present invention calculates the total average value (GRAVY) of the hydrophilic value of each HCDR3 variant, which defines the overall hydrophobic properties summed on each residue and is compared with the binding affinity. The technology of the present invention observes that there is no correlation between affinity and hydrophobicity, which further confirms that the hydrophobic effect is not the main determinant of the HCDR3 binding designed from scratch. (Table 2). Combined, these results show that the binding affinity of the designed HCDR3 is inherent to the sequence design and is not driven by a common binding mechanism. The high dependence of the binding on the sequence attribute is consistent with the low probability of accidentally designing the binder.

[0230] Verify additional targets

[0231] The present invention technology then conducted preliminary studies to demonstrate the applicability of our method to a wider set of antigens. For these additional targets, the present invention technology did not pre-screen by ACE assay. Instead, the present invention technology sampled a small number of sequences and verified binding by SPR. The present invention technology first successfully designed HCDR3 variants of therapeutic ranibizumab, which binds to human vascular endothelial growth factor A (VEGF-A), such as Fig.11AAs shown in . By SPR measurement, the binder has an affinity of 48.2 nM compared to the sub-nanomolar binding of ranibizumab (0.37 nM). In addition, the designed HCDR3 is highly different from ranibizumab with an edit distance of 13, and is novel, with a minimum of 4 mutations separating it from any HCDR3 in OAS.

[0232] The present invention technology then designed a set of HCDR3 variants of casirumab that were conditional on binding to the omicron spike protein. Casirumab binds to multiple COVID spike protein variants, especially weakly to omicron. The present invention technology measured the affinity of casirumab for omicron by SPR at 240.0nM (Table S1). The present invention technology identified an AI-designed variant that binds with similar affinity to omicron at 179.7nM ( Fig. 11B , Table S1). Interestingly, the present technology observed no binding to other spike protein variants for our AI design, indicating the possibility of controllability of target specificity in homologous antigens. The designed variants have different HCDR3 sequences compared to casirumab, with an edit distance of 6 to any HCDR3 in OAS and a minimum edit distance of 2. In addition, the edit distance of HCDR3 to any HCDR3 in CoV-AbDab is at least 4. These binding designs to two additional antigens highlight the scalability of our zero-shot design approach and indicate the possibility of achieving selective antigen controllability using generative AI.

[0233] Table S1. Measured binding affinity of Casirumonab and AI-designed variants to COVID antigens. N / A indicates lack of binding. In this context, the model is indicated to generate HCDR3 variants targeting Omicron that fit within the Casirumonab framework. These results indicate that the model successfully designed HCDR3 variants that lost activity against COVID spike protein wild type / Beta / Delta but remained active against Omicron. These results are the first step in the controllability of antibody design for specific protein variants.

[0234] wild type Beta Delta Omicron Casirimab 8.3nM 470.0nM 4.2nM 240.0nM AI-designed binders N / A N / A N / A 179.7nM

[0235] Extending to multiple CDRs

[0236] Extending antibody design beyond HCDR3 allows for increased sequence diversity and controllability and represents the next step toward completely de novo design. To this end, we applied an alternative multi-step AI design method that is different from the zero-shot method we described to simultaneously generate variants of all heavy chain variable regions (HCDR1, HCDR2, HCDR3). The present invention reports a variety of binding designs for HER2 identified in a library of less than 500 multi-step design variants (Table 2). The present invention found that these binders were again different from the examples in the model training data and the antibodies in the SAbDab and OAS databases (Figures 9, Fig.9I ).

[0237] Exemplary Materials and Methods

[0238] Library Design

[0239] For the de novo design of HCDR3 trastuzumab variants, the technology of the present invention can start from the design of two high diversity libraries (HDL), represented as HDL1, which are composed of, for example, 223,046 and 217,308 designs, respectively. HDL1 can be screened using the ACE assay and dnACE assay described herein, and the results are used to design low diversity libraries (LDL), each LDL consisting of 1,000 or less sequences screened using SPR. HDL2 is screened using ACE, and the results are used again to design LDL for SPR screening. In total, 199 binders were confirmed from HDL1 based on ACE scores, revealing an estimated value of another 3765 binders in HDL2 (Table S5). The ACE assay cannot be used to screen variants of HCDR3 ranibizumab, HCDR3 casirui monoclonal antibody, or HCDR123 trastuzumab. Instead, small LDLs can be directly screened with SPR.

[0240] Table S4. Estimated number of binders based on the ACE "grid" in HDL2. In addition to the 224 confirmed binders, an additional 3,780 designs were estimated to bind to HDL2 based on the sequence distribution in the ACE grid. Note that the total number of identified designs (41,942) is less than the total designs (217,308). The difference corresponds to designs that were not sorted in ACE, indicating that they are unlikely to bind.

[0241] ACE Grid Binding rate in SPR Number of designs identified Expected number of binders 0 0.045 35,484 1,597 1 0.194 4,459 865 2 0.522 1,359 709 3 0.932 368 343 4 0.914 200 183 5 0.942 72 68 Aggregates 41,942 3,765

[0242] Naturalness score

[0243] The naturalness score for this study can be calculated using the pre-trained antibody language model discussed herein and introduced previously. The model is based on pseudo-complexity on the extended CDR (e.g., defined by the union of IMGT and Martin definitions) of the antibody heavy chain under the language model. The metric can predict desired antibody therapeutic properties, such as developability or lack of immunogenicity. The technology of the present invention can place variant HCDR3 in a wild-type scaffold by replacing wild-type HCDR3. In addition to calculating the naturalness of the de novo binder of the present invention, the technology of the present invention can also include several controls:

[0244] 1. OAS: Consists of 1,000 HCDR3s randomly sampled from OAS. The naturalness score was calculated by grafting the HCDR3s into the trastuzumab scaffold. Given that the naturalness model was pre-trained on OAS, we expected sequences from OAS to have high naturalness scores, so we considered this as a positive control.

[0245] 2. Frequency baseline: These are 1000 sequences generated by random sampling from the length-adjusted frequency distribution of the OAS. The present technique can calculate That is, the HCDR3 in OAS has a length The probability of , and then using independent factorization based on amino acid frequencies to calculate the probability of having length The probability of sampling a particular sequence is defined as:

[0246]

[0247] The technology of the present invention may include Multiple (e.g., 1000) sequences are sampled and then automatically generated

[0248] 3. Phage display baseline: These are 1000 HCDR3s randomly sampled from the first round of phage display panning (Liu et al.). Antibody heavy chain sequences were sampled and HCDR3s extracted. Sampled from Antibodies The sampled antibody collection consisted of non-binders and binders.

[0249] 4. Scrambled OAS: This consists of 1000 HCDR3 replacement versions in the OAS control. For each such HCDR3, the technology of the present invention may include replacing the corresponding sequence 5 times, calculating the naturalness using the replaced HCDR3, and reporting the average of the 5 replacements. In some aspects, the motivation for this calculation as a negative control is that replacing the protein sequence destroys positional information. Compared to the first OAS baseline, the lower naturalness score of this baseline means that the naturalness model is able to capture positional information and not only considers amino acid composition.

[0250] We can use the two-sample t test (H0:μ1=μ2,H a The naturalness scores of the de novo designs were compared to these controls and the wild-type naturalness was taken as the population mean (H0: μ1 = μ, H2) using a one-sample t-test. a :μ1! =μ) compared with wild type. Below is a table showing the average naturalness scores in the different groups and the p-values ​​for the relevant statistical comparisons with de novo binders.

[0251] Group Average naturalness p-value (compared to de novo) From the beginning 0.175 N / A wild type 0.235 <![CDATA[3.66·10 -95 ]]> OAS 0.194 <![CDATA[3.63·10 -16 ]]> Frequency Baseline 0.143 <![CDATA[4.67·10 -53 ]]> Phage display baseline 0.136 <![CDATA[1.11·10 -87 ]]> Out-of-order OAS 0.101 <![CDATA[8.05·10 -300 ]]>

[0252] Table 2.

[0253] Table 2 depicts the mean naturalness scores for the different groups and the p-values ​​for the use of trastuzumab stents.

[0254] Furthermore, in some aspects, the present invention technology previously noted the design diversity relative to OAS, which is the pre-training data used for the model used in this article to calculate naturalness. Taking away samples from the training set may result in a reduction in naturalness. In fact, if Fig. 9B As shown in , empirical evidence shows that the naturalness of designs tends to decrease as the distance from OAS increases. Of particular interest, the average naturalness of the designs of the invention contained in OAS can be higher than the average OAS control naturalness (p=9.45e-19) and higher than the naturalness of trastuzumab (p=1.31e-4). In fact, any of the designs of the invention that differ by 1 mutation in HCDR3 in OAS have significantly higher naturalness than the OAS control sequence (p=9.83e-5).

[0255] Exemplary Binder Data

[0256] consensus sequence

[0257] As discussed, the technology of the present invention includes using training data to train the model. For example, in some aspects, the technology of the present invention can be trained using the following binder data. Tables E1 to E4 below include binder data for the following: de novo trastuzumab HCDR3 variant HER2 binders, de novo ranibizumab HCDR3 variant VEGF-A binders, de novo casirui monoclonal antibody HCDR3 variant COVID-Omicron binders, de novo trastuzumab HCDR123 variant HER2 binders, respectively. The following consensus sequences are applicable to binder data.

[0258] HER2 binders

[0259] For HER2 binders, the following information is relevant but does not change for each binder:

[0260] Parental antibody: Trastuzumab

[0261] Antigen: HER2

[0262] Parent PDB: 1N8Z

[0263] Parental heavy chain sequence (SEQ ID NO:76) : EVQLVESGGGLVQPGGSLRLSCAASGFNIKDTYIHWVRQAPGKGLEWVARIYPTNGYTRYADSVKGRFTISADTSKNTAYLQMNSLRAEDTAVYYCSRWGGDGFYAMDY WGQGTLVTVSSASTKGPSVFPLAPSSKSTSGGTAALGCLVKDYFPEPVTVSWNSGALTSGVHTFPAVLQSSGLYSLSSVVTVPSSSLGTQTYICNVNHKPSNTKVDKKVEP

[0264] Parental HCDR1 (SEQ ID NO:54) :GFNIKDTY

[0265] Parental HCDR2 (SEQ ID NO:55) :IYPTNGYT

[0266] Parental HCDR3 (SEQ ID NO:20) :SRWGGDGFYAMDY

[0267] Parent light chain sequence (SEQ ID NO:77) :DIQMTQSPSSSLSASVGDRVTITCRASQDVNTAVAWYQQKPGKAPKLLIYSASFLYSGVPSRFSGSRSGTDFTLTISSLQPEDFATYYCQQHYTTPPTFGQGTKVEI KRTVAAPSVFIFPPSDEQLKSGTASVVCLLNNFYPREAKVQWKVDNALQSGNSQESVTEQDSKDSTYSLSSTLTLSKADYEKHKVYACEVTHQGLSSPVTKSFNRGEC

[0268] Antigen chain sequence (SEQ ID NO:78):TQVCTGTDMKLRLPASPETHLDMLRH LYQGCQVVQGNLELTYLPTNASLSFLQDIQEVQGYVLIAHNQVRQVPLQRLRIVRGTQLFEDNYALAVLDNGDPLNNTTPVTGASPGGLRELQLRSLTEILKGGVLIQRNPQLCYQDTILWKDIFHKNNQLALTLIDTNRSRACH PCSPMCKGSRCWGESSEDCQSLTRTVCAGGCARCKGPLPTDCCHEQCAAGCTGPKHSDCLACLHFNHSGICELHCPALVTYNTDTFESMPNPEGRYTFGASCVTACPYNYLSTDVGSCTLVCPLHNQEVTAEDGTQRCEKCSKPC ARVCYGLGMEHLREVRAVTSANIQEFAGCKKIFGSLAFLPESFDGDPASNTAPLQPEQLQVFETLEEITGYLYISAWPDSLPLDLSVFQNLQVIRGRILHNGAYSLTLQGLGISWLGLRSLRELGSGLALIHHNTHLCFVHTVPWD QLFRNPHQALLHTANRPEDECVGEGLACHQLCARGHCWGPGPTQCVNCSQFLRGQECVEECRVLQGLPREYVNARHCLPCHPECQPQNGSVTCFGPEADQCVACAHYKDPPFCVARCPSGVKPDLSYMPIWKFPDEEGACQPCPIN

[0269] COVID, VEGF, and HER2

[0270] The shared / unchanged information for the COVID, VEGF, and HER2 affinity tables is as follows:

[0271] COVID:

[0272] ●Parental antibody: Casirikimab

[0273] ●Parental antibody PDB: 6XDG

[0274] Antigens for which binding affinity data were collected by SPR: SARS-CoV2 spike protein RBD (wild type, Beta, Delta, Omicron)

[0275] Parental heavy chain sequence (SEQ ID NO: 79): QVQLVESGGGLVKPGGSLRLSCAA SGFTFSDYYMSWIRQAPGKGLEWVSYITYSGSTIYYADSVKGRFTISRDNAKSSLYLQMNSLRAEDTAVYYCARDRGTTMVPFDYWGQGTLVTVSSASTKGPSVFPLAPSSKSTSGGTAALGCLVKDYFPEPVTVSWNSGALTSGVHTFPAVLQSSGLYSLSSVVTVPSSSLGTQTYICNVNHKPSNTKVDKKVEPKSCDKT

[0276] ● Parental HCDR3 (SEQ ID NO: 51) : ARDRGTTMVPFDY

[0277] ● Parental light chain sequence (SEQ ID NO:80) : DIQMTQSPSSLSASVGDRVTITCQA SQDITNYLNWYQQKPGKAPKLLIYAASNLETGVPSRFSGSGSGTDFTFTISGLQPEDIATYYCQQYDNLPLTFGGGTKVEIKRTVAAPSVFIFPPSDEQLKSGTASVVCLLNNFYPREAKVQWKVDNALQSGNSQESVTEQDSKDSTYSLSSTLTLSKADYEKHKVYACEVTHQGLSSPVTKSFNRGEC

[0278] ● Antigen on which the model is based: SARS-CoV-2 Omicron spike protein RBD

[0279] ○ PDB: 7QO9

[0280] ○ Sequence (SEQ ID NO: 81)

[0281] VEGF:

[0282] ●Parental antibody: Ranibizumab

[0283] ●Parental antibody PDB: 1CZ8

[0284] ●Antigen: VEGF-A

[0285] Parental heavy chain sequence (SEQ ID NO:82) :EVQLVESGGGLVQPGGSLRLSCAA SGYDFTHYGMNWVRQAPGKGLEWVGWINTYTGEPTYAADFKRRFTFSLDTSKSTAYLQMNSLRAEDTAVYYCAKYPYYYGTSHWYFDVWGQGTLVTVSSASTKGPSVFPLAPSGTAALGCLVKDYFPEPVTVSWNSGALTSGVHTFPAVLQSSGLYSLSSVVTVPSSSLGTQTYICNVNHKPSNTKVDKKVEPK

[0286] Parental HCDR3 (SEQ ID NO: 49) :AKYPYYYGTSHWYFDV

[0287] Parental light chain sequence (SEQ ID NO:83) :DIQLTQSPSSSLSASVGDRVTITCSASQDISNYLNWYQQKPGKAPKVLIYFTSSLHSGVPSRFSGSGSGTDFTLTISSLQPEDFATYYCQQYSTVPWTFGQGTKVEI KRTVAAPSVFIFPPSDEQLKSGTASVVCLLNNFYPREAKVQWKVDNALQSGNSQESVTEQDSKDSTYSLSSTLTLSKADYEKHKVYACEVTHQGLSSPVTKSFNRGE

[0288] Antigen chain sequence (SEQ ID NO:84) :VVKFMDVYQRSYCHPIETLVDIFQEYPDEIEYIFKPSCVPLMRCGGCCNDEGLECVPTEESNITMQIMRIKPHQGQHIGEMSFLQHNKCECRPK

[0289] HER2: Trastuzumab binding affinity and HCDR3 (SEQ ID NO:20) : SRWGGDGFYAMDY, 1.940000e-09(KD(M)), -8.71(log(KD(M))), 1.94(KD(nM)).

[0290] Table E1. De novo trastuzumab HCDR3 variant HER2 binders

[0291]

[0292]

[0293]

[0294]

[0295]

[0296]

[0297]

[0298]

[0299]

[0300]

[0301]

[0302] Table E2. De novo ranibizumab HCDR3 variant VEGF-A binders

[0303] HCDR3 KD(M) log(KD(M)) KD(nM) SEQ ID NO: AKYPYYYGTSHWYFDV 3.70E-10 -9.43 0.37 49 ASGNFSYPDDYGMDW 4.82E-08 -7.32 48.2 50

[0304] Table E3. De novo Casirimab HCDR3 variants COVID-Omicron binders

[0305]

[0306] Table E4. De novo trastuzumab HCDR123 variant HER2 binders

[0307]

[0308]

[0309]

[0310]

[0311]

[0312]

[0313]

[0314] clone

[0315] In some aspects, antibody variants can be cloned and expressed in Fab form. For example, in order to generate ACE assays and SPR data sets, the technology of the present invention may include using ssDNA oligonucleotide pools (oligopool) (Twist Bioscience) to synthesize DNA variants of HCDR3, HCDR1 to HCDR3 across different libraries, and using eBlocks (IDT) to synthesize double-stranded DNA formats. For each variant, codons can be randomly selected from the two most common types of Escherichia coli (E.coli) B strains. The amplification of the ssDNA oligonucleotide pool can be performed by PCR according to the advice of TwistBioscience, except that Q5 high-fidelity DNA polymerase (NEB) can be used instead of KAPA polymerase. In short, 25 μL reactions can be composed of 1 × Q5 master mix, 0.3M each of forward and reverse primers, and 10ng oligonucleotide pools. The reaction can be initially denatured at 95° C. for 3 min, followed by 13 cycles of: denaturation at 95° C. for 20 sec; denaturation at 66° C. for 20 sec; denaturation at 72° C. for 15 sec; and a final extension of denaturation at 72° C. for 1 min. DNA amplification can be confirmed by agarose gel electrophoresis, and the amplified DNA can then be purified (DNA Clean and Concentrate Kit, Zymo Research).

[0316] To construct libraries for SPR validation of model designs in independent experiments, Integrated DNA Technologies (IDT) synthesized oligonucleotides spanning the appropriate CDRs and immediate upstream / downstream flanking nucleotides.

[0317] To generate a linearized vector, the present technology may include performing a two-step PCR to separate the present invention's proprietary plasmid vector carrying trastuzumab in Fab format into two fragments in a manner that provides approximately 25 nucleotides (nt) on the 5' and 3' ends of the amplified ssDNA oligonucleotide pool or 40 nt of cloning overlap on the 5' and 3' ends of IDT eBlocks. The vector linearization reaction may be digested with DpnI (New England Biolabs) and purified from a 0.8% agarose gel (Gel DNA Recovery Kit, Zymo Research) to eliminate parental vector carryover.

[0318] Cloning reactions can consist of 50 fmol of each purified vector fragment, 100 fmol of purified library (TwistBioscience) or 10 pmol of gBlock insert (IDT) and 1× final concentration of NEBuilder HiFi DNA Assembly (New England Biolabs). Reactions can be incubated at 50°C for two hours (TwistBioscience library) or 25 min (IDT library) and subsequently purified (DNA Clean and Concentrate Kit, Zymo Research).

[0319] For HDL, Transformax EPI300 (Lucigen) E. coli can be transformed by electroporation (BioRad MicroPulser) using purified assembly reactions and grown overnight at 30°C in 20 mL of Teknova LB broth containing 50 μg / mL kanamycin, shaking at 270 rpm for 18 h at 30°C and 80% humidity. Plasmids can be extracted (Plasmid Midi kit, Zymo Research) and submitted for QC sequencing. Absci's SoluProTM host strain can be transformed with 1 ng of QC plasmid and grown at 30°C in 20 mL of Teknova LB broth containing 50 μg / mL kanamycin, shaking at 270 rpm for 18 hours at 30°C and 80% humidity.

[0320] For LDL, Absci's SoluProTM host strain can be transformed with the purified assembly reaction and grown overnight at 30°C on agar plates containing 50g / ml kanamycin and 1% glucose. Colonies can be picked for QC analysis before culture induction. In some aspects, the aforementioned experimental parameters and processing procedures can be modified.

[0321] Exemplary QC Analysis

[0322] In the present technology, the quality of the high diversity variant library can be assessed by deep sequencing. Briefly, the library plasmid pool can be amplified by PCR across the HCDR3 region and sequenced using, for example, an Illumina MiSeq platform with 20% PhiX with 2×150nt reads. The PCR reaction can be 10nM primer concentration, Q5 2× master mix (NEB) diluted in H2O, and 1ng input DNA. The reaction can be initially denatured at 98°C for 3min; followed by 30 cycles of denaturation at 98°C for 10 seconds, denaturation at 59°C for 30 seconds, and denaturation at 72°C for 15 seconds; and finally extended to denaturation at 72°C for 2min. The mutation distribution, variant representation, library complexity, and expected sequence recovery of the sequencing results can be analyzed. Metrics can include the coefficient of variation of sequence representation, the read share of the top 1% most common sequence, and the percentage of the designed library sequence observed in the library. The quality of the low-diversity variant library can be assessed by rolling circle amplification of 24 colonies (Equiphi29, Thermo Fisher Scientific) and sequencing using the Illumina DNA prep Tagmentation kit (Illumina Inc.). Each colony can be analyzed for single nucleotide polymorphisms (SNPs), the presence of multiple variants, misassembly, and / or matching with the library sequence (Geneious Prime).

[0323] Exemplary Antibody Expression in SoluProTM E. coli B Strain

[0324] After transformation and 8 hours of recovery, HDL can be grown in 50 mL of Teknova LB broth containing 50 μg / mL kanamycin at 30°C and 80% humidity with shaking at 270 rpm for 24 hours. At the end of 24 hours, the preculture can be normalized to OD600 of 0.1 in induction medium (IBM) (4.5 g / L potassium dihydrogen phosphate, 13.8 g / L ammonium sulfate, 20.5 g / L yeast extract, 20.5 g / L glycerol, 1.95 g / L citric acid) containing inducers and supplements (250 μM arabinose, 50 μg / mL kanamycin, 8 mM magnesium sulfate, 1 mM propionate, 1X Korz trace metals) and grown in 500 mL baffled flasks at 26°C and 80% humidity with shaking at 270 rpm for 16 hours. At the end of 16 hours, 500 mL aliquots can be adjusted to 20% v / v glycerol and stored at -80°C.

[0325] After conversion and QC of LDL, single colonies can be picked into deep well plates containing 400 μL Teknova LB culture fluid and 50 μg / mL kanamycin and incubated for 24 hours at 1000 rpm shaking at 30°C and 80% humidity. At the end of 24 hours, 150 μL samples can be centrifuged (3300 g, 7 min) and the supernatant can be decanted from the pre-culture plate for sequence analysis. 80 μL pre-cultures are transferred to 400 μL IBM containing inducers and supplements as described above. Cultures can be grown for 16 hours at 26°C and 80% humidity with 270 rpm shaking. At the end of 16 hours, 150 μL samples can be taken out and centrifuged (3300 g, 7 min) into precipitates, the supernatant is decanted, and then stored at -80°C.

[0326] Activity-specific cell enrichment (ACE) assay

[0327] Cell preparation

[0328] In some aspects, high-throughput quantitative selection of antigen-specific Fab-expressing cells can be adapted from conventional methods. For staining, OD600=2 thawed glycerol stocks from induced cultures can be transferred to 0.7 ml matrix tubes, centrifuged (4000 g, 5 min), and the resulting pelleted cells are washed three times with PBS (pH 7.4, 1 mM EDTA). Before fixation, washed cells can be completely resuspended in 250 μL phosphate buffer (32 mM, pH 7.4) by adding 250 μL 0.6% paraformaldehyde and 0.04% glutaraldehyde in phosphate buffer (32 mM, pH 7.4) by pipetting. After incubation on ice for 40 min, the sample can be centrifuged (4000 g, 5 min), and the pellet can be washed three times with PBS (pH 7.4, 1 mM EDTA), resuspended in non-permeabilizing buffer (20 mM Tris, 50 mM glucose, 10 mM EDTA, 5 μg / mL lysosomes) and incubated on ice for 8 min. The fixed and permeabilized cells can then be centrifuged (4000 g, 5 min) and washed multiple times (e.g., three times) with staining buffer (Perkin Elmer AlphaLISA immunoassay buffer, 25 mM HEPES, 0.1% casein, 1 mg / mL dextran-500, 0.5% Triton X-100, 0.05% Kathon).

[0329] dyeing

[0330] Before library staining, the HER2 probe can be titrated against a reference strain to determine the 75% effective concentration (EC75). After cell preparation, the library can be resuspended in 500 μL staining buffer containing 100 nM His / Avi-tagged human HER2 (Acro Biosystems) conjugated to 50 nM streptavidin-AF647 (Invitrogen) or untagged human Her2 (AcroBiosystems) directly conjugated to AF647 via free amines. The library can be incubated with the probe overnight (16 h), rotated end-to-end at 4 ° C, centrifuged (4000 g, 5 min), and the precipitate washed 3 times with PBS. The pellet can be resuspended in 500 μL staining buffer containing 26.5 nM anti-κ light chain: BV421 (BioLegend) and incubated for 2 hours by end-over-end rotation at 4°C, then centrifuged (4000 g, 5 min), washed three times with PBS and resuspended in 200 μL PBS for sorting.

[0331] Sorting

[0332] The library can be sorted by one of two methods based on binding in a quantitative ACE assay or a binary version of the ACE assay as described herein. For either method, the library can be sorted on a FACSymphony S6 (BD Biosciences) instrument. Immediately before sorting, 50 μL of stained samples were transferred to a flow tube containing 1 mL PBS + 3 μL propidium iodide.

[0333] Aggregates, debris, and impermeable cells can be removed with single peaks, size, and PI+ parent gates, respectively. Cells can then be gated to include only cells with κ light chain expression (BV421). For quantitative ACE determination, collection gates can be drawn over the logarithmic range of the binding signal for sampling. The rightmost can be set to the brightest 0.1% of the collection library, and the leftmost gate can be set to collect at the low end of the positive binding signal based on the stained control strain. Four additional gates of equal width can be distributed between them, with each gate set to approximately half of the gMFI of the right gate.

[0334] For the binary version of the ACE assay, a total of three collection gates can be set to sample at the high end of the binding range (top 0.1%), the remaining positive binding signal events, and a negative gate containing no binding signal events. The library can be sorted simultaneously on two instruments with photomultiplier tubes to normalize fluorescence intensity, and the collected events are processed independently as technical replicates.

[0335] Next-generation sequencing

[0336] Sample preparation for sorting materials

[0337] For the previously described quantitative ACE assay and the binary version of the ACE assay, sample preparation for sequencing can follow the same protocol. Specifically, the cell material from the sorting gate can be collected in a diluted PBS mixture (VWR) in a 1.5 mL tube (Eppendorf). Samples of unsorted library material can be processed for QC and ACE metric calculations. The sorted samples can be centrifuged (3,800×g) and the tube volume can be normalized to 20 μl. Amplicons covering the HCDR3 region can be generated by PCR. The reaction can be input using a 10 nM primer concentration suspended in diluted PBS (VWR), Q5 2× master mix (NEB), and 20 μl of sorted cell material. The reaction can be initially denatured at 98°C for 3 min, followed by 30 cycles of the following: denaturation at 98°C for 10 seconds; denaturation at 59°C for 30 seconds; denaturation at 72°C for 15 seconds; and finally extended to denaturation at 72°C for 2 min. After amplification, samples can be enzymatically cleaned using ExoSAP-IT (Applied Biosystems). The resulting DNA samples can be quantified by Qubitluorometer (Invitrogen), prepared for sequencing using ThruPLEX DNA-Seq kit (TakaraBio), normalized and pooled. Pool size can be verified via Tapeststation 1000HS and can be sequenced on Illumina NextSeq1000P2 (2×150nt) with 20% PhiX.

[0338] Structural modeling for computer simulations

[0339] Rosetta's FastRelax application was run with flexible backbone and side-chain degrees of freedom. Prior to the relax procedure, we first idealized all candidate structures using Rosetta's Idealize protocol to avoid steric clashes and incorrect geometries. Relax was performed using the maximum number of rotamers by passing the -EX1, -EX2, -EX3, and -EX4 flags at initialization. The flag -packing:repack_only was also included to disable the design, as well as the flags -no_his_his_pairE and -multi_cool_annealer 10 to set the number of annealing iterations. To rank the conformations in FastRelax, Rosetta's ref2015 energy function was used. It is known that running relax on a structure will often shift the backbone by a few angstroms, so based on the conformations of the initial structure, additional terms were included containing harmonic distance constraints for all C / 3 atom pairs within a distance of 10 for any atom that is not part of or in a CDR loop. These constraints were given a weight of 10-4. The protocol was run ten times for each target and the decoy with the lowest energy in the HCDR3 loop region was finally selected.

[0340] Antibody

[0341] As described herein, the technology of the present invention may include a computer system, a computer-implemented method, and / or one or more computer-readable media that predicts or generates structural information of a biomolecule (such as an antibody). The term "antibody" as used herein refers to a complete antibody that interacts with an epitope on a target antigen (e.g., by binding, steric hindrance, stabilization / destabilization, spatial distribution). A naturally occurring "antibody" is a glycoprotein comprising at least two heavy (H) chains and two light (L) chains interconnected by disulfide bonds. Each heavy chain consists of a heavy chain variable region (abbreviated herein as VH) and a heavy chain constant region. The heavy chain constant region consists of three domains CH1, CH2, and CH3. Each light chain consists of a light chain variable region (abbreviated herein as VL) and a light chain constant region. The light chain constant region consists of one domain CL. The VH region and the VL region can be further subdivided into hypervariable regions, called complementary determining regions (CDRs), interspersed with more conservative regions, called framework regions (FRs). Each VH and VL is composed of three CDRs and four FRs, arranged in the following order from the amino terminus to the carboxyl terminus: FR1, CDR1, FR2, CDR2, FR3, CDR3, FR4. The variable regions of the heavy and light chains contain binding domains that interact with antigens. The constant region of an antibody can mediate the binding of an immunoglobulin to host tissues or factors, including various cells of the immune system (e.g., effector cells) and the first component (Clq) of the classical complement system. The term "antibody" includes, for example, monoclonal antibodies, human antibodies, humanized antibodies, camelized antibodies, chimeric antibodies, single-chain Fvs (scFv), disulfide-linked Fvs (sdFv), Fab fragments, F(ab') fragments, and anti-idiotypic (anti-Id) antibodies (including, for example, anti-Id antibodies of the antibodies of the present invention), and epitope-binding fragments of any of the above. The antibody can be of any isotype (e.g., IgG, IgE, IgM, IgD, IgA, and IgY), class (e.g., IgG1, IgG2, IgG3, IgG4, IgA1, and IgA2), or subclass. The antibody or epitope-binding fragment can be a multispecific molecule or a component of a multispecific molecule.

[0342] The light chain and heavy chain are divided into structural and functional homologous regions. The terms "constant" and "variable" are used functionally. In this regard, it should be understood that the variable domains of the light chain (VL) and heavy chain (VH) portions determine antigen recognition and specificity. In contrast, the constant domains of the light chain (CL) and heavy chain (CH1, CH2 or CH3) confer important biological properties, such as secretion, transplacental mobility, Fc receptor binding, complement fixation, etc. By convention, the numbering of the constant region domain increases as it becomes more distant from the antigen binding site or the amino terminus of the antibody. The N terminus is the variable region and the C terminus is the constant region; the CH3 domain and the CL domain actually contain the carboxyl termini of the heavy chain and the light chain, respectively.

[0343] As used herein, the phrase "antibody fragment" refers to one or more portions of an antibody that retain the ability to specifically interact with a target epitope (e.g., by binding, steric hindrance, stabilization / destabilization, spatial distribution). Examples of binding fragments include, but are not limited to: a Fab fragment, a monovalent fragment consisting of a VL domain, a VH domain, a CL domain, and a CH1 domain; a F(ab)2 fragment, a bivalent fragment comprising two Fab fragments connected by a disulfide bridge at the hinge region; a Fd fragment consisting of a VH domain and a CH1 domain; a Fv fragment consisting of a VL domain and a VH domain of a single antibody arm; a dAb fragment (Ward et al., (1989) Nature 341: 544-546), which consists of a VH domain; and an isolated complementary determining region (CDR). In addition, although the two domains VL and VH of the Fv fragment are encoded by separate genes, they can be connected by synthetic linkers using recombinant methods to enable them to be made into a single protein chain, in which the VL region and the VH region are paired to form a monovalent molecule (called single-chain Fv (scFv); see, for example, Bird et al. (1988) Science 242: 423-426; and Huston et al. (1988) Proc. Natl. Acad. Sci. 85: 5879-5883). Such single-chain antibodies are also intended to be encompassed within the term "antibody fragment". These antibody fragments are obtained using conventional techniques known to those skilled in the art, and the utility of the fragments is screened in the same manner as intact antibodies.

[0344] As described herein, antibodies may include biologically active derivatives or variants or fragments. As used herein, "biologically active derivatives" or "biologically active variants" include any derivative or variant of an antibody that has substantially the same function and / or biological properties as the antibody (e.g., WT antibody), such as binding properties, and / or the same structural basis, such as a peptide backbone or basic polymeric unit, including a framework region.

[0345] An "analog," such as a "variant" or "derivative" is an antibody that is substantially similar in structure to a naturally occurring antibody or WT antibody or another reference antibody and has the same biological activity (although to a different extent in some cases), as will be understood by those skilled in the art. For example, an antibody variant refers to an antibody that shares a substantially similar structure with a reference antibody and has the same biological activity. A variant or analog has a different amino acid sequence composition than the reference antibody from which the analog was derived, based on one or more mutations involving (i) deletion of one or more amino acid residues at one or more termini of the antibody and / or one or more internal regions (e.g., fragments) of the antibody sequence, (ii) insertion or addition of one or more amino acids at one or more termini of the antibody (usually "additions" or "fusions"), and / or one or more internal regions (usually "insertions") of the antibody sequence, or (iii) substitution of one or more amino acids for other amino acids in the antibody sequence. For example, a "derivative" is a type of analog and refers to an antibody that shares the same or substantially similar structure with a reference antibody that has been modified (e.g., chemically modified).

[0346] In some aspects, variants or sequence variants are mutants in which 1, 2, 3, 4, 5, 6 or more amino acids within one or more CDRs are mutated relative to a reference antibody. In some aspects, a CDR on a light chain, a heavy chain, or both a heavy chain and a light chain is mutated. In some aspects, one or more framework amino acid residues are mutated relative to a reference antibody.

[0347] In substitution variants, one or more amino acid residues of an antibody (e.g., in a CDR region) are removed and replaced with an alternative residue. In one aspect, the substitution is conservative in nature and this type of conservative substitution is well known in the art. Alternatively, the disclosure encompasses substitutions that are also non-conservative. Exemplary conservative substitutions are described in Lehninger, [Biochemistry, 2nd Edition; Worth Publishers, Inc., New York (1975), pp. 71-77].

[0348] The antibodies considered herein include full-length antibodies, biologically active subunits or fragments of full-length antibodies, and biologically active derivatives and variants of any of these forms of therapeutic proteins. Therefore, antibodies include those (1) having an amino acid sequence having an amino acid sequence identity of greater than about 60%, about 65%, about 70%, about 75%, about 80%, about 85%, about 90%, about 91%, about 92%, about 93%, about 94%, about 95%, about 96%, about 97%, about 98%, or about 99% or more with a reference antibody (e.g., encoded by a reference nucleic acid or amino acid sequence described herein) in a region of at least about 25, about 50, about 100, about 200, about 300, about 400 or more amino acids. According to the present disclosure, the term "recombinant protein" or "recombinant antibody" includes any protein obtained by recombinant DNA technology. In some aspects, the term encompasses antibodies as described herein.

[0349] In some embodiments, the antibodies or antibody variants described herein are expressed from one or more expression constructs and / or expressed in a cell or cell line described herein.

[0350] Exemplary wild-type or reference antibodies include commercially available or other known antibodies, including therapeutic monoclonal antibodies. Reference antibodies according to the present disclosure may include any antibody now known or later developed, including antibodies that are not clinically and / or commercially available.

[0351] Cells and expression constructs

[0352] cell

[0353] In some embodiments, antibodies of the disclosure, including wild-type (WT) antibodies and variant antibodies, are produced in cells. Cells comprising one or more expression constructs described herein are contemplated in various embodiments of the disclosure.

[0354] Prokaryotic Host Cells. In some embodiments of the present disclosure, expression constructs designed to express gene products, including the fusion proteins described herein, are provided in a host cell, such as a prokaryotic host cell. Prokaryotic host cells can include archaea (such as Haloferax volcanii, Sulfolobus solfataricus), gram-positive bacteria (such as Bacillus subtilis, Bacillus licheniformis, Brevibacillus choshinensis, Lactobacillus brevis, Lactobacillus buchneri, Lactococcus lactis, and Streptomyces lividans), or gram-negative bacteria, including Alphaproteobacteria (Agrobacterium tumefaciens, Caulobacter crescentus, Rhodobacter sphaeroides, and Streptomyces lividans). sphaeroides and Sinorhizobium meliloti), Betaproteobacteria (Alcaligenes eutrophus), and Gammaproteobacteria (Acinetobacter calcoaceticus, Azotobacter vinelandii, Escherichia coli, Pseudomonas aeruginosa, and Pseudomonas putida). Preferred host cells include gammaproteobacteria of the Enterobacteriaceae family, such as Enterobacter, Erwinia, Escherichia (including Escherichia coli), Klebsiella, Proteus, Salmonella (including Salmonella typhimurium), Serratia (including Serratia marcescans), and Shigella.

[0355] Eukaryotic Host Cells. Many additional types of host cells can be used in the expression systems of the present disclosure, including eukaryotic cells, such as yeast (Candida shehatae, Kluyveromyces lactis, Kluyveromyces fragilis, other Kluyveromyces, Pichia pastoris, Saccharomyces cerevisiae, Saccharomyces pastorianus, also known as Saccharomyces carlsbergensis, Schizosaccharomyces pombe, Dekkera / Brettanomyces species, and Yarrowia lipolyticd); other fungi (Aspergillus nidulans, Aspergillus niger, and Aspergillus niger); niger), Neurosporacrassa, Penicillium, Tolypocladium, Trichoderma reesei); insect cell lines (Drosophila melanogaster Schneider 2 cells and Spodoptera frugiperda Sf9 cells); and mammalian cell lines, including immortalized cell lines (Chinese hamster ovary (CHO) cells, HeLa cells, baby hamster kidney (BHK) cells, monkey kidney cells (COS), human embryonic kidney (HEK, 293 or HEK-293) cells) and human hepatocellular carcinoma cells (Hep G2). The above host cells can be obtained from the American Type Culture Collection.

[0356] As described in WO / 2017 / 106583, which is incorporated herein by reference in its entirety, the production of gene products such as therapeutic proteins on a commercial scale and in a soluble form is addressed by providing suitable host cells that can grow at high cell densities in fermentation cultures and can produce soluble gene products in an oxidative host cytoplasm by highly controlled inducible gene expression. The host cells of the present disclosure having these qualities are produced by combining some or all of the following features. (1) The host cell is genetically modified to have an oxidative cytoplasm by increasing the expression or function of an oxidative polypeptide in the cytoplasm, and / or by reducing the expression or function of a reducing polypeptide in the cytoplasm. Specific examples of such genetic changes are provided herein. Optionally, the host cell can also be genetically modified to express chaperones and / or cofactors and / or glycosylated polypeptide gene products that contribute to the production of the desired gene product. (2) The host cell comprises one or more expression constructs designed to express one or more gene products of interest; in certain embodiments, at least one expression construct comprises an inducible promoter and a polynucleotide encoding a gene product to be expressed from the inducible promoter. (3) The host cell contains additional genetic modifications designed to improve certain aspects of the expression of the gene product of the expression construct. In specific embodiments, the host cell (A) has an alteration in the gene function of at least one gene encoding a transporter protein for an inducer of at least one inducible promoter, and as another example, wherein the gene encoding the transporter protein is selected from the group consisting of araE, araE, araG, araH, rhaT, xylF, xylG and xylH, or particularly araE, or wherein the alteration in gene function is more particularly expression of araE from a constitutive promoter; and / or (B) the gene function level of at least one gene encoding a protein that metabolizes the inducer of at least one inducible promoter is reduced, and as a further example An example of an embodiment, wherein the gene encoding the protein that metabolizes at least one inducer of the inducible promoter is selected from the group consisting of araA, araB, araD, prpB, prpD, rhaA, rhaB, rhaD, xylA and xylB; and / or (C) the gene function level of at least one gene encoding a protein involved in the biosynthesis of the inducer of at least one inducible promoter is reduced, and the gene in another embodiment is selected from the group consisting of scpA / sbm, argK / ygfD, scpB / ygfG, scpC / ygfH, rmlA, rmlB, rmlC and rmlD.

[0357] Host cells with oxidative cytoplasm. The expression systems of the present disclosure are designed to express gene products; in certain embodiments of the present disclosure, the gene products are expressed in host cells. Examples of host cells are provided that allow for efficient and cost-effective expression of gene products (including components of multimeric products). In addition to isolated cells in culture, host cells may also include cells that are part of a multicellular organism, or cells that grow in different organisms or biological systems. In certain embodiments of the present disclosure, the host cell is a microbial cell, such as a yeast (Saccharomyces, Schizosaccharomyces, etc.) or a bacterial cell, or a Gram-positive bacterium or a Gram-negative bacterium, or Escherichia coli, or an Escherichia coli B strain, or an Escherichia coli (B strain) EB0001 cell (also known as an Escherichia coli ASE (DGH) cell), or an Escherichia coli (B strain) EB0002 cell. In growth experiments of Escherichia coli host cells with oxidative cytoplasm, in particular Escherichia coli B strain Express (NEB Catalog No. C3028H) and T7Express (NEB catalog number C3029H) and Escherichia coli K strain T7 (NEB Catalog No. C3026H), these E. coli B strains with an oxidative cytoplasm are able to grow to much higher cell densities than the closest corresponding E. coli K strain (WO / 2017 / 106583).

[0358] Changes in host cell gene function. Certain changes may be made to the gene function of a host cell comprising an inducible expression construct to facilitate effective and uniform induction of the host cell population by the inducer. Preferably, as measured by the method of Khlebnikov et al. described in Example 9 of WO / 2017 / 106583, the combination of expression constructs, host cell genotypes and induction conditions results in at least 75% (more preferably at least 85%, most preferably at least 95%) of cells in the culture expressing gene products from each induced promoter. For host cells other than E. coli, these changes may involve the function of genes that are structurally similar to E. coli genes, or genes that perform functions similar to E. coli genes in host cells. Changes to host cell gene function include eliminating or reducing gene function by completely deleting the gene protein coding sequence, or deleting a sufficiently large portion of the gene, inserting the sequence into the gene, or otherwise changing the gene sequence so that a functional gene product with a reduced level of production is produced by the gene. The change of host cell gene function also includes enhancing gene function by, for example, changing natural promoter to produce stronger promoter to instruct higher level of transcription of gene, or introducing missense mutation in protein coding sequence, thereby leading to more active gene product. The change of host cell gene function also includes increasing gene function by: for example, changing natural promoter to produce stronger promoter that instructs higher transcription level of described gene, or introducing missense mutation in protein coding sequence, thereby producing more active gene product. The change of host cell gene function includes changing gene function in any way, including, for example, changing natural inducible promoter to produce constitutively activated promoter. In addition to the expression of the change of gene function of inducer transport and metabolism and / or the change of chaperone protein as described herein about inducible promoter, it is also possible to change the reduction-oxidation environment of host cell.

[0359] Host cell reduction-oxidation environment. In bacterial cells such as E. coli, proteins requiring disulfide bonds are usually exported to the periplasm, where disulfide bond formation and isomerization are catalyzed by the Dsb system comprising DsbABCD and DsbG. Increased expression of cysteine ​​oxidase DsbA, disulfide isomerase DsbC, or a combination of Dsb proteins that are usually all transported to the periplasm has been used in the expression of heterologous proteins requiring disulfide bonds (Makino et al., Microb Cell Fact 2011 May 14; 10:32). It is also possible to express cytoplasmic forms of these Dsb proteins, such as cytoplasmic forms of DsbA and / or DsbC ('cDsbA' or 'cDsbC'), which lack a signal peptide and are therefore not transported to the periplasm. Cytoplasmic Dsb proteins (such as cDsbA and / or cDsbC) can be used to make the host cell's cytoplasm more oxidative and thus more conducive to disulfide bond formation in heterologous proteins produced in the cytoplasm. The host cell's cytoplasm can also be made less reductive, thereby more oxidative, by directly altering the thioredoxin and glutaredoxin / glutathione enzyme systems: mutants deficient in glutathione reductase (gor) or glutathione synthetase (gshB) together with thioredoxin reductase (trxB) result in cytoplasmic oxidation. These strains are unable to reduce ribonucleotides and therefore are unable to grow in the absence of exogenous reducing agents such as dithiothreitol (DTT). Suppressor mutations in the gene ahpC encoding the peroxiredoxin AhpC (such as ahpC* and ahpCA, Lobstein et al., MicrobCell Fact 2012 May 8;11:56; doi:10.1186 / 1475-2859-11-56) convert it to a disulfide reductase, thereby generating reduced glutathione, allowing electrons to be channeled to ribonucleotide reductase, and enabling cells deficient in gor and trxB or deficient in gshB and trxB to grow in the absence of DTT. Different classes of mutant forms of AhpC allow strains deficient in gamma-glutamylcysteine ​​synthetase (gshA) activity and deficient in trxB to grow in the absence of DTT; these include AhpC V164G, AhpC S71F, AhpC E173 / S71F, AhpC E171Ter, and AhpC dupl62-169 (Faulkner et al., Proc Natl Acad Sci USA 2008 May 6;105(18):6735-6740, Epub 2008 May 2). In such strains with an oxidative cytoplasm, exposed protein cysteines are easily oxidized in a process catalyzed by thioredoxin, thereby reversing their physiological function and leading to the formation of disulfide bonds.Other proteins that may help reduce the effects of oxidative stress in host cells with oxidative cytoplasm are the HPI (hydroperoxidase I) catalase-peroxidase encoded by E. coli katG and the HPII (hydroperoxidase II) catalase-peroxidase encoded by E. coli katE, which disproportionately convert peroxide to water and O2 (Farr and Kogoma, Microbiol Rev. 1991 Dec; 55(4): 561-585; Review). Increasing the amount of KatG and / or KatE protein in host cells, either by inducing co-expression or by increasing constitutive expression levels, is an aspect of some embodiments of the present disclosure.

[0360] Another modification that can be made to the host cell is to express the sulfhydryl oxidase Ervlp from the inner membrane space of yeast mitochondria in the host cytoplasm, which has been shown to increase the production of various complex disulfide-binding proteins of eukaryotic origin in the cytoplasm of E. coli, even in the absence of mutations in gor or trxB (Nguyen et al., Microb Cell Fact 2011 Jan 7;10:1).

[0361] The host cell comprising the expression construct preferably also expresses cDsbA and / or cDsbC and / or Ervlp; lacks trxB gene function; also lacks gor, gshB or gshA gene function; optionally has increased levels of katG and / or katE gene function; and expresses an appropriate mutant form of AhpC such that the host cell can grow in the absence of DTT.

[0362] Chaperone. In some embodiments, the desired gene product is co-expressed with other gene products (such as chaperones that are conducive to the production of the desired gene product). Chaperones are proteins that assist the non-covalent folding or unfolding and / or assembly or decomposition of other gene products, but do not appear in the structure when the resulting monomer or polymer gene product structure performs its normal biological function (completing the folding and / or assembly process). Chaperones can be expressed from an inducible promoter or constitutive promoter in the expression construct, or can be expressed from the host cell chromosome; preferably, the expression of chaperones in the host cell is at a sufficiently high level to produce a co-expressed gene product that is correctly folded and / or assembled into the desired product. Examples of chaperones present in Escherichia coli host cells are folding factors DnaK / DnaJ / GrpE, DsbC / DsbG, GroEL / GroES, IbpA / IbpB, Skp, Tig (trigger factor) and FkpA, which have been used to prevent protein aggregation in cytoplasmic or periplasmic proteins. DnaK / DnaJ / GrpE, GroEL / GroES and ClpB can act synergistically to assist protein folding, so the combined expression of these chaperones has been shown to be beneficial for protein expression (Makino et al., Microb Cell Fact 2011 May 14; 10:32). When expressing eukaryotic proteins in prokaryotic host cells, eukaryotic chaperones, such as protein disulfide isomerase (PDI) from the same or related eukaryotic species, are co-expressed or inducibly co-expressed with the desired gene product in certain embodiments of the present disclosure.

[0363] One chaperone protein that can be expressed in a host cell is a protein disulfide isomerase from Humicola insolens, a soil mycelium (soft rot fungus). The amino acid sequence of Humicola insolens PDI is shown in SEQ ID NO: 1 of WO / 2017 / 106583; it lacks the signal peptide of the native protein, so that it remains in the cytoplasm of the host cell. The nucleotide sequence encoding PDI is optimized for expression in Escherichia coli; SEQ ID NO: 2 of WO2017 / 106583 contains a GCTAGC Nhel restriction site at its 5' end, an AGGAGG ribosome binding site at nucleotides 7 to 12, a PDI coding sequence at nucleotides 21 to 1478, and a GTCGACSail restriction site at its 3' end. The nucleotide sequence of SEQ ID NO: 2 is designed to be inserted immediately downstream of a promoter (such as an inducible promoter). The NheI and SalI restriction sites in SEQ ID NO:2 can be used to insert it into a vector multiple cloning site, such as the multiple cloning site of the pSOL expression vector (SEQ ID NO:3 of WO / 2017 / 106583), described in published U.S. patent application US2015353940A1, the entire contents of which are incorporated herein by reference. Other PDI polypeptides can also be expressed in host cells, including PDI polypeptides from various species (Saccharomyces cerevisiae (UniProtKB PI 7967), Homo sapiens (UniProtKB P07237), Mus musculus (UniProtKB P09103), Caenorhabditis elegans (UniProtKB Q 17770 and Q17967), Arabdopsis thaliana (UniProtKB 048773, Q9XI01, Q9S G3, Q9LJU2, Q9MAU6, Q94F09 and Q9T042), Aspergillus niger (UniProtKB Q12730)) and modified forms of such PDI polypeptides. In certain embodiments of the present disclosure, the PDI polypeptide expressed in the host cell of the present disclosure has at least 70% or 80% or 90% or 95% amino acid sequence identity over at least 50% (or at least 60%, or at least 70%, or at least 80%, or at least 90%) of the length of SEQ ID NO: 1 of WO / 2017 / 106583, wherein the amino acid sequence identity is determined according to Example 10 of WO / 2017 / 106583.

[0364] Cellular transport of cofactors. When using the expression systems of the present disclosure to produce enzymes that require cofactors to function, it is helpful to use host cells that can synthesize the cofactors from available precursors or take up the cofactors from the environment. Common cofactors include ATP, coenzyme A, flavin adenine dinucleotide (FAD), NAD+ / NADH, and heme. Polynucleotides encoding cofactor transport polypeptides and / or cofactor synthesis polypeptides can be introduced into host cells, and such polypeptides can be constitutively expressed or inducibly co-expressed with gene products produced by the methods of the present disclosure.

[0365] Glycosylation of polypeptide gene products. A host cell can be altered in its ability to glycosylate polypeptides. For example, a eukaryotic host cell may have eliminated or reduced gene function in a glycosyltransferase and / or oligosaccharyltransferase gene, thereby impairing normal eukaryotic glycosylation of a polypeptide to form a glycoprotein. Prokaryotic host cells, such as E. coli, which do not normally glycosylate polypeptides, can be altered to express a set of eukaryotic and prokaryotic genes that provide glycosylation functions (DeLisa et al., WO2009089154A2, July 16, 2009).

[0366] Available host cell strains with altered gene function. In order to generate preferred strains of host cells to be used in the expression systems and methods of the present disclosure, it is useful to start with a strain that already contains the desired gene alteration (Table A; WO / 2017 / 106583).

[0367] Table A. Exemplary host cell lines

[0368]

[0369] Expression constructs

[0370] In some embodiments of the present disclosure, it is contemplated that an inducible promoter is used with an expression construct. Exemplary promoters are described herein, and are also described in WO / 2016 / 205570, which is incorporated herein by reference in its entirety. As described herein, cells comprising one or more expression constructs may optionally include one or more inducible promoters to express antibodies of the present disclosure, including wild-type antibodies and variant antibodies.

[0371] Expression construct. Expression construct is a polynucleotide designed to express one or more antibodies, and is therefore not a naturally occurring molecule. Expression construct can be integrated into the host cell chromosome, or maintained in the host cell as a polynucleotide molecule (such as a plasmid or an artificial chromosome) that replicates independently of the host cell chromosome. An example of an expression construct is a polynucleotide produced by inserting one or more polynucleotide sequences into the host cell chromosome, wherein the inserted polynucleotide sequence changes the expression of the chromosome coding sequence. An expression vector is a plasmid expression construct specifically for expressing one or more gene products (such as one or more antibodies). One or more expression constructs can be integrated into the host cell chromosome or maintained on an extrachromosomal polynucleotide (such as a plasmid or an artificial chromosome). The following is a description of a specific type of polynucleotide sequence in an expression construct that can be used for antibody expression or co-expression.

[0372] Origin of replication. The expression construct must include an origin of replication, also referred to as a replicon, so as to be maintained in a host cell as an independently replicated polynucleotide. Different replicons using the same replication mechanism cannot be maintained together in a single host cell by repeated cell division. Therefore, according to the origin of replication contained in the plasmid, the plasmid can be divided into incompatibility groups, as shown in Table 2 of WO / 2016 / 205570. The origin of replication for the expression construct can be selected based on incompatibility groups, copy number and / or host range and other criteria. As described above, if two or more different expression constructs are used in the same host cell to co-express a variety of antibodies or antibody components (e.g., heavy chains and light chains, including fragments, as described herein), in one embodiment, different expression constructs contain origins of replication from different incompatibility groups: for example, a pMB1 replicon in one expression construct and a p15A replicon in another expression construct. The average copy number of the expression construct in the cell relative to the number of host chromosome molecules is determined by the origin of replication contained in the expression construct. The copy number can range from a few copies per cell to hundreds of copies (Table 2 of WO / 2016 / 205570). In one embodiment of the present disclosure, different expression constructs are used, which include inducible promoters activated by the same inducer but with different replication origins. By selecting a replication origin that maintains each different expression construct in a certain approximate copy number in the cell, it is possible to regulate the overall production level of an antibody component or fragment (e.g., a heavy chain or light chain) expressed from one expression construct relative to another antibody component or fragment (e.g., a heavy chain or light chain) expressed from a different expression construct. As an example, in order to co-express subunits A and B of a multimeric protein (including, for example, a heavy chain and a light chain), an expression construct is produced that includes a colE1 replicon, an am promoter, and a coding sequence A of subunit A expressed from the am promoter: 'colE1-Para-A.

[0373] Another expression construct is generated, which comprises the p15A replicon, the am promoter and the coding sequence of subunit B: 'p15A-Para-B'. These two expression constructs can be maintained together in the same host cell, and the expression of subunits A and B is induced by adding an inducer, arabinose, to the growth medium. If it is necessary to significantly increase the expression level of subunit A relative to the expression level of subunit B so that the stoichiometric ratio of the expression amounts of the two subunits is closer to the desired ratio, for example, a new expression construct for subunit A can be generated, which has a modified pMB 1 replicon as found in the replication origin of the pUC9 plasmid ('pUC9ori'): pUC9ori-Para-A. The expression of subunit A from a high copy number expression construct such as pUC9ori-Para-A should increase the amount of subunit A produced relative to the expression of subunit B from p15A-Para-B. In a similar manner, the use of an origin of replication that maintains the expression construct at a lower copy number, such as pS001 (WO2016 / 205570), can reduce the overall level of gene products expressed from the construct. The selection of an origin of replication can also determine which host cells can maintain an expression construct comprising the replicon. For example, expression constructs containing the colE1 origin of replication have a relatively narrow range of viable hosts, namely species within the Enterobacteriaceae family, whereas expression constructs containing the RK2 replicon can be maintained in Escherichia coli, Pseudomonas aeruginosa, Pseudomonas sp., Azotobacter vinelandii, and Alcaligenes eutrophus, and if the expression construct contains the RK2 replicon and some regulatory genes from the RK2 plasmid, the expression construct can be maintained in diverse host cells, such as Sinorhizobium meliloti, Agrobacterium tumefaciens, Caulobacter crescentus, Acinetobacter calcoaceticus, and Rhodobacter sphaeroides (Kiies and Stahl, Microbiol Rev 1989 Dec; 53(4):491-516).

[0374] Similar considerations can be used to generate expression constructs for inducible expression or co-expression in eukaryotic cells. For example, the 2-micron circular plasmid of Saccharomyces cerevisiae is compatible with plasmids from other yeast strains, such as pSR1 (ATCC Accession Nos. 48233 and 66069; Araki et al., J Mol Biol 1985 Mar. 20; 182(2): 191-203) and pKD1 (ATCC Accession No. 37519; Chen et al., Nucleic Acids Res 1986 Jun. 11; 14(11): 4471-4481).

[0375] Selectable marker. Expression constructs usually include a selection gene, also referred to as a selectable marker, which encodes a protein necessary for survival or growth of host cells in a selective culture medium. Host cells that do not contain an expression construct comprising a selection gene will not survive in the culture medium. Typical proteins encoded by selection genes confer resistance to antibiotics or other toxins, or supplement the auxotrophic defects of host cells. An example of a selection scheme utilizes drugs (such as antibiotics) to prevent the growth of host cells. Those cells containing an expression construct comprising a selective marker produce proteins that confer drug resistance and survive in the selection scheme. Some examples of antibiotics commonly used in selecting selective markers (and abbreviations representing genes providing antibiotic resistance phenotypes) are: ampicillin (ampicillin) (AmpR), chloramphenicol (chloramphenicol) (CmlR or CmR), kanamycin (kanamycin) (KanR), spectinomycin (spectinomycin) (SpcR), streptomycin (streptomycin) (StrR) and tetracycline (tetracycline) (TetR). Many plasmids in Table 2 of WO / 2016 / 205570 contain selective markers, such as pBR322 (AmpR, TetR); pMOB45 (CmR, TetR); pACYClW (AmpR, KanR); and pGBMl (SpcR, StrR). The native promoter region of the selection gene is usually included as part of the selective marker portion of the expression construct together with the coding sequence of its gene product. Alternatively, the coding sequence of the selection gene can be expressed from a constitutive promoter.

[0376] In various aspects, suitable selectable markers include, but are not limited to, neomycin phosphotransferase (nptII), hygromycin phosphotransferase (hpt), dihydrofolate reductase (dhfr), zeocin, phleomycin, bleomycin resistance gene (ble), gentamycin acetyltransferase, streptomycin phosphotransferase, mutant forms of acetolactate synthase (als), bromoxynil nitrilase, phosphinothricin acetyltransferase (bar), enolpyruvylshikimate-3-phosphate (EPSP) synthase (aro A), muscle-specific tyrosine kinase receptor molecule (MuSK-R), copper-zinc superoxide dismutase (sod1), metallothionein (cup1, MT1), β-lactamase (BLA), puromycin N-acetyl-transferase (pac), blasticidin acetyltransferase (bls), blasticidin deaminase (bsr), histidinol dehydrogenase (HDH), N-succinyl-5-aminoimidazole-4-carboxamide ribonucleic acid (SAICAR) synthase (ade1), argininosuccinate lyase (arg4), β-isopropylmalate dehydrogenase (leu2), invertase (suc2), orotidine-5'-phosphate (OMP) decarboxylase (ura3), and orthologs of any of the foregoing.

[0377] Inducible promoters. As described herein, there are several different inducible promoters that can be included in expression constructs as part of the inducible co-expression system of the present disclosure. Preferred inducible promoters share at least 80% polynucleotide sequence identity (more preferably at least 90% identity, and most preferably at least 95% identity) with at least 30 (more preferably at least 40, and most preferably at least 50) consecutive bases of the promoter polynucleotide sequence defined in Table 1 of WO / 2016 / 205570 with reference to the Escherichia coli K-12 substrain MG1655 genomic sequence, wherein the polynucleotide sequence identity percentage is determined using the method of Example 11 of WO / 2016 / 205570. Preferred inducible promoters have a strength of at least 75% (more preferably at least 100%, and most preferably at least 110%) of the corresponding 'wild-type' inducible promoter of E. coli K-12 substrain MG1655 under 'standard' inducing conditions (see Example 5 of WO / 2016 / 205570) as determined using the quantitative PCR method of De Mey et al. (Example 6 of WO / 2016 / 205570). Within the expression construct, the inducible promoter is located at the 5' end (or 'upstream') of the coding sequence of the gene product (e.g., antibody or antibody fragment) to be inducibly expressed, such that the presence of the inducible promoter will direct transcription of the gene product coding sequence in the 5' to 3' direction relative to the coding strand of the polynucleotide encoding the gene product.

[0378] Ribosome binding site. For some antibodies or antibody fragments, the nucleotide sequence of the region between the transcription start site and the start codon of the coding sequence of the gene product to be inducibly expressed corresponds to the 5' untranslated region ('UTR') of the mRNA of the polypeptide gene product. Preferably, the region of the expression construct corresponding to the 5'UTR contains a polynucleotide sequence similar to the consensus ribosome binding site (RBS, also known as the Shine-Dalgarno sequence) found in the host cell species. In prokaryotes (archaea and bacteria), the RBS consensus sequence is GGAGG or GGAGGU, and in bacteria (Escherichia coli), the RBS consensus sequence is AGGAGG or AGGAGGU. The RBS is usually spaced 5 to 10 intervening nucleotides from the start codon. In the expression construct, the RBS sequence is preferably at least 55% identical to the AGGAGGU consensus sequence, more preferably at least 70% identical, most preferably at least 85% identical, and is separated from the start codon by 5 to 10 intervening nucleotides, more preferably 6 to 9 intervening nucleotides, and most preferably 6 or 7 intervening nucleotides. The ability of a given RBS to produce a desired translation initiation rate can be calculated using the RBS calculator on the website salis.psu.edu / software / RBSLibraryCalculatorSearchMode; the same tool can be used to optimize synthetic RBSs to obtain translation rates spanning a 100,000+ fold range (Salis, Methods Enzymol 2011; 498: 19-42).

[0379] Multiple cloning sites. A multiple cloning site (MCS), also known as a polylinker, is a polynucleotide containing multiple restriction sites that are very close to or overlapping each other. The restriction sites in an MCS usually occur once within the MCS sequence, and preferably do not occur within the rest of the plasmid or other polynucleotide construct, allowing restriction enzymes to cut the plasmid or other polynucleotide construct only within the MCS. Examples of MCS sequences are those in the pBAD series of expression vectors, including pBAD18, pBAD18-Cm, pBAD18-Kan, pBAD24, pBAD28, pBAD30, and pBAD33 (Guzman et al., J Bacteriol. 1995 Jul; 177(14):4121-4130); or those in the pPRO series of expression vectors derived from the pBAD vectors, such as pPRO18, pPRO18-Cm, pPRO18-Kan, pPRO24, pPRO30, and pPRO33 (U.S. Pat. No. 8178338 B2; May 15, 2012; Keasling, Jay). A multiple cloning site can be used to generate an expression construct: by placing the multiple cloning site at the 3' end (or downstream) of the promoter sequence, the MCS can be used to insert the coding sequence of the gene product to be expressed or co-expressed into the construct at an appropriate position relative to the promoter so that transcription of the coding sequence will occur. Depending on which restriction enzymes are used for cutting within the MCS, some portion of the MCS sequence may remain in the expression construct after the coding sequence or other polynucleotide sequence is inserted into the expression construct. Any remaining MCS sequence may be upstream or downstream or on both sides of the inserted sequence. The ribosome binding site can be placed upstream of the MCS, preferably immediately adjacent to the MCS or separated from the MCS by only a few nucleotides, in which case the RBS will be located upstream of any coding sequence inserted into the MCS. Another alternative is to include a ribosome binding site within the MCS, in which case the choice of restriction enzyme used for cutting within the MCS will determine whether the RBS is retained and how it relates to the inserted sequence. A further alternative is to include an RBS within the polynucleotide sequence to be inserted into the expression construct at the MCS, preferably in an appropriate relationship to any coding sequence to stimulate translation from the transcribed messenger RNA.

[0380] Expression from a constitutive promoter. The expression construct of the present disclosure may also include a coding sequence expressed from a constitutive promoter. Unlike inducible promoters, constitutive promoters initiate continuous gene product production under most growth conditions. An example of a constitutive promoter is an example of the Tn3 bla gene, which encodes beta-lactamase and is responsible for the ampicillin resistance (AmpR) phenotype conferred to host cells by many plasmids, including pBR322 (ATCC 31344), pACYClW (ATCC 37031) and pBAD24 (ATCC 87399). Another constitutive promoter that can be used in expression constructs is the promoter of the E. coli lipoprotein gene Ipp, which is located at positions 1755731-1755406 (plus strand) in E. coli K-12 substrain MG1655 (Inouye and Inouye, Nucleic Acids Res 1985 May 10; 13(9):3101-3110). Another example of a constitutive promoter that has been used for heterologous gene expression in E. coli is the trpLEDCBA promoter, which is located at positions 1321169-1321133 (minus strand) in E. coli K-12 substrain MG1655 (Windass et al., Nucleic Acids Res 1982 November 11; 10(21):6639-6657). Constitutive promoters can be used in expression constructs for expressing selectable markers, as described herein, and also for constitutive expression of other gene products suitable for co-expression of desired products. For example, transcriptional regulators of inducible promoters (such as AraC, PrpR, RhaR and XylR) can be alternatively expressed from a constitutive promoter on an expression construct identical to the inducible promoter that regulates it or on a different expression construct if they are not expressed from a bidirectional inducible promoter. Similarly, to give a few examples, gene products that can be used to produce or transport inducers, such as PrpEC, AraE or Rha, or proteins that modify the redox environment of the cell, can be expressed from a constitutive promoter in an expression construct. Gene products that can be used to produce co-expressed gene products and the resulting desired products also include chaperones, cofactor transporters, and the like.

[0381] Signal peptide. The antibody or antibody fragment expressed or co-expressed by the method of the present disclosure may contain a signal peptide or lack a signal peptide, depending on whether such gene products need to be exported from the cytoplasm of the host cell to the periplasm, or retained in the cytoplasm. The structural feature of the signal peptide (also referred to as a signal sequence, a leader sequence or a leader peptide) is an extension of hydrophobic amino acids, about five to twenty amino acids long, and generally about ten to fifteen amino acids in length, tending to form a single alpha helix. This hydrophobic extension is usually immediately before a shorter extension rich in positively charged amino acids (particularly lysine). The signal peptide to be cleaved from the mature polypeptide usually ends with an amino acid extension recognized and cleaved by a signal peptidase. The functional feature of the signal peptide is that it can directly transport the polypeptide through the plasma membrane of a prokaryotic organism (or the inner membrane of a gram-negative bacterium such as Escherichia coli) co-translation or translation, or transport the polypeptide to the endoplasmic reticulum of a eukaryotic cell. The extent to which a signal peptide enables transport of a polypeptide to the periplasmic space of a host cell such as E. coli can be determined, for example, by separating periplasmic proteins from proteins retained in the cytoplasm using methods such as those described in Example 12 of WO / 2016 / 205570.

[0382] The following is a description of some genetic modifications that can be performed on host cells containing such expression constructs and inducible promoters that can be used to express or co-express gene products in expression constructs. Unless otherwise specified, examples of these inducible promoters and related genes are from Escherichia coli (E.coli) strain MG1655 (American Type Culture Collection ATCC 700926), which is a substrain of Escherichia coli K-12 (American Type Culture Collection ATCC 10798). Table 1 of WO / 2016 / 205570 lists the nucleotide sequences of these examples of inducible promoters and related genes in the genomic position in Escherichia coli MG1655. Nucleotides and other genetic sequences cited by the genomic position in Table 1 of WO / 2016 / 205570 are explicitly incorporated herein by reference. Additional information about E. coli promoters, genes and strains described herein can be found in many public sources, including online EcoliWiki resources, which are located at ecoliwiki.net.

[0383] Arabinose promoter. (As used herein, "arabinose" means L-arabinose). Several E. coli operons involved in arabinose utilization can be induced by arabinose araBAD, araC, arciE and araFGH, but the terms 'arabinose promoter' and 'ara promoter' are generally used to specify araBAD promoter. Several additional terms have been used to represent E. coli araBAD promoters, such as Para, ParaB, ParaBAD and PBAD. 'ara promoter' used herein or any alternative term given above means E. coli araBAD promoter. From the use of another term 'araC-araBAD promoter', it can be seen that araBAD promoter is considered to be a part of a bidirectional promoter, wherein araBAD promoter controls the expression of araBAD operon in one direction, and araC promoter, which is very close to araBAD promoter and is located on the opposite chain, controls the expression of araC coding sequence in another direction. AraC protein is a positive transcriptional regulator and a negative transcriptional regulator of araBAD promoter. In the absence of arabinose, the AraC protein inhibits the transcription of PBAD, but in the presence of arabinose, the AraC protein changes its conformation when bound to arabinose and becomes a positive regulatory element that allows the transcription of PBAD. The protein encoded by the araBAD operon metabolizes L-arabinose by converting L-arabinose to D-xylulose-5-phosphate through the intermediates L-ribulose and L-ribulose-phosphate. For the purpose of maximizing the induction of expression from an arabinose inducible promoter, it is useful to eliminate or reduce the function of AraA, which catalyzes the conversion of L-arabinose to L-ribulose, and optionally eliminate or reduce at least one of the functions of AraA, AraB and AraD. By eliminating or reducing the ability of the cell to convert arabinose to other sugars, the ability of the host cell to reduce the effective concentration of arabinose in the cell is eliminated or reduced, so that more arabinose can be used for the induction of the arabinose inducible promoter. Genes encoding transporters that move arabinose into the host cell are araE, which encodes a low affinity L-arabinose proton symporter, and the araFGH operon, which encodes a subunit of the ABC superfamily high affinity L-arabinose transporter. Other proteins that can transport L-arabinose into cells are certain mutants of the LacY lactose permease: the LacY(AlWC) and LacY(AlWV) proteins, which have the amino acid cysteine ​​or valine instead of alanine at position 177, respectively (Morgan-Kiss et al., Proc Natl Acad Sci USA 2002 May 28;99(11):7373-7377). In order to achieve homogeneous induction of the arabinose-inducible promoter, it is useful to allow arabinose to be transported into the cell independently of the regulation of arabinose.This can be achieved by eliminating or reducing the activity of the AraFGH transporter and changing the expression of araE so that it is only transcribed from a constitutive promoter. The constitutive expression of araE can be achieved by eliminating or reducing the function of the natural araE gene and introducing an expression construct into the cell, wherein the expression construct includes the coding sequence of the AraE protein expressed by the constitutive promoter. Alternatively, in cells lacking the AraFGH function, the promoter controlling the expression of the host cell chromosome araE gene can be changed from an arabinose inducible promoter to a constitutive promoter. In a similar manner, as another alternative to the homogeneous induction of the arabinose inducible promoter, the host cell lacking the AraE function can have any functional AraFGH coding sequence expressed by a constitutive promoter present in the cell. As another alternative, it is possible to express the araE gene and the araFGH operon from a constitutive promoter by replacing natural araE and araFGH promoters with the constitutive promoter in the host chromosome. It is also possible to eliminate or reduce the activity of AraE and AraFGH arabinose transporters, in which case it is possible to use the mutation in the LacY lactose permease so that the protein can transport arabinose. Since the expression of the lacY gene is not normally regulated by arabinose, the use of LacY mutants (such as LacY(A177C) or LacY(A177V)) will not result in an 'all or nothing' induction phenomenon when the arabinose-inducible promoter is induced by the presence of arabinose. Since the LacY(A177C) protein appears to be more efficient in transporting arabinose into cells, the use of a polynucleotide encoding the LacY(A177C) protein is preferred to the use of a polynucleotide encoding the LacY(A177V) protein.

[0384] Propionate promoter. The 'propionate promoter' or 'prp promoter' is the promoter for the Escherichia coli prpBCDE operon, and is also referred to as PPΦB. Like the ara promoter, the prp promoter is part of a bidirectional promoter that controls the expression of the prpBCDE operon in one direction, and the prpR promoter controls the expression of the prpR coding sequence in the other direction. The PrpR protein is a transcriptional regulator of the prp promoter, and transcription of the prp promoter is activated when the PrpR protein binds 2-methylcitrate ('2-MC'). Propionate (also called propanoate) is a member of the propionic acid (or 'propanoic acid') family. The ion CH3CH2COO— of propionic acid)') is the smallest 'fat' with the general formula H(CH2)"COOH, which shares some properties of this class of molecules: an oily layer when salted out of water and a soap-like potassium salt. Commercially available propionate is usually sold as a monovalent cationic salt of propionic acid (such as sodium propionate (CH3CH2COONa)) or as a divalent cationic salt (such as calcium propionate (Ca(CH3CH2COO)2)). Propionate is membrane permeable and is metabolized to 2-MC by conversion of propionate to propionyl-CoA by PrpE (propionyl-CoA synthetase), which is then converted to 2-MC by PrpC (2-methylcitrate synthase). Other proteins PrpD (2-methylcitrate dehydratase) and PrpB (2-methylisocitrate lyase), encoded by the prpBCDE operon, are involved in the further catabolism of 2-MC into smaller products. , such as pyruvate and succinate. In order to maximize the induction of the propionate inducible promoter by the propionate added to the cell growth medium, it is therefore necessary to have a host cell with PrpC and PrpE activity to convert propionate into 2-MC, and to eliminate or reduce PrpD activity, and optionally eliminate or reduce PrpB activity to prevent 2-MC from being metabolized. Another operon encoding proteins involved in the biosynthesis of 2-MC is the scpA-argK-scpBC operon, also known as the sbm-yg / DGH operon. These genes encode proteins required for converting succinate into propionyl-CoA, which can then be converted into 2-MC by PrpC. Eliminating or reducing the function of these proteins will remove the parallel pathways that produce 2-MC inducers, and therefore may reduce the background expression level of the propionate inducible promoter, and increase the sensitivity of the propionate inducible promoter to exogenously provided propionate.It has been found that the deletion of sbm-ygfD-ygfG-ygfH-ygfl introduced into E. coli BL21(DE3) to create strain JSB (Lee and Keasling, Appl Environ Microbiol 2005 Nov;71(11):6856-6862) helps to reduce background expression in the absence of exogenous inducers, but this deletion also reduces the overall expression of the prp promoter in strain JSB. However, it should be noted that the deletion of sbm-ygfD-ygfG-ygfH-ygfl apparently also affects ygfl, which encodes a putative LysR family transcriptional regulator of unknown function. The gene sbm-yg / DGH is transcribed as an operon, and ygfl is transcribed from the opposite strand. The 3' ends of the ygfti and ygfl coding sequences overlap by several base pairs, so a deletion that removes all of the sbm-yg / DGH operon apparently also removes ygfl coding function. Eliminating or reducing the function of a subset of the sbm-ygfDGH gene products, such as YgfG (also known as ScpB, methylmalonyl-CoA decarboxylase), or deleting most of the sbm-yg / DGH (or scpA-argK-scpBC) operon while leaving enough of the 3' end of the ygfli (or scpC) gene so that expression of ygfli is not affected, may be enough to reduce background expression from the propionate-inducible promoter without reducing the maximum level of induced expression.

[0385] Rhamnose promoter. (As used herein, 'rhamnose' means L-rhamnose). 'Rhamnose promoter' or 'rha promoter' or PrhaSR is the promoter of the Escherichia coli rhaSR operon. Like the ara and prp promoters, the rha promoter is part of a bidirectional promoter, controlling the expression of the rhaSR operon in one direction, and the rhaBAD promoter controls the expression of the rhaBAD operon in the other direction. However, the rha promoter has two transcriptional regulators involved in regulating expression: RhaR and RhaS. The RhaR protein activates the expression of the rhaSR operon in the presence of rhamnose, while the RhaS protein activates the expression of the L-rhamnose catabolism and transport operons rhaBAD and rhaT, respectively (Wickstrum et al., J Bacteriol 2010 January; 192(1): 225-232). Although the RhaS protein can also activate the expression of the rhaSR operon, in fact, RhaS negatively self-regulates this expression to a higher level by interfering with the ability of the cyclic AMP receptor protein (CRP) to co-activate the expression of RhaR. The rhaBAD operon encodes the following proteins for rhamnose catabolism: RhaA (L-rhamnose isomerase), which converts L-rhamnose to L-rhamnose; RhaB (rat urokinase), which phosphorylates L-rhamnose to form L-rhamnose-1-P; and RhaD (rhamnose-1-phosphate aldolase), which converts L-rhamnose-1-P to L-formaldehyde and DHAP (dihydroxyacetone phosphate). In order to maximize the amount of rhamnose available in the cell to induce expression from a rhamnose-inducible promoter, it is necessary to reduce the amount of rhamnose that is catalytically decomposed by eliminating or reducing the function of at least one of RhaA, or optionally RhaA, and RhaB and RhaD. E. coli cells can also synthesize L-rhamnose from α-D-glucose-1-P through the activity of proteins RmlA, RmlB, RmlC, and RmlD (also referred to as RfbA, RfbB, RfbC, and RfbD, respectively) encoded by the rmlBDACX (or rfbBDACX) operon. In order to reduce background expression from a rhamnose-inducible promoter and enhance the sensitivity of exogenously provided rhamnose to induction of a rhamnose-inducible promoter, it may be useful to eliminate or reduce the function of one or more of RmlA, RmlB, RmlC, and RmlD.

[0386] RmlD protein. L-rhamnose is transported into the cell by RhaT (rhamnose permease or L-rhamnose: proton symporter). As described above, the expression of RhaT is activated by the transcriptional regulator RhaS. In order to make the expression of RhaT independent of the induction of rhamnose (which induces RhaS expression), the host cell can be changed so that all functional RhaT coding sequences in the cell are expressed by a constitutive promoter. In addition, the coding sequence of RhaS can be deleted or inactivated so that functional RhaS will not be produced. By eliminating or reducing the function of RhaS in the cell, due to the absence of RhaS negative autoregulation, the expression level of the rhaSR promoter is increased, while the expression level of the rhamnose catalytic operon rhaBAD is reduced, further increasing the ability of rhamnose to induce the expression of the rha promoter.

[0387] Xylose promoter. (As used herein, "xylose" means D-xylose.) The xylose promoter or 'xylose promoter' or PxyiA means the promoter for the E. coli xylAB operon. The xylose promoter region is organizationally similar to other inducible promoters, as both the xylAB operon and the xylFGHR operon are expressed in opposite directions from adjacent xylose-inducible promoters on the E. coli chromosome (Song and Park, J Bacteriol. 1997 Nov; 179(22): 7025-7032). The transcriptional regulator for both the PxyiA and PxyiF promoters is XylR, which activates expression of these promoters in the presence of xylose. The xylR gene is expressed as part of the xylFGHR operon or from its own weak promoter, which is not induced by xylose, and is located between the xylH and xylR protein coding sequences. D-xylose is metabolized by XylA (D-xylose isomerase), which converts D-xylose into D-xylulose, which is then phosphorylated by XylB (xylulose kinase) to form D-xylulose-5-P. In order to maximize the amount of xylose that can be used in cells to induce expression from a xylose inducible promoter, it is necessary to reduce the amount of xylose that is catalytically decomposed by eliminating or reducing the function of at least XylA or optionally both XylA and XylB. The xylFGHR operon encodes XylF, XylG and XylH, which are subunits of the ABC superfamily high-affinity D-xylose transporter. The xylE gene encoding the Escherichia coli low-affinity xylose-proton symporter represents a separate operon, and its expression can also be induced by xylose. In order to make the expression of the xylose transporter independent of the induction of xylose, the host cell can be changed so that all functional xylose transporters are expressed by a constitutive promoter. For example, the xylFGHR operon can be altered so that the xylFGH coding sequence is deleted, XylR is the only active protein expressed from the xylose-inducible PxyiF promoter, and the xylE coding sequence is expressed from a constitutive promoter rather than its native promoter. As another example, the xylR coding sequence is expressed from the PxyiA or promoter in the expression construct, while the xylFGHR operon is deleted and xylE is constitutively expressed, or alternatively the xylFGH operon (lacking the xylR coding sequence as it is present in the expression construct) is expressed from a constitutive promoter, and the xylE coding sequence is deleted or altered so that it does not produce active protein.

[0388] Lactose promoter. The term 'lactose promoter' refers to a lactose-inducible promoter for the lacZYA operon, also known as lacZpl; this lactose promoter is located at about 365603-365568 (negative strand, with a NA polymerase binding ('-35') site at about 365603-365598, a Pribnow box ('-10') at 365579-365573, and a transcription start site at 365567) in the genomic sequence of Escherichia coli K-12 substrain MG1655 (NCBI reference sequence NC 000913.2, January 11, 2012). In some embodiments, the inducible co-expression system of the present disclosure may comprise a lactose-inducible promoter, such as the lacZYA promoter. In other embodiments, the inducible co-expression system of the present disclosure comprises one or more inducible promoters that are not lactose-inducible promoters.

[0389] Alkaline phosphatase promoter. The terms 'alkaline phosphatase promoter' and 'phoA promoter' refer to the promoter of the phoApsiF operon, i.e., a promoter induced under phosphate starvation conditions. The phoA promoter region is located at approximately 401647-401746 in the genomic sequence of Escherichia coli K-12 substrain MG1655 (NCBI reference sequence NC 000913.3, December 16, 2014) (positive strand, where the Pribno box ('-10') is at 401695-401701 (Kikuchi et al., Nucleic Acids Res 1981 November 11; 9(21): 5671-5678)). The transcriptional activator of the phoA promoter is PhoB, which is a transcriptional regulator that, together with the sensor protein PhoR, forms a two-component signal transduction system in Escherichia coli. PhoB and PhoR are transcribed from the phoBR operon located at about 417050-419300 (positive strand, with PhoB coding sequence at 417,142 to 417,831 and PhoR coding sequence at 417,889 to 419,184) in the genomic sequence of Escherichia coli K-12 substrain MG1655 (NCBI reference sequence NC 000913.3, December 16, 2014). The phoA promoter differs from the above-mentioned inducible promoters in that it is induced by the lack of a substance - intracellular phosphate rather than by the addition of an inducer. Therefore, the phoA promoter is generally used to direct the transcription of gene products produced at a stage where the host cell is depleted of phosphate (e.g., the late stage of fermentation). In some embodiments, the inducible co-expression system of the present disclosure may include the phoA promoter. In other embodiments, the inducible co-expression system of the present disclosure includes one or more inducible promoters that are not the phoA promoter.

[0390] Affinity determination

[0391] As described herein, the technology of the present invention may include a computer system, a computer-implemented method, and / or a computer-readable medium / media that predicts or generates structural information of a biomolecule (such as an antibody). In some aspects, a computer system predicts, for example, the binding affinity between an antibody and an antigen. However, wet lab techniques are contemplated for (i) confirming predicted affinity or (ii) generating data to train an affinity model described herein. Antibody binding and antibody affinity determination analysis are well known in the art.

[0392] In one embodiment, the activity-specific cell enrichment method (ACE) can be used to identify host cells expressing "active" antibodies rather than "inactive materials". The difference between active antibodies and inactive antibodies is the ability of active antibodies to specifically bind to binding partner molecules (such as antigens or epitopes). The ACE analysis scheme is described in WO / 2021 / 146626, which is incorporated herein by reference. It will be appreciated by those of ordinary skill in the art that ACE can not only distinguish between active / inactive in a binary manner, but also calculate scores proportional to affinity. Therefore, ACE provides quantitative analysis information, not just binary / Boolean information, which enables the modeling techniques herein to perform regression techniques. This richer modeling represents a favorable improvement over the limited binary classification of conventional techniques.

[0393] In another embodiment, the HiPR binding assay described in WO / 2021 / 163349 and incorporated herein by reference is used in combination with the methods provided herein.

[0394] Binding assays, such as assays measuring protein-protein interactions (including antibody-antigen interactions and including measuring binding affinity) are well known in the art. For example, surface plasmon resonance (SPR), dual polarisation interferometry (DPI), static light scattering (SLS), dynamic light scattering (DLS), flow-induced dispersion analysis (FIDA), fluorescence polarization / anisotropy, fluorescence resonance energy transfer (FRET), bio-layer interferometry (BLI), isothermal titration calorimetry (ITC), microscale thermophoresis (MST), single colour reflectometry (SCORE) are contemplated. In addition, bimolecular fluorescence complementation (BiFC), affinity electrophoresis, label transfer, phage display, tandem affinity purification (TAP), cross-linking, quantitative immunoprecipitation combined with knock-down (QUICK) and proximity ligation assay (PLA) are other well-known assays that provide information on protein-protein interactions.

[0395] In some embodiments, the binding affinity of the antibodies described herein is measured by array surface plasmon resonance (SPR) according to standard techniques (Abdiche et al. (2016) MAbs 8: 264-277). Briefly, antibodies are immobilized on HC 30M chips at four different densities / antibody concentrations. Antibody targets at different concentrations (0-500 nM) are then bound to the captured antibodies. Kinetic analysis is performed using Carterra software to extract the association and dissociation rate constants (k, respectively) for each antibody. a and k d ). According to kd / k a The apparent affinity constant (K D In some embodiments, kinetics and affinity are determined using Carterra LSA. In other embodiments, binding affinity can be determined by surface plasmon resonance (e.g., BIAcore) using, for example, an IBISMX96 SPR system from IBIS Technologies or a Carterra LSA SPR platform. TM ) or, for example, using the Octet TM System, measured by bio-layer interferometry. In some embodiments, a biosensor instrument such as Octet RED384, ProteOn XPR36, IBIS MX96 and Biacore T100 (Yang, D. et al. J. Vis. Exp., 2017, 122: 55659) is used.

[0396] K D is the equilibrium dissociation constant, i.e., k between the antibody and its antigen off / k on The ratio of K D Inversely proportional to affinity. K D The value is related to the antibody concentration, so K D The lower the value (the lower the concentration), the higher the affinity of the antibody. The K values ​​of antibodies (including reference antibodies and variant antibodies) according to various embodiments of the present disclosure are shown in Table 1. D It can be, for example, in the micromolar range (10 -4 Up to 10 -6 ) in the nanomolar range (10 -7 Up to 10 -9 ) within the picomolar range (10 -10 Up to 10 -12 ) or in the femtomolar range (10 -13 Up to 10 -15 ). In some embodiments, the antibody affinity of the variant antibody is improved by about 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, or 50% or more relative to the reference antibody. The improvement can also be relative to the fold change (e.g., 2×, 4×, 6×, or 2-fold, 3-fold, 4-fold, 5-fold, 6-fold, 7-fold, 8-fold, 9-fold, 10-fold or more increase in binding activity, etc.) and / or order of magnitude (e.g., 10 7 , 10 8 , 10 9 etc.) to indicate.

[0397] In some embodiments, data generated from the antibodies and assays described herein are used to train one or more models, as will be described below.

[0398] Other considerations

[0399] Before further describing the present disclosure, it should be understood that the present disclosure is not limited to the specific embodiments described, and therefore may of course vary. It should also be understood that the terminology used herein is for the purpose of describing specific embodiments only, and is not intended to be limiting, as the scope of the present disclosure will be limited only by the appended claims.

[0400] Where a range of values ​​is provided, it is understood that each intervening value between the upper and lower limits of the range (to one-tenth of the unit of the lower limit, unless the context clearly indicates otherwise) and any other stated value or intervening value within the stated range are encompassed within the disclosure. The upper and lower limits of these smaller ranges may be independently included in the smaller ranges and are also encompassed within the disclosure, subject to any specifically excluded limits in the stated ranges. When a stated range includes one or both of the limits, ranges excluding either or both of those included limits are also encompassed within the disclosure.

[0401] Unless otherwise defined, all technical terms and scientific terms used herein have the same meaning as those commonly understood by those of ordinary skill in the art to which the disclosure belongs. Although any methods and materials similar or equivalent to those described herein may also be used in the practice or testing of the disclosure, preferred methods and materials are now described. All publications mentioned herein are incorporated herein by reference to disclose and describe the methods and / or materials to which the referenced publications are combined.

[0402] It must be noted that, as used herein and in the appended claims, the singular forms "a / an," and "the" include plural referents unless the context clearly dictates otherwise. Thus, for example, reference to a "conformational switching probe" includes a plurality of such conformational switching probes, and reference to a "microfluidic device" includes reference to one or more microfluidic devices and equivalents thereof known to those of ordinary skill in the art, and so on. It should be further noted that the claims may be drafted to exclude any element, for example, any optional element. Therefore, this statement is intended to serve as a precondition for the use of exclusive terms such as "solely," "only," or the use of a "negative" limitation in conjunction with the recitation of claim elements.

[0403] As will be apparent to those skilled in the art upon reading this disclosure, each of the individual embodiments described and illustrated herein has discrete components and features that can be readily separated or combined with the features of any of the other several embodiments without departing from the scope or spirit of the present disclosure. Any recited method can be performed in the order of events recited or in any other order that is logically possible. This is intended to provide support for all such combinations.

[0404] The various embodiments described above can be combined to provide additional embodiments. The entire contents of all U.S. patents, U.S. patent application publications, U.S. patent applications, foreign patents, foreign patent applications, and non-patent publications mentioned in this specification and / or listed in the application data sheet are incorporated herein by reference. Aspects of the embodiments can be modified as necessary to adopt the concepts of various patents, applications, and publications to provide additional embodiments.

[0405] These and other changes can be made to the embodiments in light of the above detailed description. In general, in the appended claims, the terms used should not be interpreted as limiting the claims to the specific embodiments disclosed in the specification and claims, but should be interpreted as including all possible embodiments and the full range of equivalents granted by these claims. Therefore, the claims are not limited by the present disclosure.

Claims

1. A computing system for training a machine learning model to generate structural information of a target biomolecule, comprising: one or more processors; and one or more non-transitory computer-readable media having stored thereon instructions that, when executed by the one or more processors, cause the computing system to: receiving one or more training inputs, including one or more of: (i) input biomolecule structure information, (ii) input biomolecule binding partner structure information, or (iii) input biomolecule-input binding partner binding complex structure information; processing the one or more training inputs with a machine learning-based biomolecular prediction model to generate predicted biomolecular structure information; evaluating a loss function comparing the predicted biomolecular structure information to a ground truth; and One or more values ​​of one or more parameters of the machine-learned model are modified based at least in part on the loss function. 2 . The computing system of claim 1 , wherein the target biomolecule is an antibody or a protein.

3. The computing system of claim 1, wherein the input biomolecule structure information comprises three-dimensional coordinates of a primary sequence.

4. The computing system of claim 1, wherein the input biomolecule binding partner is an antigen, a receptor, a ligand, or a cell membrane.

5. The computing system of claim 1 , wherein the training inputs are respectively represented in the non-transitory computer-readable medium as at least one of: (i) Protein Data Bank (PDB) data format, (ii) JSON data format, or (iii) XML data format. 5a. The computing system of claim 1, wherein the training input comprises at least 1, at least 5, at least 10, or at least 100 binding agents from Table E1, Table E2, Table E3, and / or Table E4 above.

6. The computing system of claim 1, wherein the input biomolecule binding partner is a protein, and the input biomolecule binding partner structural information comprises a three-dimensional structure.

7. The computing system of claim 1 , the one or more non-transitory computer-readable media having stored thereon additional instructions that, when executed by the one or more processors, cause the computing system to: The training input is received from a database comprising antibody structures and / or antigen structures.

8. The computing system of claim 1 , the one or more non-transitory computer-readable media having stored thereon additional instructions that, when executed by the one or more processors, cause the computing system to: The training input is received from the Structural Antibody Database (SAbDab).

9. A computing system for generating structural information of a target biomolecule, comprising: one or more processors; and One or more non-transitory computer-readable media having stored thereon: A machine learning-based biomolecule prediction model that is trained to predict structural information of biomolecules based on input; and instructions that, when executed by the one or more processors, cause the computing system to: (1) receiving a target input, wherein the target input comprises one or more of the following: a target binding partner primary sequence, a three-dimensional coordinate of a target binding partner, a target binding partner epitope primary sequence, or a three-dimensional coordinate of a target binding partner epitope primary sequence, or a fragment or portion of any of the foregoing; and (2) Predicting the structural information of the target biomolecule by processing the target input with the machine learning-based biomolecule prediction model. 9a. The computing system according to claim 9, The target input includes CDRH1 and CDRH2 of trastuzumab-Bh1 and the structure of VEGF. 9b. The computing system according to claim 9, The target input includes the CDRH1 and CDRH2 of trastuzumab and the structure of HER2.

10. The computing system of claim 9, wherein the structural information is represented in the computer-readable medium as at least one of: (i) Protein Data Bank (PDB) data format, (ii) JSON data format, or (iii) XML data format.

11. The computing system of claim 1 or claim 9, wherein the machine learning biomolecule prediction model is an artificial neural network.

12. The computing system of claim 11, wherein the artificial neural network comprises at least one of a geometric neural network, a transformer network, or a geometric vector perceptron network.

13. The computing system of claim 9, wherein the structural information of the target biomolecule comprises data representing one or more of: (i) amino acids of the target biomolecule; (ii) the peptide sequence of the target biomolecule; (iii) the polypeptide sequence of the target biomolecule; (iv) the primary sequence of the target biomolecule; (v) one or more secondary structures of the target biomolecule; (vi) one or more tertiary structures of the target biomolecule; (vii) one or more quaternary structures of the target biomolecule; or (viii) the three-dimensional coordinates of the primary sequence of the target biomolecule.

14. The computing system of claim 13, the one or more non-transitory computer-readable media having stored thereon additional instructions that, when executed by the one or more processors, cause the computing system to: Output the data representing one or more of the following: (i) the amino acid of the target biomolecule; (ii) the peptide sequence of the target biomolecule; (iii) the polypeptide sequence of the target biomolecule; (iv) the primary sequence of the target biomolecule; (v) the one or more secondary structures of the target biomolecule; (vi) the one or more tertiary structures of the target biomolecule; (vii) the one or more quaternary structures of the target biomolecule; or (viii) the three-dimensional coordinates of the primary sequence of the target biomolecule.

15. The computing system of claim 14, wherein the data is encoded in (i) a Protein Data Bank (PDB) data format, (ii) a JSON data format, (iii) an XML data format, or (iv) another suitable data format.

16. The computing system of claim 14, wherein the polypeptide sequence of the target biomolecule corresponds to a complementarity determining region of an antibody.

17. The computing system of claim 14, wherein the structural information of the target biomolecule is selected from the group consisting of: alpha helix, beta pleated sheet, and coil.

18. The computing system of claim 9, wherein the target biomolecule is an antibody and the target binding partner is an antigen.

19. The computing system of claim 9, the one or more non-transitory computer-readable media having stored thereon additional instructions that, when executed by the one or more processors, cause the computing system to: Step 2 is repeated until the structural information of the target biomolecule is complete.

20. The computing system of claim 19, wherein the repeating is performed iteratively.

21. The computing system of claim 9, the one or more non-transitory computer-readable media having stored thereon additional instructions that, when executed by the one or more processors, cause the computing system to: receiving an input parameter specifying a desired length of amino acids of the target biomolecule; and A target biomolecule having the desired length is generated.

22. The computing system of claim 9, the one or more non-transitory computer-readable media having stored thereon additional instructions that, when executed by the one or more processors, cause the computing system to: The structural information of the target biomolecule is updated by folding one or more proteins.

23. The computing system of claim 9, wherein the target binding partner is at least one of (i) a primary amino acid sequence, (ii) an antigenic epitope, and / or (iii) a known three-dimensional structure of an antigen.

24. The computing system of claim 9, wherein the structural information of the target biomolecule is predicted in at least one of the following: (i) an N-terminus to C-terminus order, (ii) a C-terminus to N-terminus order, (iii) by random sampling; or (iv) by non-random sampling.

25. The computing system of claim 9, The one or more non-transitory computer-readable media also store thereon: a machine learning-based docking model trained to generate docked complexes from two or more input three-dimensional biomolecules; and The one or more non-transitory computer-readable media have additional instructions stored thereon that, when executed by the one or more processors, cause the computing system to: The predicted structural information of the target input and the target biomolecule is processed using the docking model to generate a docked complex comprising the target input and the target biomolecule.

26. The computing system of claim 25, the one or more non-transitory computer-readable media having stored thereon additional instructions that, when executed by the one or more processors, cause the computing system to: The docked complex is provided as a docked complex output.

27. The computing system of claim 25, The one or more non-transitory computer-readable media also store thereon: a machine learning-based binding affinity prediction model trained to predict binding affinity from an input docking complex; and The one or more non-transitory computer-readable media have additional instructions stored thereon that, when executed by the one or more processors, cause the computing system to: The docked complex is processed using the trained affinity prediction model to generate a binding affinity value.

28. The computing system of claim 27, wherein the machine learning model is an artificial neural network.

29. The computing system of claim 28, wherein the artificial neural network is at least one of a graph convolutional neural network, a message passing neural network, or a geometric vector perceptron network.

30. The computing system of claim 28, the one or more non-transitory computer-readable media having stored thereon additional instructions that, when executed by the one or more processors, cause the computing system to: The node information contained within the artificial neural network is aggregated using a pooling operation to generate the binding affinity value.

31. The computing system of claim 28, the one or more non-transitory computer-readable media having stored thereon additional instructions that, when executed by the one or more processors, cause the computing system to: The docked complex is provided as a docked complex output.

32. The computing system of claim 9, the one or more non-transitory computer-readable media having stored thereon additional instructions that, when executed by the one or more processors, cause the computing system to: The machine learning-based biomolecule prediction model is trained by providing sequence data as input to force teach the biomolecule prediction model.

33. The computing system of claim 25, the one or more non-transitory computer-readable media having stored thereon additional instructions that, when executed by the one or more processors, cause the computing system to: The docking model is trained by providing antibody structural data as input to force teach the docking model.

34. The computing system of claim 27, the one or more non-transitory computer-readable media having stored thereon additional instructions that, when executed by the one or more processors, cause the computing system to: The affinity prediction model is trained by providing sequence data as input to force teach the affinity prediction model. 34a. The computing system of claim 27, the one or more non-transitory computer-readable media having additional instructions stored thereon that, when executed by the one or more processors, cause the computing system to: The machine learning-based biomolecule prediction model is trained using at least 1, at least 5, at least 10, or at least 100 binding agents from Table E1, Table E2, Table E3, and / or Table E4 above.

35. A computational system for predicting affinity of a target biomolecule, comprising: One or more processors ; as well as One or more non-transitory computer-readable media having stored thereon: Machine learning-based affinity prediction artificial neural network, including: (i) one or more biomolecule prediction layers trained to predict biomolecule structural information from target inputs; (ii) one or more docking layers trained to generate docked complexes from two or more input three-dimensional biomolecules; and (iii) one or more affinity prediction layers trained to predict affinities from input docking complexes; wherein the one or more biomolecule prediction layers, the one or more docking layers, and the one or more affinity prediction layers are connected; and instructions that, when executed by the one or more processors, cause the computing system to: receiving a target input comprising one or more of a target binding partner sequence, a target binding partner, or a target epitope; and The target input is processed using the affinity prediction artificial neural network to generate a docking complex corresponding to the target input and a corresponding structural affinity value.

36. The computing system of claim 35, the one or more non-transitory computer-readable media having stored thereon additional instructions that, when executed by the one or more processors, cause the computing system to: receiving a sequence affinity prediction value from an affinity sequencing prediction model; and The structural affinity values ​​are averaged with the sequence affinity predictions to generate an overall affinity prediction.

37. The computing system of claim 35, the one or more non-transitory computer-readable media having stored thereon additional instructions that, when executed by the one or more processors, cause the computing system to: The one or more biomolecule prediction layers are pre-trained using bound or unbound structures.

38. The computing system of claim 35, the one or more non-transitory computer-readable media having stored thereon additional instructions that, when executed by the one or more processors, cause the computing system to: The one or more biomolecule prediction layers and the one or more docking layers are pre-trained using the combined structures.

39. The computing system of claim 35, the one or more non-transitory computer-readable media having stored thereon additional instructions that, when executed by the one or more processors, cause the computing system to: The machine learning affinity prediction artificial neural network is pre-trained using affinity training data.

40. The computing system of claim 35, wherein the affinity training data comprises an affinity score that is proportional to activity.

41. The computing system of claim 35, the one or more non-transitory computer-readable media having stored thereon additional instructions that, when executed by the one or more processors, cause the computing system to: receiving an external docking complex corresponding to the target input; and The external docked complex is compared to the generated docked complex.

42. The computing system of claim 35, the one or more non-transitory computer-readable media having stored thereon additional instructions that, when executed by the one or more processors, cause the computing system to: The gradient flow of the machine-learned affinity prediction artificial neural network is controlled by applying a stop gradient function in at least one of the one or more biomolecule prediction layers, the one or more docking layers, or the one or more affinity prediction layers.

43. The computing system of claim 35, the one or more non-transitory computer-readable media having stored thereon additional instructions that, when executed by the one or more processors, cause the computing system to: The neural network is trained using at least 1, at least 5, at least 10, or at least 100 binding agents from Table E1, Table E2, Table E3, and / or Table E4 above.