Unlocking de novo antibody design with generative artificial intelligence

Abstract

A computing system for generating structural information of a biomolecule, comprising: a processor; 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 receive training inputs, process the training inputs with a machine-learned biomolecule prediction model, evaluate a loss function, and modify parameters of the machine-learned model. A computing system for generating structural information of a target biomolecule, comprising: a processor; and one or more non-transitory computer-readable media having stored thereon a machine-learned biomolecule prediction model and instructions that, when executed by the one or more processors, cause the computing system to receive a target input and predict structural information of the target biomolecule. 1. A computing system for predicting affinity of a target biomolecule, comprising: a processor; and one or more non-transitory computer readable media comprising: a machine-learned affinity prediction artificial neural network comprising: (i) one or more biomolecular prediction layers trained to predict biomolecular 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 from the input docking complexes, wherein the one or more biomolecular 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, that cause the computing system to receive a target input and process 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.

Description

[Technical field]

[0001] Incorporation by Reference of Electronically Submitted Materials A sequence listing, which is part of this disclosure, is submitted contemporaneously herewith in computer readable format. The name of the file containing the sequence listing is "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.

[0002] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims priority to U.S. Provisional Patent Application No. 63 / 478,933, entitled UNLOCKING DE NOVO ANTIBODY DESIGN WITH GENERATIVE ARTIFICIAL INTELLIGENCE, filed January 7, 2023, and U.S. Provisional Patent Application No. 63 / 308,495, entitled STRUCTURE-BASED DESIGN OF BINDING PARTNER-TARGETED BIOMOLECULES, filed February 9, 2022, the contents of each of which are incorporated herein by reference in their entirety. [Background technology]

[0003] Antibodies are diverse proteins that naturally bind to and are used by the immune system to neutralize foreign bodies such as viruses, fungi, and bacteria. Thus, antibodies have the potential to serve as drug candidates for many infectious diseases and cancers. Conventional techniques for in silico automated design and / or antibody improvement are inadequate, for example, because such deep learning-based protein design techniques blindly design proteins, use only open source data, and do not allow for targeted design to specified binding partners. Such techniques also have not demonstrated the ability to tailor the design with other criteria / goals. Conventional laboratory-based techniques are time-consuming and expensive. For example, conventional de novo antibody discovery requires time- and resource-intensive screening of large immune or synthetic libraries. These methods also offer little control over the output sequence, which can result in lead candidates with suboptimal binding and poor development suitability. Conventional generative antibody design techniques have not demonstrated de novo antibody design with experimental validation. Thus, improved techniques for protein design are needed. Summary of the Invention

[0004] In one aspect, 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 storing instructions that, when executed by the one or more processors, cause the computing system to: (1) receive one or more training inputs, the 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; (2) process the one or more training inputs with a machine-learned biomolecule prediction model to generate predicted biomolecule structural information; (3) evaluate a loss function that compares the predicted biomolecule structural information to ground truth values; and (4) modify one or more values ​​of one or more parameters of the machine-learned model based at least in part on the loss function.

[0005] In another aspect, 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-learned biomolecular 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 a target binding partner primary sequence, a three-dimensional coordinate of the target binding partner, a target binding partner epitope primary sequence, or a three-dimensional coordinate of the 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 biomolecule by processing the target input with the machine-learned biomolecular prediction model.

[0006] In yet another aspect, a computing system for predicting affinity of a target biomolecule, comprising: one or more processors; and one or more non-transitory computer readable media comprising a machine-learned affinity prediction artificial neural network comprising: (i) one or more biomolecular prediction layers trained to predict biomolecular 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 from the input docking complexes, wherein the one or more biomolecular prediction layers, the one or more docking layers, and the one or more affinity prediction layers are connected; and instructions, which 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 corresponding to the target input and a corresponding structural affinity value. [Brief description of the drawings]

[0007] [Figure 1A] Illustrating zero-shot generative AI for de novo antibody design, deep learning models trained on antibody-antigen interactions combined with ultra-high throughput wet-lab experiments enable the design of binders to undiscovered antigens without the need for further affinity maturation or lead optimization. [Figure 1B] 1 illustrates an exemplary computing environment for implementing the techniques, according to some aspects. [Figure 2A] 1 illustrates an exemplary computer-implemented method for training a machine learning model to generate structural information for biomolecules, according to some embodiments. [Figure 2B]FIG. 1 illustrates a data flow block diagram for training and computing one or more machine learning models, according to some embodiments. [Figure 2C] 1 illustrates a computer-implemented method for generating structural information of a target biomolecule, according to some embodiments. [Figure 3A] FIG. 1 illustrates a block flow diagram of an exemplary computer-implemented method for performing in silico de novo biomolecular discovery, according to some embodiments. [Figure 3B] 1 illustrates an exemplary block diagram according to some aspects. [Figure 3C] 3 illustrates a block diagram of exemplary CDR3 in silico discovery inputs and outputs, according to some embodiments, where block 310b includes SEQ ID NO:1 and block 310c includes SEQ ID NO:2. [Figure 3D] FIG. 1 illustrates an exemplary block diagram of machine learning based drug candidate design, according to some embodiments. [Figure 3E] FIG. 1 illustrates an exemplary block flow diagram of antigen encoding, according to some embodiments. [Figure 3F] 3 illustrates an exemplary block flow diagram of antibody generation training, according to some embodiments, where block 322a includes SEQ ID NO:3 and SEQ ID NO:4, and block 322c includes SEQ ID NO:5. [Figure 3G] Illustrates an exemplary block flow diagram of antibody generation training according to some embodiments, including SEQ ID NO:6, SEQ ID NO:7, and SEQ ID NO:8. [Figure 3H] Illustrates an exemplary block flow diagram of antibody generation training according to some embodiments, including SEQ ID NO:8 and SEQ ID NO:9. [Figure 3I] Illustrates an exemplary block flow diagram of antibody generation training according to some embodiments, including SEQ ID NO:3, SEQ ID NO:4, and SEQ ID NO:5. [Figure 3J] 1 illustrates an exemplary block flow diagram of antibody generation training according to some embodiments, including SEQ ID NO:3 and SEQ ID NO:4. [Figure 3K]Illustrates an exemplary block flow diagram of antibody generation training according to some embodiments, including SEQ ID NO:3, SEQ ID NO:4, and SEQ ID NO:5. [Figure 3L] FIG. 1 illustrates an example block diagram illustrating supervised machine learning model output, according to some aspects. [Figure 4] 1 illustrates an example graphical user interface including visualized ground truth structural information and predicted structural information, according to some aspects. [Figure 5A] FIG. 1 illustrates an exemplary block flow diagram for complete in silico biomolecule design, according to some embodiments. [Figure 5B] FIG. 1 illustrates an exemplary block diagram of a fully in silico de novo biomolecular design artificial intelligence model, according to some embodiments. [Figure 5C] FIG. 1 illustrates an exemplary block flow diagram of an in silico biomolecular design pipeline, according to some embodiments. [Figure 6A] 1 illustrates exemplary in silico biomolecular engineering involving binder prediction based on optimization of CDRH3, while keeping the remaining CDRs unchanged, according to some embodiments comprising SEQ ID NOs:553-555. [Figure 6B] 1 illustrates an exemplary visualization of ranibizumab and trastuzumab bH1, according to some embodiments. [Figure 6C] 1 illustrates exemplary sensorgrams corresponding to surface plasmon response results generated for predicted in silico binding of trastuzumab-bH1 CDR3 variants that bind VEGF, according to some embodiments. [Figure 6D] Illustrates exemplary tabular data corresponding to surface plasmon response results generated for predicted in silico binding of Trastuzumab-bH1 CDR3 variants binding to VEGF in FIG. 6C according to some embodiments, including SEQ ID NOs: 10-22. [Figure 6E]1 illustrates exemplary sensorgrams corresponding to surface plasmon response results generated for predicted in silico binding of Trastuzumab CDR3 variants that bind to Her2, according to some embodiments. [Figure 6F] Illustrates exemplary tabular data corresponding to surface plasmon response results generated for predicted in silico binding of Trastuzumab CDR3 variants that bind to Her2 in FIG. 6E according to some embodiments, including SEQ ID NOs: 23-38. [Figure 7A] FIG. 1 illustrates an example block diagram of early laboratory validation of modeling predictions, according to some embodiments. [Figure 7B] FIG. 1 illustrates an exemplary block diagram of in silico generation of de novo designed trastuzumab CDR3 with comparable or improved binding affinity to VEGF without multispecificity, according to some embodiments. [Figure 7C] FIG. 1 illustrates an exemplary block diagram of in silico generation of de novo designed trastuzumab CDR3 that binds HER2 with high specificity, according to some embodiments. [Figure 7D] FIG. 1 illustrates an exemplary block diagram of a negative control having a scrambled CDR3 sequence that does not bind to the antigen, according to some embodiments, including SEQ ID NO: 39 and SEQ ID NO: 40. [Figure 7E] Illustrates an exemplary block diagram of generated CDR3 sequences exhibiting high diversity, according to some embodiments, including SEQ ID NO:20. [Figure 7F] 1 illustrates an exemplary tree diagram showing the high degree of diversity of in silico generation of de novo designed variants, according to some embodiments. [Figure 8A] Illustrates a logo plot of HCDR3 of 421-linked trastuzumab variants, where the higher diversity observed in the center of the CDR (hypervariable domain) corresponds to the D germline gene, while the tails correspond to the J and V germline genes, which tend to be less diverse. [Figure 8B]FIG. 1 illustrates the binding affinities of AI-generated zero-shot binders, including 71 designs with affinity comparable to trastuzumab (<10 nM) and 3 designs with tighter binding, according to some embodiments. [Figure 8C] Designed variant binding affinity versus edit distance to trastuzumab is depicted; edit distances range from 2 mutations (84.6% sequence identity) to 12 mutations (7.7% sequence identity), illustrating the novelty of the design; although binding affinity tends to decrease as edit distance increases, the technique still identifies binders with comparable (within 0.5 logs) or higher binding affinity in all groups. [Figure 8D] The pairwise edit distance between binders (e.g., maximum 15 mutations, median 8, mean 7.7, standard deviation 2.1) is illustrated. [Figure 8E] 1 illustrates the pairwise edit distance distribution (log scale) of pairwise HCDR3 edit distances of AI-designed binders to HER2, according to some embodiments. [Figure 8F] Illustrates the length distribution of HCDR3 lengths of AI-designed binders to HER2. [Figure 9A] Binder minimum edit distance to training data HCDR3 is depicted, and across designs binders are observed that are up to 8 mutations away from any HCDR3 in the training data. [Figure 9B] 1 illustrates the binder minimum edit distance to any HCDR3 in the SAbDab of AI assigned binders to HER2, according to some embodiments. [Figure 9C] 1 illustrates the binder minimum edit distance of OAS HCDR3 according to some embodiments. [Figure 9D] 1 illustrates the naturalness scale of designed binders of the invention versus their diversity relative to trastuzumab, according to some embodiments. [Figure 9E] 1 illustrates the naturalness of designed binders versus controls, according to some embodiments. [Figure 9F]Illustrates the distribution of minimum sum of edit distances to HCDR3 in the OAS of AI-designed binders to HER2, and to trastuzumab HCDR1 and HCDR2 to OAS HCDR1 and HCDR2 sequences, respectively. [Figure 9G] FIG. 1 illustrates naturality scores versus minimum edit distance to HCDR3 in OAS for minimum OAS ED distribution of naturality scores for de novo designed HER2 binders. [Figure 9H] Illustrates naturalness score versus minimum edit distance to any HCDR 123 in the OAS. [Figure 9I] Illustrates distribution of minimum edit distance to HCDR 3 in the OAS of multi-stage, multi-HCDR, AI-designed binders to HER2. [Figure 9I] 1 illustrates the minimum edit distance to any HCDR3 in an OAS, according to some embodiments. [Figure 9J] Illustrates the distribution of minimum edit distances to HCDR1, HCDR2, and HCDR3 in the OAS of a multi-stage, multi-HCDR, AI-designed binder for HER2. [Figure 10A] FIG. 1 illustrates the conformational flexibility of de novo designed HCDR3 Her2 binders, including SEQ ID NOs: 41-48. [Figure 10B] Illustrates a space-filling representation of the HCDR3 loop interacting with epitope residues with HCDR3 atoms of 5 A. The Trastuzumab HCDR3 loop is in red and the de novo Her2 binder HCDR3 loop is in blue, showing two distinct epitope pockets that interact differentially with each HCDR3 residue, with the epitope interaction surface varying based on HCDR3 sequence and conformation. [Figure 10C] Illustrated is a comparison of the Trastuzumab-Her2 structure with a de novo designed HCDR3-Her2 complex showing the superposition of the Trastuzumab-Her2 structure with a de novo designed HCDR3-Her2 complex, illustrating the conformational differences in the HCDR3 backbone, with the main chain backbone illustrated as ribbons and the spatially conserved side chains shown as sticks. [Figure 10D]Illustrates a stick representation of the HCDR3 loop interacting with epitope residues with HCDR3 atoms at 5 Å. The trastuzumab HCDR3 loop is in red and the de novo Her2 binder HCDR3 loop is in blue. The 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. [Figure 11A] Illustrates alignment of designed VEGF binders to the ranibizumab sequence, including SEQ ID NO:49 and SEQ ID NO:50. [Figure 11B] Illustrates alignment of SARS-CoV-2 Omicron binder to casirivimab. The designed HCDR3s are diverse and novel, as they are 6 mutations away from the HCDR3 of casirivimab, at least 2 mutations away from any HCDR3 in OAS, and at least 4 mutations away from any HCDR3 in CoV-AbDab (a database of antibodies capable of binding to coronaviruses), including SEQ ID NO:51 and SEQ ID NO:52. [Figure 12A] 1 illustrates designed VEGF binder structures, according to some embodiments. [Figure 12B] 1 illustrates a wild-type binder structure, according to some embodiments. [Figure 13A] Illustrates sensorgrams according to several embodiments, including SEQ ID NOs: 41-48. [Figure 13B] 1 illustrates a sensorgram, according to some embodiments. [Figure 13C] 1 illustrates a sensorgram, according to some embodiments. [Figure 13D] 1 illustrates a sensorgram, according to some embodiments. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0008] The present disclosure addresses the need for artificial intelligence (AI) and machine learning (ML) models 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 involves using a de novo generative deep learning model to de novo design antibodies against three separate targets in a zero-shot fashion, with all designs being the result of a single round of model generation with no follow-up 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 for binding to human epidermal growth factor receptor 2 (HER2).

[0009] From these screens, the technology can further characterize the biophysics of 421 binders using surface plasmon resonance (SPR), finding that three bind more tightly than the therapeutic antibody trastuzumab. The binders are highly diverse and have low sequence identity with known antibodies. In addition, these binders score highly in our previously introduced naturalness index (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 are likely to have a desirable developability profile and low immunogenicity. These results pave the way for accelerated drug discovery against novel therapeutic targets using generative AI combined with high-throughput experiments.

[0010] A particularly challenging aspect of antibody drug discovery is the first step of lead candidate identification, due to the labor-intensive and uncontrollable nature of traditional screening methods. Generative AI-based de novo design has the potential to disrupt these shortcomings of the current drug discovery process. The zero-shot nature of the AI ​​design approach removes the need for tedious library screening to identify binding molecules, resulting in significant time and cost savings. Furthermore, the controllable nature of model-based design allows for the creation of proteins optimized for developability and immunogenicity properties, reducing downstream developability risks.

[0011] The present technology provides a step towards fully de novo antibody design by demonstrating the ability to generate novel antibody variants that confer comparable and in some cases better binding and native sequence properties than the parent antibody in a zero-shot fashion. The AI-generated sequences are different from any observed in the model training set, and the majority are different from known sequences in the Observed Antibody Space (OAS) database, yet still maintain a high naturalness score, indicating that the model can design antibody sequences along biologically feasible variants. Furthermore, the designed sequences are highly distinct from each other, demonstrating the ability to design a diverse set of binding molecule solutions. Furthermore, the inventors demonstrate progress in designing de novo multiple CDRs by creating and validating binders with up to three novel heavy chain CDRs using a modified multi-step approach. The present technology is generalizable, as demonstrated by the deployment of the present generative design method to distinct antigens. In addition, the development of epitope specificity across multiple antigens for antibody design may enable precise interactions with biologically relevant target regions associated with disease mechanisms of action. In addition to advances in generative modeling, the speed and scale of wet-lab validation of AI-generated designs will increase incrementally as the time and cost of DNA synthesis continue to decrease.

[0012] Antibodies are a growing class of therapeutic molecules, especially due to their attractive drug-like properties, including high target selectivity and minimal immunogenic effects. Antibody drug development generally begins with the discovery of an initial lead molecule. Existing approaches for lead discovery typically consist of randomly searching a large combinatorial sequence space by screening a large library of antibody variants against a target antigen. Techniques such as phage, yeast display, hybridoma screening or immunization combined with 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 antibody, and often produce suboptimal leads. Applying generative artificial intelligence (AI) to design de novo antibodies in a zero-shot and controllable manner, rather than screening and developing lead molecules, could significantly reduce the time and resources required for therapeutic antibody development.

[0013] 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 studies have shown that models trained on these data can be used for the de novo design of certain classes of proteins. These studies screen tens to thousands of protein designs representing 2-4 orders of magnitude fewer proteins than have been validated in our study. Moreover, despite the enormous therapeutic relevance of antibody-based therapeutics, which will account for 30% of FDA-approved biologics in 2022, we have yet to achieve de novo design of antibodies with wet-lab validation. De novo antibody design is particularly interesting because the major determinants of antibody function emerge from the complementarity determining regions (CDRs) of the sequence. These hypervariable regions directly interact with antigens, and among these, the heavy chain CDR3 (HCDR3) is often the most critical for binding, but also the most variable, making it particularly challenging to model. Several studies with experimental validation have attempted to optimize antibodies using supervised learning, but none have attempted zero-shot or de novo design.

[0014] Many groups have recognized the potential of zero-shot generative AI to impact antibody design. Several promising methods have recently emerged leveraging ideas from language modeling to geometric learning for antibody design. However, such methods have failed to demonstrate de novo antibody design in a zero-shot manner with laboratory validation. The present technology integrates the ideas of generative modeling with high-throughput experimental capabilities in wet labs. Recent advances in DNA synthesis and sequencing, E. coli.-based antibody expression, and fluorescence-activated cell sorting have made it possible to experimentally evaluate hundreds of thousands of individual designs rapidly and in parallel.

[0015] The technology demonstrates zero-shot antibody design with extensive wet-lab experiments. As a first step toward full de novo antibody design, the technology shows that HCDR3 can be designed in a generative AI manner using the model system of trastuzumab and its target antigen, human epidermal growth factor receptor 2 (HER2), as a model system. In some embodiments, all antibodies that bind to HER2 or homologs of HER2 can be removed from the training set. The technology includes de novo design of many (e.g., about 440,000 or more) unique HCDR3 variants of trastuzumab, which can be screened for binding to HER2 using a proprietary activity-specific cell enrichment (ACE) assay. As used herein, the term "quantitative affinity activity specific cell enrichment or qaACE assay" refers to a high-throughput assay for obtaining affinity and sequence data for biomolecular variants (U.S. Provisional Patent Application No. 63 / 371,474, filed August 15, 2022, and PCT / US23 / 60167, filed January 5, 2023, each of which is incorporated by reference in its entirety).

[0016] For example, quantitative affinity ACE ("qaACE") and ACE analysis, further described herein and known as "de novo ACE" or "dnACE", are methods for sampling antibody variant binding in a high throughput manner using flow cytometry and next generation sequencing. The main goal of this method is to generate high throughput binding information and / or training data for AI models to perform sequence-based binding predictions. This method can be applied to any antibody format, mab, fab, scFv, scFAB, VHH, nanobody, etc., and possibly other binding drug formats.

[0017] In one embodiment, the first step in the qaACE process is to generate a mutationally diverse antibody library that evenly samples the sequence space around the starting antibody molecule, containing variants that span a range of mutational distances from the original sequence.

[0018] In some embodiments, including the examples herein, the method provides flow cytometry readout from antibodies, expressed in SoluPro E.coli, bound to fluorescently labeled antigen probes. In the qaACE assay, the expression of antibody molecules is normalized so that changes in the fluorescent signal in cells are due to the different affinities of the expressed antibody variants in the cells that bind to the fluorescent antigen probe. This normalization is achieved through a common target molecule probe that binds to all variants and whose signal is in an orthogonal fluorescent channel to the antigen probe. In this setting, we show that the fluorescent signal of the variants is proportional to the measured KD of the antibody variant within a range. Given this proportionality, FACS can be used to sort cells containing antibody variants across a range (e.g., distribution) of affinities.

[0019] After sorting across a range of affinity values ​​with gating across the library population distribution, the cellular material is sequenced and quantified for the prevalence of observed variants across affinity gates (bins, tubes). The quantification is used to calculate an enrichment score for each variant. The enrichment scores generated via qaACE are an ideal data type for AI modeling purposes due to their accuracy and throughput.

[0020] In one exemplary workflow, the present disclosure provides a qaACE assay that includes some or all of the following general steps:

[0021] 1) SoluPro Generation of antibody or other drug molecule libraries for screening qaACEs expressed in host cells such as E. coli. 2) Identification of antigen or binding partner probes that are fluorescently labeled for use in, for example, FACS, throughout the early cell assay development process. 3) Use of a generic probe for a target molecule variant that allows detection of expression levels within cells, which is used to gate homogeneously expressing populations to disambiguate affinity and expression signals associated with epitope binding signals. 4) Sorting of cells across affinity distributions. 5) Sequencing of cells sorted across affinity distributions. 6) DNA barcodes or UMIs can be added via PCR amplification of the region of interest during sequencing. These UMIs will allow absolute quantification of variants obtained from the gate. 7) Generation of affinity correlated enrichment scores for each observed variant. 8) AI model training using enrichment scores and antibody variant sequences.

[0022] Additional ACE assay analyses are discussed further below.

[0023] From these designs, the technique can be used to functionally validate many (e.g., 421 or more) binders using SPR and to infer the presence of many more binders in total (e.g., about 4,000). Not only are the designed binder sequences different from those found in the training dataset, they are also very different and dissimilar to those previously observed in structural antibody databases or large datasets of known antibodies. Furthermore, according to the naturalness index mentioned above (Bachas S et al.), the designed binders are likely to be developable and have favorable immunogenic properties. We demonstrate the scalability of our approach by designing and validating binding molecules against two additional antigens: human vascular endothelial growth factor A (VEGF-A) and SARS-CoV-2 spike protein (COVID-19 Omicron variant).

[0024] While the primary focus of this work is 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 using a multi-step generative AI approach to design antibodies with any antibody domain / sequence that differs from that of the parent antibody (e.g., all three heavy chain CDRs (HCDR1, HCDR2, HCDR3), any light chain, etc.). Taken together, this work advances the path to rapid progress toward fully de novo antibody design using generative AI, which has the potential to transform the availability of therapeutics for patients.

[0025] This technology represents an important advancement in in silico antibody design that has the potential to transform the availability of effective therapeutics for patients. Generative AI-designed antibodies will significantly reduce development timelines by generating molecules with desired qualities 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, leading to safer and more effective treatments for patients than those enabled by traditional development approaches. The core platform of generative AI design methods and ultra-high throughput wet-lab screening capabilities will continue to drive progress in the field and unlock new capabilities in the rapidly accelerating field of protein therapeutic design.

[0026] Screening hundreds of thousands of model-generated sequences for binding The present technology leverages the ACE assay described above to screen large antibody variant libraries containing hundreds of thousands of members expressed in fragment antigen binding (Fab) format. The technology validates the ACE assay in a de novo discovery workflow by sampling sequences for follow-up analysis by SPR, the gold standard in binding affinity measurement and detection. Empirical evidence finds that the ACE assay can correctly classify binders with nearly 60% accuracy and >95% recall (Table S3). Table S3. Precision and recall of the binary ACE (bACE) assay. The bACE assay predicted 353 variants to bind by SPR, and 209 binders were confirmed (59.2% precision). An additional 4 binders were not predicted to bind by bACE (98.1% recall). [Table 1]

[0027] This allows for a powerful workflow in which a large set of predicted populations can be first screened by ACE assay, and then the predicted binding populations can be screened via SPR to remove false positives and collect high-quality binding affinity measurements (Figure 1A).

[0028] Exemplary Computer-Implemented Machine Learning Training and Computation 1B illustrates an exemplary computing environment 100 for training and / or computing one or more machine learning (ML) models, according to some embodiments. The environment 100 includes a client computing device 102, a structure prediction server 104, an assay device 106, and an electronic network 108. Some embodiments may include multiple client devices 102, multiple molecular modeling servers 104, and / or multiple assay devices 106. Generally, the one or more molecular modeling servers 104 operate to perform the training and computing of in silico molecular modeling, fully or partially, as described herein.

[0029] The client computing device 102 may be an individual server, a group (e.g., a cluster) of multiple servers, 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 included in a respective 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). In general, the processor is configured to execute software instructions stored in a memory. The memory may include one or more persistent memories (e.g., hard drives / 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 .

[0030] The client computing device 102 may include a respective input device and a respective output device. The respective input devices may include any suitable device 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 respective output devices may include any suitable device for communicating output, such as hardware speakers, a computer monitor, a touch screen, etc. In some cases, the input devices and the output devices may be combined 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 network interface controller, such as a wired / wireless controller (e.g., an Ethernet controller), to facilitate bidirectional / multiplexed networking over a network between the client computing device 102 and other components of the environment 100.

[0031] 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 further include a data repository 180. The data repository 180 may be a structured query language (SQL) database (e.g., MySQL database, Oracle database, etc.) or another type of database (e.g., a non-SQL only (NoSQL) database). In some embodiments, the data repository 180 may include a file system (e.g., EXT file system, Apple File System (APFS), Networked File System (NFS), local file system, etc.), an object store (e.g., Amazon Web Services S3), data flakes, etc. The data repository 180 may include multiple data types, such as pre-trained data, pre-trained data, and fine-tuning data obtained from public data sources (e.g., SAbDab data, OAS data). The fine-tuning data may be proprietary affinity data obtained from quantitative assays ACE, Carterra, or any other suitable source. Data repository 180 may contain machine learning model training data represented in any suitable data format, such as Protein Data Bank (PDB) format, JavaScript® Object Notation (JSON) format, eXtensible Markup Language (XML) format, etc.

[0032] The server 104 may include a library of client bindings for accessing the data repository 180. In some embodiments, the data repository 180 is located remotely from the structure prediction server 104. For example, the data repository 180 may be implemented using a RESTdb.IO database, Amazon Relational Database Service (RDS), or the like in some embodiments. In some embodiments, the structure prediction server 104 may include a client-server platform technology, such as Python, PHP, ASP.NET, Java J2EE, Ruby on Rails, Node.js, Web Services 77, or an online API, for receiving and responding to electronic requests. Furthermore, the structure prediction server 104 may include a set of instructions that may be integrated with the client-server platform technology for performing the machine learning operations discussed below.

[0033] The assay device 106 may be a surface plasmon resonance (SPR) machine, such as, for example, a Carterra SPR machine. The device 106 may be physically connected to either the structure prediction server 104 or the data repository 180, as shown. The device 106 may be located in a laboratory and may be accessible from one or more computers in the laboratory (not shown) and / or from the structure prediction server 104. The device 106 may generate data and upload the data to the data repository 180 directly and / or via a laboratory computer. The assay device 106 may include instructions for receiving one or more sequences (e.g., mutant sequences) and for synthesizing those sequences. The synthesis may be performed via another technique (e.g., via a different device or via a human). In some aspects, the device 106 may not be configured as a device, but as an alternative assay that may measure protein-protein interactions as listed in other sections of this application. For example, the device 106 may instead be configured as a series of devices / workflows including plate and liquid handling. In general, device 106 may be replaced with suitable hardware and / or software, optionally including a human operator, to generate affinity data.

[0034] 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). The network 108 may enable, for example, bidirectional communication between the client computing devices 102 and the structure prediction server 104.

[0035] Processor 150 may include any suitable number of processors and / or processor types, such as one or more graphic 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 memory) and stores one or more sets of computer-executable instructions / modules 160, including an input / output (I / O) module 162, a machine learning training module 164, and a machine learning computation module 166. In some aspects, more or fewer modules may be included, and in some aspects, one or more of the models may be combined or aggregated into fewer modules.

[0036] Each of the modules 160 implements certain functions related to the present technology, as described further below. The modules 160 may store machine-readable instructions, including one or more applications, one or more software components, and / or one or more APIs, that may be implemented to facilitate or perform features, functions, or other disclosures described herein, such as any method, process, element, or limitation, as illustrated, shown, or described with respect to various flowcharts, examples, diagrams, figures, and / or other disclosures herein. In some aspects, multiple modules 160 may implement certain technologies in a coordinated manner. For example, the machine learning computing module 166 may read information from one or more other models before, during, and / or after the inference operation is started. Thus, the models 160 may exchange data via suitable technologies, such as, for example, inter-process communication (IPC), representational state transfer (REST) ​​APIs, etc., within a single computing device, such as the structure prediction server 104. 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 over a network, such as network 108. Module 160 of FIG. 1B will now be described in more detail.

[0037] In general, the I / O module 162 includes instructions that allow 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 uses the ML training model 164 to train one or more ML models in preparation for using one or more trained ML models to generate outputs used in an antibody prediction project, a docking complex prediction project, and / or an affinity value prediction project. Once 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 to initiate a molecular modeling process. The I / O module 162 may include instructions for generating one or more graphical user interfaces (GUIs) (not shown) that collect and store parameters related to biomolecular modeling, such as specific reference proteins, biomolecules, binding partners, etc., from lists stored in the data repository 180.

[0038] 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 in a computer-usable storage medium, or a tangible non-transitory computer-readable medium (e.g., a standard random access memory (RAM), an optical disk, a universal serial bus (USB) drive, etc.) having such computer-readable program code or computer instructions embodied therein, and the computer-readable program code or computer instructions may be installed on or otherwise adapted to be executed by the processor 150 (e.g., in operation in connection with a respective operating system in memory 154) to facilitate, implement, or perform the machine-readable instructions, methods, processes, elements, or limitations as illustrated, depicted, or described with respect to the various flowcharts, illustrations, diagrams, figures, 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, bytecode, interpretable source code, etc. (e.g., via Golang, Python, C, C++, C#, Objective-C, Java, Scala, ActionScript, JavaScript, HTML, CSS, XML, etc.).

[0039] The technology may include a generative model for automatically designing or improving antibodies against a particular target of interest (e.g., a target biomolecule). The model may be trained on antibody-antigen structures or a large collection of existing antibody sequences. The model may also be trained using unique high-throughput binding interaction data. The model may, in some embodiments, be validated by designing an antibody that binds to a receptor protein involved in cancer. The technology may allow for zero-shot generation of antibodies tailored to bind to any target. The technology may, in some embodiments, allow for redesigning the complementarity determining regions (CDRs) of an antibody, which confers binding specificity to the target. The generative model may automatically design or improve antibodies against a particular target of interest, and may be pre-trained on a collection of antibody-antigen complexes mined from public data sources (e.g., Protein Data Bank).

[0040] In some embodiments, a model trained according to the present technology can be trained to receive an antigen structure and predict the structure, sequence, and / or affinity of an antibody based on the received antigen. In some embodiments, the prediction can be performed by iteratively refining the model's prediction of the structure while autoregressively deciphering the sequence. In such cases, generalized autoregressive or bidirectional training can be used to capture long-range patterns of the sequence. The antibody and antigen modules can capture independent SE(3) equivalence, allowing data-efficient protein-protein rigid-body docking.

[0041] In some embodiments, the high GPU memory of 80GB A100 can be used to achieve the deciphering of long structural graphs. A scalable implementation of pairwise SE(3) equivalence can be useful in some embodiments to improve training efficiency. After pre-training, the technology can fine-tune the model to predict the binding affinity of the antibody to the antigen of interest. The affinity prediction module can predict the binding affinity of the antibody to the antigen using the structural and sequence features inferred by the model. In particular, embeddings at each residue can be inferred and propagated to the affinity prediction module using attention. The affinity prediction module can combine high-throughput binding interaction data (e.g., data generated on Absci's Integrated Drug CreationTM platform). Thus, advantageously, the technology achieves a novel coupling of the structural model of antibody binding with real-world binding affinity data.

[0042] The technology may validate the trained model by designing antibodies to neutralize Her2, an important target in breast cancer. Starting with trastuzumab, a well-understood Her2 binder, the technology may include designing a library containing trastuzumab variants with primary, secondary, and hypermutations. The binding affinity of the variants in the library may be measured by high-throughput screening, and the library may be scored using a generative model to demonstrate high correlation between the model predictions and the high-throughput data. In addition to or alternatively to pre-training on structural data, the technology may pre-train a language model on a large database of existing antibody sequences. The language model and the predictive model may be effectively assembled to further improve prediction of binding affinity data. After training, the model may generate diverse new antibodies against unseen antigens by iteratively decoding the sequences in an autoregressive manner. In contrast to previous work on deep learning-based protein design, in which proteins are blindly designed, our model allows for targeted design against specific antigens.

[0043] In some aspects, the ML training module 164 may include a set of computer-executable instructions implementing 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 the data repository 180 and / or memory 154, which are accessible to or otherwise communicatively coupled to the structure prediction server 104.

[0044] For example, the ML training module 164 may train one or more ML models (e.g., artificial neural networks (ANNs)). One or more training datasets may be used for model training in the present technology, as discussed herein. The input data may have a particular shape that may affect the ANN network architecture. The elements of the training dataset may include tensors that are scaled to small values ​​(e.g., in the range of (-1.0, 1.0)). In some aspects, a pre-processing layer may be included in the training (and computation) that applies principal component analysis (PCA) or another technique to the input data. PCA or another dimensionality reduction technique may be applied during training to reduce the dimensionality from a high number to a relatively small number. Reducing the dimensionality may result in a substantial reduction in the computational resources (e.g., memory and CPU cycles) required to train and / or analyze the input data.

[0045] In general, training an ANN may include establishing a network architecture, or topology, and adding layers including an activation function, a loss function, and an optimizer for each layer (e.g., “leaky” rectified linear unit (ReLU), softmax, hyperbolic tangent, etc.). In one embodiment, the ANN may use different activation functions at each layer or between the hidden layer and the output layer. Suitable optimizers may include Adam and Nadam optimizers. In one embodiment, 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 split into training, validation, and test data. For example, 20% of the training dataset may be retained for later validation and / or testing. In that example, 80% of the training dataset may be used for training. In that example, the training dataset data may be shuffled before being so split. Splitting 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 coded in an N-dimensional tensor, array, matrix, and / or other suitable data structure. In some embodiments, training may be performed by continuous evaluation (e.g., loop) of the network using labeled training samples. The process of training the ANN may cause the weights or parameters of the ANN to change. The weights may be initialized to random values. The weights may be adjusted as the network is continuously trained to converge the values ​​output by the network to expected or "learned" values ​​by using one or more gradient descent algorithms. In one embodiment, regression without an activation function may be used, in which the input data may be normalized by mean centering, and a mean squared error loss function, in addition to the mean absolute error, may be used to determine an appropriate loss and to quantify the accuracy of the output.

[0046] In some embodiments, 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 known as transfer learning and may enable training of a base model that is universal in the sense that it can be used as a common grammar for all antibody sequences, for example. In some examples, pre-training may be used to train multiple models of independent artificial neural networks and / or multiple respective layers of a single artificial neural network (e.g., an artificial neural network used to predict affinity values ​​from biological structure information). The term "pre-training" may be used to describe scenarios in which a second training may occur (i.e., when a model may be "fine-tuned"). Transfer learning refers to the ability of a model to leverage the results (weights) of a first pre-training to better initialize a second training, which may otherwise require random initialization. The second training, i.e., fine-tuning, may be performed using affinity data as discussed herein. The technique of combining pre-training and fine-tuning advantageously boosts performance in that the results of training on affinity data perform better after pre-training than when pre-training is not performed (e.g., using natural antibody structures from SAbDab). Model fine-tuning may be performed in some embodiments for a given antibody-antigen pair. In some embodiments, ML model self-supervised learning may be performed during pre-training to give the model an understanding of antibody grammar.

[0047] In general, the ML model may be trained as described herein using supervised, semi-supervised, or unsupervised machine learning programs or algorithms. The machine learning programs or algorithms may employ neural networks, which may include one or more of graph convolutional neural networks, message passing neural networks, geometric vector perceptron networks, recurrent neural networks, deep learning neural networks, convolutional neural networks, deep learning neural networks, transformers, autoencoders, and / or combined modules or programs that learn on two or more features or feature datasets (e.g., structured data, unstructured data, etc.). The machine learning programs or algorithms may also include natural language processing, semantic analysis, automated inference, regression analysis, support vector machine (SVM) analysis, decision tree analysis, random forest analysis, K-nearest neighbor analysis, naive Bayes analysis, clustering, reinforcement learning, and / or other machine learning algorithms and / or techniques (e.g., generative algorithms, genetic algorithms, etc.).

[0048] In some aspects, an ML algorithm or technique may be selected for a particular input based on the program set size of the input. In some aspects, the artificial intelligence and / or machine learning based algorithm may be based on or otherwise incorporate aspects of one or more machine learning algorithms included as libraries or packages executed on the server 104. For example, the libraries may include TensorFlow-based libraries, Pytorch libraries (e.g., PyTorch Lightning), Keras libraries, Jax libraries, the HuggingFace ecosystem (e.g., transformer, dataset, and / or tokenizer libraries therein), and / or scikit-learn Python libraries. However, these popular open source libraries, while convenient, are not required. The techniques may be implemented using other frameworks / languages.

[0049] Machine learning may involve identifying and recognizing patterns in existing data (e.g., structural information, docking functions, binding affinities, etc.) to facilitate making predictions, classifications, and / or identifications for subsequent data (e.g., using the trained model to generate antibodies based on the input antibodies, predict the ability of the generated antibodies to dock to the input antigen (or other antigens), and / or predict the binding affinity of the generated antibodies, etc.). Machine learning models can be created and trained based on example data (e.g., "training data") inputs or data (which may be referred to as "features" and "labels") to make valid and reliable predictions for new inputs. In supervised machine learning, a machine learning program running on a server, computing device, or other processor may provide example inputs (e.g., "features") and their associated or observed outputs (e.g., "signals") for the machine learning program or algorithm to determine or discover rules, relationships, patterns, or other machine learning "models" that map the example inputs (e.g., "features") to outputs (e.g., signs), for example, by determining and / or assigning weights or other metrics to the model across its various feature categories. Such rules, relationships, or other models may then be provided subsequent inputs for the model to run on the server, computing device, or other processor to predict expected outputs based on the discovered rules, relationships, or models.

[0050] For example, the ML training module 164 may analyze the labeled data at an input layer of a model having 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 may be, for example, structural information of an antibody. In some embodiments, the output may be a sequence of a new protein. During training, the labeled data may be propagated through one or more connected deep layers of the ML model to establish weights of one or more nodes or neurons of the respective layers. Initially, the weights may be initialized to random values, and one or more suitable activation functions may be selected for the training process, as will be understood by those skilled in the art. The ML training module 164 may include training an output layer of each of the one or more machine learning models. The output layer may be trained to output a prediction. For example, the ML model trained herein may predict structural information of an antibody by analyzing the labeled examples provided during training. In some embodiments, the binding affinity may be expressed as a real number (e.g., in a regression analysis). In some embodiments, multiple ANNs may be trained and / or computed separately. For example, individual models can be fine-tuned (i.e., trained) based on a pre-trained model using transfer learning for multiple different antibody-antigen pairs.

[0051] In unsupervised or semi-supervised machine learning, a server, computing device, or other processor is required to find its unique structure in the unlabeled example input, for example, multiple training iterations are performed by the server, computing device, or other processor to train the generation of multiple models until a satisfactory model is generated. In the present technology, semi-supervised learning can be used, among others, for natural language processing purposes and to learn the grammar of antibody sequences using objects such as masked language model objects. Supervised learning and / or unsupervised machine learning can also include retraining, retraining, or otherwise updating the model with new or different information, which may include information received, captured, generated, or otherwise used over time. In various aspects, training the ML model herein can include generating an ensemble model that includes multiple models or sub-models, including models that are trained by the same and / or different AI algorithms described herein and configured to operate together.

[0052] Once model machine learning training module 164 initializes one or more ML models, which may be ANNs or recurrent networks, for example, model machine learning training module 164 trains the ML models by inputting labeled data into the models (e.g., antibody / antigen complexes, docking biomolecules, etc.) The trained ML models provide accurate predictions for inputs that the models have not seen before (i.e., not used during training).

[0053] The model machine learning training module 164 may split the training data into respective training and test data sets. The model machine learning training module 164 may train the ANN using the labeled data. The model machine learning training module 164 may use the test data to calculate accuracy / error metrics (e.g., cross entropy) and test the corresponding set of labels. The model machine learning training module 164 may serialize the trained models and store the trained models in a database (e.g., data repository 180). Of course, it will be understood by those skilled in the art that the model machine learning training module 164 may train and store more than one model. For example, the model machine learning training module 164 may train a separate model for performing antibody generation, another model for performing antibody docking, and yet another model for performing affinity prediction. It should be understood that depending on the embodiment, the structure of the described network may vary. For example, in some aspects, antibody generation, antibody docking, and affinity prediction may each correspond to a respective layer of a monolithic machine learning model. For example, in some embodiments, the trained ML model predicts many pieces of data: antibody sequence, antibody affinity, docking configuration. The present technology allows supervision of any one of these components, and the remaining components can be improved through training. This ability is an advantageous result of the end-to-end training of the present technology.

[0054] The machine learning computing module 166 may include a set of computer-executable instructions that implements machine learning loading, configuration, initialization, and / or computing functions. The ML computing module 166 may include instructions for storing trained models (e.g., in the electronic data repository 180, as pickled binaries, etc.). Once trained, the trained ML model may be operated in an inference mode, where the model may output one or more predictions, classifications, etc., when provided with de novo inputs for which the model has not been previously provided, as described herein. In some aspects, a loss minimization function may be used, for example, to train the ML model to generate outputs that resemble known outputs (i.e., ground truth examples).

[0055] Once a model has been trained by the model machine learning training module 164, the model computation module 166 may load one or more trained models (e.g., from the data repository 180). The model computation module 166 generally applies new data to the trained model that the trained model has not previously analyzed. For example, the model computation module 166 may load a serialized model, deserialize the model, and load the model into memory 154. The model computation module 166 may load new data (e.g., binding partner structure information) that was not used to train the trained model. For example, the new data may include data described herein (e.g., antigen sequence data, etc.) coded as input tensors. The model computation module 166 may apply one or more input tensors to the trained ML model. The model computation module 166 may receive outputs (e.g., tensors, feature maps, etc.) from the trained ML model. The output of the ML model may 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 may include the docking complex and binding affinity values. In this way, the present technology freely provides a means to, for example, quantitatively design a target antibody corresponding to an input target antigen, as well as to quantify the docking of the input antigen and the designed target antibody and to quantify the binding affinity of the two. These technologies are much more accurate and data-rich than traditional industry practices. By using ML, the present technology avoids the expensive and time-consuming process of laboratory experiments due to in silico performance without the need for the continuous use of wet labs.

[0056] The model computation module 166 may be accessed by another element (e.g., a web service) of the structure prediction server 104. The ML computation module 166 may pass its output to another module for further processing / analysis. In some embodiments, a user may interact with the ML model during training and / or computation using command line tools, application programming interfaces (APIs), software development kits (SDKs), Jupyter notebooks, etc.

[0057] With respect to modules 160, it will be understood by those skilled in the art that in some aspects the software instructions comprising modules 160 may be organized differently and more / fewer modules may be included. For example, one or more of modules 160 may be omitted or combined. In some aspects additional modules may be added (e.g., a localization module). In some aspects software libraries (e.g., Python code) implementing one or more modules may be combined such that, for example, ML training module 164 and ML computation module 166 are a single set of executable instructions used to perform training and prediction.

[0058] Exemplary ACE Assay Analysis In operation, the assay device 106 may be used to generate binding scores from the leads. For example, in one embodiment, the following procedure may be used.

[0059] 1. Paired-end reads can be merged with the maximum allowed overlap set according to amplicon size and sequencing read length (e.g., 150 bases for all libraries described herein) using FLASH2. 2. Primers could be removed from both ends of merged reads using the cut-adapt tool, discarding reads where primers were not detected. 3. Reads can be aggregated across all FACS sorting gates and aligned to a reference sequence (parental version of the amplicon) in amino acid space. Alignment can be performed using the Needleman-Wunsch algorithm implemented in Biopython. In an 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 embodiments, parameters can be selected by manual inspection across several processed libraries. 4. Reads can then be discarded if (1) the average base quality falls below a certain number (e.g., 20) or (2) the sequence (in DNA space) is less than the number of reads that span all gates (e.g., 10). 5. The present technology may further include flagging (1) sequences that align with the reference with low score (e.g., defined as less than 0.6 of the score obtained by aligning the reference to itself), (2) sequences that contain stop codons outside the region of interest, and (3) sequences that contain frameshift insertions or deletions. In some embodiments, the flagged sequences may not be included in any mutation-related statistics, but may be used for count normalization for binding score calculation. FastQC and MultiQC may be used to generate sequencing quality control indicators. 6. For each gate, the occurrence rate 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., millions). 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. In all experiments, weights can be assigned linearly using an integer scale, with the gate capturing the lowest fluorescent signal assigned a weight of 1, the next lowest a weight of 2, and so on. 8. Any detected sequences that were not present in the originally designed and synthesized library can be dropped. 9. ACE scores can be averaged across independent FACS sorts, dropping sequences with standard deviations of replicates greater than 1.25. Amino acid variants can only be retained if the technique collects a number (e.g., at least three) of independent QC passing observations between the synonymous DNA variants and replicate FACS sorts.

[0060] Exemplary binary assay (dnACE) analysis Enrichment scores can be calculated for each variant screened by the binary version of the ACE assay. For example, in some embodiments, the following procedure can be used.

[0061] 1. Paired-end reads can be merged using Fastp with quality filtering and base correction in the merged regions enabled. 2. Primers could be removed from both ends of merged reads using the cut-adapt tool, discarding reads where primers were not detected. 3. Unique sequences can be tallied to provide a raw count of each variant observed in each sample. Sequences that do not match the designed sequences in the library can be discarded. 4. For each sample, the relative abundance can be calculated for each variant. The enrichment score can be calculated by dividing the relative abundance of each variant in the gate by its relative abundance in the unsorted library sample.

[0062] Calculating the expected binding percentage for the ACE assay The ACE score is discretized into a set of 6 bins, and for each bin, the set of variants belonging to the bin are screened via SPR. The percentage of variants in a bin that bind is used to estimate the expected binding rate for that bin. These values ​​can be found in Table S2. Table S2. SPR binding rate corresponding to ACE score of sequences in HDL2. SPR binding rate increases as sequences fall into higher ACE bins. This validates the ACE assay as a useful method to select binders (i.e., select sequences from the highest gates). The data can also be used to estimate the total number of binders in each bin. [Table 2]

[0063] Exemplary Surface Plasmon Resonance (SPR) Embodiments As mentioned above, SPR can be used for a variety of techniques, including functional validation of binders.

[0064] surface preparation For example, in some embodiments, post-induction samples may be transferred to a plate (e.g., a 96-well plate (e.g., Greiner Bio-One)), pelleted, and lysed in 50 μL of lysis buffer (e.g., 1X BugBuster Protein Extraction Reagent containing 0.01KU Benzonase Nuclease and 1X Protease Inhibitor Cocktail). The plate may be incubated at 30°C for 15-20 minutes and then centrifuged to remove insoluble debris. After lysis, samples may be adjusted to a final volume of 260 μL with 200 μL of 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 filtered into a 96-well plate. The lysed samples can then be transferred from the 96-well plate to a larger plate (e.g., a 384-well plate) for high-throughput SPR, for example, using a Hamilton STAR automated liquid handler. Colonies can be prepared in two sets of independent replicates prior to lysis, with each replicate measured in two separate experimental runs. In some examples, a single replicate can be used, as shown.

[0065] SPR High-throughput SPR experiments were performed 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, The experiments may be performed on a microfluidic Carterra LSA SPR instrument using 100 mM NaCl, 3 mM EDTA, 0.01% w / v Tween-20). Carterra LSA SAD200M chips may be pre-functionalized with 20 μg / mL of biotinylated antibody capture reagent for 600 seconds before performing the experiment. Lysed samples in a 384-well block may be immobilized on the chip surface for 600 seconds, followed by a 60 second wash step for baseline stabilization.

[0066] Antigen binding may be performed using a non-regeneration kinetic method with a 300 s association phase followed by a 900 s dissociation phase. For analyte injection, six leading blanks may be introduced to create a consistent baseline before monitoring antigen binding kinetics. After the leading blanks, five concentrations of HER2 extracellular domain antigen (e.g., ACRO Biosystems, prepared in 3-fold serial dilutions from a starting concentration of 500 nM) may be injected into the instrument and the time series response recorded. In some experiments, measurements of individual DNA variants may be repeated four times. Each experimental run may consist of two complete measurement cycles (ligand immobilization, leading blank injection, analyte injection, chip regeneration), which may provide two duplicate measurement attempts per clone in one run. In some experiments, technical replicates measured in separate runs may further double the number of measurement attempts per clone to four.

[0067] Exemplary Sequencing Techniques Low diversity library sequencing To identify the DNA sequence of each antibody variant evaluated by SPR, duplicate plates can be provided for sequencing. A portion of the pelleted material can be transferred to a 96-well PCR (e.g., Thermo-Fisher) plate via a pinner (e.g., Fisher Scientific) that can contain reagents for performing an initial phase PCR of a two-phase PCR for the addition of Illumina adapters and sequencing. The reaction volume used can be, for example, 12.5 μl. During the initial PCR phase, partial Illumina adapters can be added to the amplicon via four PCR cycles. The second phase PCR can add the remaining portion of the Illumina sequencing adapters, as well as Illumina i5 and i7 sample indexes. The initial PCR reaction can use, for example, a UMI primer concentration of 0.45 μM, 6.25 μl of Q5 2× master mix (NEB). Reactions may be first denatured at 98° C. for 3 minutes, followed by 4 cycles of 98° C. for 10 seconds, 59° C. for 30 seconds, and 72° C. for 30 seconds, with a final extension of 2 minutes at 72° C. After the initial PCR, 0.5 μM secondary sample index primer may be added to each reaction tube.

[0068] The reaction can then be denatured at 98°C for 3 minutes, followed by 29 cycles of 98°C for 10 seconds, 62°C for 30 seconds, and 72°C for 15 seconds, with a final extension at 72°C for 2 minutes. The reaction can then be pooled into 1.5 mL tubes (Eppendorf). The pooled samples can be size-selected using 1x AMPure XP (Beckman Coulter) bead procedure. The resulting DNA samples can be quantified by Qubit fluorometer. The pool size can in some embodiments be verified via Tapestation 1000 HS and sequenced on an Illumina MiSeq Micro (2x150nt) for HCDR3 libraries, or an Illumina MiSeq Reagent Kit v3 (2x300nt) for HCDR1-HCDR3 libraries with 20% PhiX.

[0069] After sequencing, amplicon reads can be merged using Fastp and trimmed by cutadapt

[44] to enumerate each unique sequence. A custom R script can then be applied to calculate the sequence frequency ratio between the most abundant sequence and the second most abundant sequence in each sample. Levenshtein distances can be calculated between the two sequences. These distance values ​​can be used for downstream filtering to ensure that clonal populations have been measured by SPR. Dominant sequences in each sample can be compared to the designed sequences and discarded if they do not match any expected sequences. Dominant sequences can then be combined with their companion Carterra SPR measurements.

[0070] Exemplary Computer-Implemented Method In some embodiments, the techniques may include training one or more machine learning models to generate target biomolecular structures and / or sequences.

[0071] 2A illustrates an exemplary computer-implemented method 200 for training a machine learning model to generate structural information for a biomolecule, according to some embodiments. The method 200 may include receiving one or more training inputs (block 202), 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.

[0072] In general, the one or more training inputs may include data representing one or more of: (i) amino acids, (ii) peptide sequences, (iii) polypeptide sequences, (iv) primary sequences, (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.

[0073] The method 200 may include processing one or more training inputs with a machine learning biomolecular prediction model to generate predicted biomolecular structural information (block 204). The method 200 may include evaluating a loss function that compares the predicted biomolecular structural information to ground truth values ​​(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). The method 200 may be implemented in some embodiments by the ML training module 164 of FIG. 1.

[0074] In some embodiments, comparing the predicted biomolecular structure information to ground truth values ​​to evaluate a loss function to train the machine learning model in block 206 may include configuring a second model to act as a critic to train the machine learning model to predict according to the ground truth data. Such techniques as adversarial architectures (e.g., generative adversarial networks) may be used in this context.

[0075] In some aspects, the biomolecule may be an antibody or a protein. In some aspects, the biomolecule structural information includes three-dimensional coordinates of a 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 one or more of (i) a Protein Data Bank (PDB) data format, (ii) a JSON data format, or (iii) an XML data format, respectively. Other suitable formats may be used in some aspects. In some aspects, the biomolecule binding partner is a protein and the biomolecule binding partner structural information includes a three-dimensional structure. In some aspects, the method 200 may further include receiving the training input from a database including antibody structures and / or antigen structures. For example, the training input may be received / obtained from a structural antibody database (SAbDab). It will be appreciated by those skilled in the art that data processing steps may be useful for implementing the present technology. For example, to assess whether an ML model can generalize to a specific binding partner (e.g., an antigen) that is very different from the one the model was trained on, all biomolecules in the training set that are within 40% sequence identity with any protein in the corresponding validation and testing sets can be removed.

[0076] FIG. 2B illustrates a data flow block diagram 220 for training and computing one or more machine learning models according to some aspects of the present technology. In FIG. 220, the antibody generation step occurs at block 222a. During training, as shown, training information flows from the antibody generation block 222a to the docking block 222b, where the docking step occurs. Training information also flows from the docking block 222b to the affinity prediction block 222c, where affinity prediction occurs. During training, the training signal also flows in the reverse direction, as shown. For example, FIG. 2B may illustrate the flow of data during training using the method 200 of FIG. 2A. In some aspects, the training and / or computing of one or more antibody generation machine learning models may 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 appreciated by those of skill in the art, this technique lacks the use of an input binding partner (e.g., an antigen). In some embodiments, 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].

[0077] FIG. 2C illustrates a computer-implemented method 230 for generating structural information of a target biomolecule according to some embodiments. 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 further include predicting structural information of the target biomolecule by processing the target input with a biomolecule prediction model (block 234). In some embodiments, 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 embodiments, the machine-learned biomolecule prediction model of method 200 and / or 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.

[0078] Typically, the structural information of the target biomolecule includes (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.

[0079] Method 230 may include storing (e.g., in memory 154, data repository 180) and / or outputting data representing a portion or all of the structural information of the biomolecule. For example, the data may represent one or more of: (i) amino acids of the target biomolecule, (ii) a peptide sequence of the target biomolecule, (iii) a polypeptide sequence of the target biomolecule, (iv) a 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) three-dimensional coordinates of the primary sequence of the 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 molecule 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 completed. In some aspects, the method 230 may be performed iteratively.

[0080] In some aspects, the method 230 may include receiving input parameters specifying a desired length of amino acids of the target biomolecule and generating a target biomolecule having the desired length. In some aspects, a machine-learned biomolecule prediction model may determine the length by predicting a sequence end token. In some aspects, the 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 antigen epitope, and / or (iii) a known three-dimensional structure of the 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) via random sampling. Advantageously, the technology is not limited to predicting in any fixed order.

[0081] In some embodiments, the method 230 may further include computing and / or training a machine-learned docking model trained to generate a docking complex from two or more input three-dimensional biomolecules. In some embodiments, the method 230 may further include processing the target input and predicted structural information of the target biomolecule using the docking model to generate a docking complex including the target input and the target biomolecule. In some embodiments, the method 230 may include providing the docking complex as a docking complex output. In some embodiments, the method 230 may further include computing and / or training a trained machine-learned binding affinity prediction model to predict binding affinity from the input docking complex. The method 230 may include processing the docking complex using the trained affinity prediction model to generate a binding affinity value. In some embodiments, the machine-learned model is an artificial neural network. In some embodiments, 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 embodiments, the method 230 may include summarizing the node information contained within the artificial neural network using a pooling operation to generate a binding affinity value. In some embodiments, the method 230 may include providing the docking complex as a docking complex output. In some embodiments, the machine-learned model may be trained to predict binding specificity in addition to or instead of binding affinity.

[0082] In some embodiments, the method 230 may include training a machine-learned biomolecule prediction model by providing sequence data as input to supervise the biomolecule prediction model, training a docking model by providing antibody structure data as input to supervise the docking model, and / or training an affinity prediction model by providing sequence data as input to supervise the affinity prediction model. Specifically, the biomolecule prediction model may be designing one or more new antibodies. In the early stages of training, the design ability may not be high performance (e.g., the model may never reach a point where the training signal is significant). With supervised supervision, the problem becomes easier. At each stage, the model is trained to gaze at the last ground truth position, making the model focus on the information that is most important in the supervised training stage.

[0083] The model may be trained using an antibody / antigen structure (e.g., an antibody / antigen complex in a protein data bank such as SAbDab). The model may be configured to issue an initial guess for the drug molecule or antibody. The guess may be generated by instructions that iteratively output one residue or structure at a time based on an input binding partner (e.g., an antigen). Initially, the trained model may predict a first structure of a target biomolecule whose corresponding sequence is unknown but whose target antigen is known (e.g., provided as an input to the trained model). The ML model may iteratively predict a first amino acid, update the structure of the target biomolecule, then the second amino acid, and so on, until the structure of the target biomolecule is completed or another stopping condition occurs. The stopping condition may include the trained model making a prediction for each corresponding input, convergence criteria, filling in a portion of a CDR, etc. The prediction of the trained ML model may work in any direction, including (i) N-terminal to C-terminal order, (ii) C-terminal to N-terminal order, or (iii) via random sampling. The antigen provided as input to the trained ML model may be the primary amino acid sequence, the antigen epitope, and / or a known antigen three-dimensional structure, optionally including the antigen and epitope. The model may design the entire structure, or a portion thereof (e.g., CDR3). However, empirical studies may demonstrate that good results have shown that coarse-graining residues or averaging residues may perform poorly, i.e., not equivalent to aspects of the technology that analyze the entire structure (e.g., the entire antibody).

[0084] In some aspects, the method 230 may include training a machine-learned affinity prediction artificial neural network, including (i) one or more biomolecular prediction layers trained to predict biomolecular 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 from the input docking complexes. The one or more biomolecular prediction layers, the one or more docking layers, and the one or more affinity prediction layers may be connected (fully or otherwise). 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, and processing the target input using the affinity prediction artificial neural network to generate a docking complex and a corresponding structure affinity value corresponding to the target input. In some aspects, the method 230 may include receiving a sequence affinity prediction value from an affinity sequencing prediction model, and averaging the structure affinity values ​​with the sequence affinity prediction value to generate an ensemble affinity prediction value.

[0085] In some embodiments, method 230 may include training a machine-learned affinity predictive artificial neural network (or layers thereof) to generate diversity. For example, given an antigen, the predictive ANN may generate millions of antibodies. To rank / filter them and select the ones the inventors believe are most likely to work, the techniques may include ranking the antibodies using other techniques, including those described in U.S. Provisional Patent Application No. 63 / 297,679, the entirety of which is incorporated herein by reference.

[0086] In some aspects, the method 230 may include pre-training one or more biomolecular prediction layers using bound or unbound structures, pre-training one or more biomolecular 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, the method 230 may include receiving an external computational docking complex corresponding to the target input and comparing the external computational docking complex with the generated docking complex. Thus, the present techniques may be advantageously used in validation processes. In some aspects, the method 230 may include affinity training data derived from an assay (e.g., an ACE assay). The affinity training data may include an affinity score proportional to activity. The method 230 may include controlling the gradient flow of the machine-learned affinity prediction artificial neural network by applying a stopping gradient function in at least one of the one or more biomolecular prediction layers, the one or more docking layers, or the one or more affinity prediction layers. For example, in PyTorch, stopping gradients can be implemented using ".detach()", which detaches the associated variables from the computation graph, thereby creating a new graph. In aspects where the same network weights are applied to each iteration, stopping gradients can be employed such that each iteration is a separate graph. The gradients can then be averaged. In this context, conceptually, stopping gradients can be viewed as a strategy for augmenting limited data.

[0087] In general, method 230 may be used to perform de novo discovery of biomolecules (e.g., drug candidates such as antibodies) in silico. FIG. 3A illustrates a block flow diagram of an exemplary computer-implemented method 300 for performing de novo in silico biomolecule discovery, according to some embodiments. Method 300 may include receiving a scaffold and an input target (block 302a). Method 300 may include analyzing the scaffold and the input target (e.g., via method 200) and generating one or more biomolecules (e.g., one or more antibodies) using a trained machine learning model (e.g., via method 230). Method 230 may further include analyzing the output of the machine learning model at block 302b using a lead optimization machine learning model (block 302c). The lead optimization machine learning model may generate one or more candidates that method 230 may validate (block 302d). Method 230 may further include generating one or more drug candidates and / or one or more respective manufacturing cell lines (block 302e). In Figure 3A, the ML model in block 302(b) that designs biomolecules (e.g., one or more antibodies) against a target of selection (e.g., a target antigen) may include binders. These binders may be input into a machine learning model in block 302c to further optimize potential candidates.

[0088] FIG. 3B illustrates an exemplary block diagram 304 according to some embodiments. The blocks of the block diagram 304 may correspond to various elements of the computing environment 100 of FIG. 1, for example. The block diagram 304 includes an antibody generation block 306a. In block 306a, the antibody generation and structurally tailored antibody design may be performed. For example, the block 306 may correspond to training / operating one or more biomolecule generation machine learning models in block 222a of FIG. 2B. The block diagram 304 includes an end-to-end docking block 306b, which may include geometric learning (i.e., graph-based learning) in some embodiments. For example, the block 306b may correspond to training / operating one or more docking (e.g., antibody-antigen docking) machine learning models in block 222b of FIG. 2B. The block diagram 304 may include an affinity (e.g., binding affinity) prediction block 306c, which may perform graph-based affinity prediction. In some embodiments, block 306c may correspond to affinity prediction machine learning model training / computation in block 222c of FIG. 2B. In some embodiments, block diagram 304 may include a data source block 306d, which may include both public data sources (e.g., public data banks) and proprietary assay data for performing high-throughput experiments. Block 306e illustrates a computer architecture that may be used to implement the computer-implemented techniques of the present invention. For example, block 306e illustrates a GPU that may correspond to GPU 150 of FIG. 1.

[0089] FIG. 3C illustrates an exemplary CDR3 in silico discovery input and output block diagram 308. Block diagram 308 includes an objective (block 310a). In the illustrated example, the objective is to design a CDR3 targeting epidermal growth factor receptor (EGFR) that works well in a VHH scaffold (see, e.g., https: / / www.rcsb.org / structure / 4krl). However, one of ordinary skill in the art would understand that other objectives are envisioned, including other binding partners (e.g., other antigens, epitopes, etc.) and other scaffolds. Block diagram 308 further includes a target input (block 310b), including a target input including one or more of a target binding partner sequence, a target binding partner, or a target epitope. In block 310b, one or more trained ML models can analyze the target input to predict structural information. Starting with an EGFR antigen and a given scaffold (e.g., a VHH), a model can be trained to generate a CDR3 of interest. At this point, the model may never have seen a similar CDR3 or a similar antigen (i.e., it has not been provided with training data representing them). In some embodiments, those sequences that share more than a certain percentage with the target binding partner (e.g., 40% sequence homology) may also be removed from the training data. In some embodiments, antigen structures of the same class, architecture, topology, or homologous superfamily with the target binding partner may be removed from the training data. Block 310c may include an antibody scaffold. Given the input antigen and antibody scaffold sequences, the ML model may output a desired sequence (e.g., the sequence of the CDR3 of a VHH scaffold against the target EGFR).

[0090] FIG. 3D illustrates an exemplary block diagram 312 of machine learning based drug candidate design, according to an embodiment. FIG. 312 includes block 314a illustrating an equivalent representation of a biomolecular structure. In particular, block 314a illustrates structure prediction, rotational equivalence, and local coordinate frame of biomolecular structural information. Furthermore, block 314b illustrates that an attention layer may enable non-linear modeling of potential epitope-paratope interactions, and various modeling architectures that may be used. Still further, block 34c illustrates iterative design of drug candidates.

[0091] FIG. 3E illustrates an exemplary block flow diagram 316 of antigen encoding, according to some embodiments. Block diagram 316 includes an ML model that is trained by receiving one or more training inputs, which may be 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 (EGFR antigen, to continue the above example). Block diagram 316 may inject the antigen onto the graph (block 318b). The ML model may be trained by passing the training inputs around the graph (e.g., by ML training module 164 of FIG. 1). Thus, every "node" or residue of the graph in block 318b may be summarized, for example, by its neighbors. The ML model may retain this information, and over the course of training, it will learn how to extract meaningful information, such as point cloud relationships (block 318d).

[0092] FIG. 3F illustrates an exemplary block flow diagram 320 of antibody generation training according to some embodiments. In FIG. 320, the ML model starts with an initial guess of the antibody structure based on the input sequence (e.g., SAbDab sequence) (block 322a). The ML model may predict the target biomolecule structural information (e.g., the next sequence of the antibody structure) based on the antigen code information learned in block flow diagram 316 (block 322b). Depending on whether the prediction is correct or incorrect, the model may output a prediction that may be compared to ground truth data to adjust the weights of the model (block 322c). For example, the ML model may evaluate a loss function in block 322c and modify one or more parameters of the ML model based on the loss function. As discussed herein, the antibody training in block flow diagram 320 may continue until the structural information of the target biomolecule is completed.

[0093] FIG. 3G illustrates an exemplary block flow diagram 324 of antibody generation training / computation according to some embodiments. FIG. 324 may include analyzing an input antibody scaffold that includes an unknown region (e.g., CDR3 region), illustrated by a string that includes the letter "X". Block diagram 324 may include making a first prediction using the trained ML model, where the output is a probability distribution for the sequence that comes next to the unknown region. In some embodiments, an attention function may be used with a softmax activation function. The first prediction may be the next sequence of CDR3. FIG. 3H illustrates an exemplary block flow diagram of antigen generation training / computation according to some embodiments. FIG. 3H illustrates a second prediction (i.e., prediction t+1), where the second sequence of CDR3 is filled in by the model. In some embodiments, the ML model may make all predictions in a single step rather than iteratively. For example, in some embodiments, the predicted structure is a random structure. In other embodiments, the same starting structure is used every time the model runs. In some aspects, other machine learning architectures may work, including those used in other disciplines / technologies (e.g., visual artificial intelligence / machine learning fields). In particular, those skilled in the art will understand that diffusion models and other generative modeling techniques that work well for continuous systems (e.g., image data) may be adapted for use with the present technology.

[0094] FIG. 3I illustrates an exemplary block flow diagram of antibody generation training according to some embodiments. In particular, FIG. 3I illustrates an example of a prediction that matches a known ground truth. Thus, FIG. 3I is an example of negative reinforcement. FIG. 3J illustrates an exemplary block flow diagram of antibody generation training according to some embodiments. Specifically, FIG. 3J illustrates an example of a generated antibody that is a three-dimensional structure / framework. For example, the illustrated three-dimensional structure can be depicted using three-dimensional coordinates corresponding to the primary sequence of the target biomolecule. FIG. 3K illustrates an exemplary block flow diagram of antibody generation training according to some embodiments. In particular, FIG. 3K illustrates a third correct answer, for example, against the ground truth values ​​of FIG. 3F and FIG. 3I. FIG. 3L illustrates an exemplary block diagram illustrating a supervised machine learning model output according to some embodiments. Block diagram 334 includes both a prediction of structural information of the target biomolecule and one or more sequences thereof. Block diagram 334 also illustrates that additional loss functions may be applied to the physical properties, naturalness, confidence, docking, affinity, etc. of the structural information of the target biomolecule.

[0095] Illustrative Results FIG. 4 illustrates an exemplary graphical user interface 400 including visualized predicted structure information 402a and visualized ground truth structure information 402b. The visualized predicted structure information 402a may correspond to the output of the affinity prediction block 222a. The visualized ground truth structure information 402b may correspond to data received from a database or other source (e.g., protein data bank, wet lab crystallography, etc.). The visualized predicted structure information 402a may be generated by inputting a target input including 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 one or more of the fragments or portions of any of the foregoing described herein. For example, in some examples, the input to the ML model may be one or more protein sequences. The visualized predicted structure information 402a and the visualized ground truth structure information 402b may be generated by computing a pairwise distance algorithm between individual points of the visualized predicted structure information 402a and the visualized ground truth structure information 402b. Lighter values ​​may indicate values ​​that are further apart, while darker values ​​may indicate values ​​that are closer to each other. In particular, the diagonal lines (coordinates i, i) represent the same residue.

[0096] Regarding the results of the present technique, the fact that the visualized predicted structure information 402a and the visualized ground truth structure information 402b appear visually similar indicates that the trained ML model / ANN is functioning as expected, in particular the protein folding component of the network is functioning and designing the output along the way.

[0097] Exemplary in silico biomolecular designs 5A illustrates an exemplary block flow diagram 500 of a complete in silico biomolecular design according to some embodiments. Specifically, a group of patient samples may be collected, each of the patients having one or more diagnosed diseases (block 502). One or more AI models may be applied to the patient samples to discover novel biomolecular targets (block 504, block 506). For example, the AI ​​models may correspond to those trained by the ML training module 164 of FIG. 1.

[0098] Figure 5B illustrates an exemplary block diagram 500 of a fully in silico de novo biomolecular design artificial intelligence model, according to some embodiments. As discussed herein with respect to Figure 3A, an antibody scaffold can be processed by the ML model to generate one or more designed CDRs 516 (e.g., corresponding to one or more respective drug candidates) (block 512). Figure 5B further illustrates a lead optimization model including in silico selection of amino acid identities within the Trastuzumab CDR3 to increase antibody affinity for the antigen (Her2) (block 514).

[0099] FIG. 5C illustrates an example block flow diagram 520 of an in silico biomolecule design pipeline, according to some embodiments. A group of patient samples may be collected, with the patients having one or more diagnosed diseases (block 522). De novo discovery may be performed in silico using one or more trained ML models discussed herein that are provided with target inputs to predict structural information of the target biomolecule (block 526-A). In some embodiments, the ML model may receive additional information (e.g., information from an ACE assay) (block 524). In some embodiments, lead optimization may be performed using a lead optimization machine learning model discussed herein (block 526-B). The ML model may output novel biomolecules targeted to specific epitopes with desired scaffolds and / or desired attributes (block 528). These outputs may be used to generate advanced producing cell lines and create drug candidates (block 530). Those skilled in the art will appreciate that the in silico design techniques of the present invention can benefit from the use of supercomputing resources (e.g., accelerated neural network training environments such as the GNN SE(3) Transformer, hosted solutions for AI development (e.g., DGX Foundry), and graphic processing unit (GPU) supercomputers (e.g., AX 100).

[0100] Example machine learning modeling results As discussed herein, the technology includes ML modeling to automatically design or improve biomolecules against a particular target of interest. For example, in some cases, models can be pre-trained to predict the structure, sequence, and / or affinity of an antibody given an antigen structure (e.g., using biomolecular complexes mined from protein data banks). Recent studies have quantified that the technology is highly effective even without the added benefit of pre-training ML models to predict the binding affinity of target binding partners and predicted target biomolecular structural information.

[0101] Figure 6A illustrates an example of in silico biomolecular engineering including binder prediction based on optimization of CDRH3 while keeping the remaining CDRs unchanged, according to some embodiments. Specifically, Figure 6A illustrates wild-type heavy chain biomolecules, namely, trastuzumab wild-type light chain HER2, bH1 light chain HER2, and bH1 light chain VEGF-A. Each of these biomolecules has distinct K of 2 nM, 26 nM, and 300 nM, respectively. d Figure 6B illustrates an exemplary visualization of ranibizumab and trastuzumab bH1, according to some embodiments.

[0102] FIG. 6C illustrates an exemplary sensorgram corresponding to SPR results generated for predicted in silico binding of Trastuzumab-bH1 CDR3 variants to VEGF, according to some embodiments. FIG. 6D illustrates exemplary tabular data corresponding to SPR results generated for predicted in silico binding of Trastuzumab-bH1 CDR3 variants to VEGF in FIG. 6C, according to some embodiments. Specifically, FIG. 6C and FIG. 6D illustrate the output corresponding to inputting the CDRH1 and CDRH2 of Trastuzumab-Bh1 and the structure of VEGF into a trained ML model to design the CDR3 of Trastuzumab-Bh1. The heavy chain CDRs can be the same for Trastuzumab and Trastuzumab-Bh1, but the light chain CDRs can be different. The results demonstrated that the predicted variants specifically bind to VEGF, as shown in Figure 6D, where the name column indicates the control biomolecule (e.g., ABS26144=trastuzumab and ABS26148=trast-bH1). Trastuzumab bH1 is a known variant of trastuzumab designed to bind to 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 inSpite of Structural Plasticity. PLoS ONE 6(4):e17887.doi:10.1371 / journal.pone.0017887). The structure of bH1 bound to VEGF was used to predict the CDR3 variants that bind to VEGF. The Name column also illustrates predicted biomolecules that begin with the prefix "Trast_".

[0103] FIG. 6E illustrates an exemplary sensorgram corresponding to SPR results generated for predicted in silico binding of Trastuzumab CDR3 variants binding to Her2 according to some embodiments. FIG. 6F illustrates exemplary tabular data corresponding to SPR results generated for predicted in silico binding of Trastuzumab CDR3 variants binding to Her2 in FIG. 6E according to some embodiments. Specifically, FIG. 6E and FIG. 6F illustrate the output corresponding to inputting the CDRH1 and CDRH2 of Trastuzumab and the structure of Her2 into a trained ML model to design the CDR3 of Trastuzumab. In particular, the CDR3 of the heavy chain (the other five CDRs are from Trastuzumab) can be predicted using the structure of Trastuzumab bound to HER2 to predict the CDR3 variants that will bind to HER2. Antibodies can be produced and affinities can be measured using SPR to verify the predicted binding. The results demonstrate that the predicted variants specifically bind to Her2.

[0104] In general, the data presented in Figures 6A-6D demonstrate that the present technique for in silico design of target biomolecules can predict 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 the time-consuming and resource-intensive traditional wet-lab approach. Furthermore, the results identified thus far do not necessarily require the docking complex and affinity prediction techniques described herein for fine-tuning. That is, the results in Figures 6A-6D are achievable through one or more biomolecular prediction layers trained to predict biomolecular structural information from target input. Adding additional 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 affinities from the input docking complexes) is expected to produce even higher quality results.

[0105] Screening hundreds of thousands of model-generated sequences for binding In some embodiments, the technology leverages the ACE assay (e.g., Bachas S et al.; Liu J. Activity-specific cell enrichment; Patent Publication No. WO2021 / 146626, 22.07.2021, etc.) to screen large antibody variant libraries containing hundreds of thousands of members (e.g., represented in Fab64 format). After screening with ACE, the technology may validate the assay for the de novo discovery technology described herein by sampling sequences for follow-up analysis by SPR, the gold standard in binding affinity measurement and detection. Experimental results have demonstrated that the ACE assay has high recall, correctly classifying >95% of SPR binders. Furthermore, the assay has high accuracy, with 60% of ACE binders being SPR validated. This may enable a powerful workflow in which a large predicted population can be first screened by the ACE assay, as discussed herein, and then the expected binding population can be screened via SPR to remove false positives and collect high-quality binding affinity measurements.

[0106] The present technology demonstrates the ability of generative AI to design de novo antibodies and antigens of interest and generate new HCDR3 sequences zero shot for known antibodies. The present technology focuses on the design of the HCDR3 region, which is a major determinant of antibody function, due to its high sequence diversity and high density of paratope residues in the immune repertoire. For example, the present technology may select trastuzumab, which binds to HER2, as a scaffold antibody for testing HCDR3 designed sequences. HCDR3 designs are generated by a model tuned to the antigen-only modified HER2 3-D structure (PDB: 1N8Z (chain C)) and the sequence of the trastuzumab scaffold, except for HCDR3. The present technology may remove any antibodies known to bind the target or any homologs (>40% sequence identity or part of the same homologous superfamily) for the models used herein. In some settings, we instead remove all antibodies from the training set that have >40% sequence identity to the wild-type antibody. In both cases, we observed binders. A total of 440,354 antibody variants were generated and screened in the ACE assay to identify binding variants. The technology found a total of about 4000 predicted binders based on the expected ACE assay binding rate, and is proceeding with a subset for further characterization, confirming HER2 binding for a total of 421 AI designs via SPR. [Table 3]

[0107] The confirmed binders exhibited a range of affinities for HER2, with 71 designs exhibiting affinities <10 nM (Figure 8A). Strikingly, three zero-shot designs exhibited tighter binding than trastuzumab, with one binding in the sub-nanomolar affinity range. These high affinity design examples were zero-shot generated from the model without additional affinity maturation, thus skipping a typically critical step in the therapeutic antibody development process. The ability to generate desirable antibodies without the need for additional optimization could significantly shorten development timelines.

[0108] Figure 8A illustrates the binding affinity of AI-generated zero-shot binders, finding 71 designs with comparable affinity to trastuzumab (<10 nM) and 3 designs with tighter binding. Figure 8B illustrates the designed variant binding affinity versus edit distance to trastuzumab. Edit distances range from 2 mutations (84.6% sequence identity) to 12 mutations (7.7% sequence identity), illustrating the novelty of the designs. Figure 8C illustrates the logo plot of the HCDR3 of 421 binding affinity trastuzumab variants. Greater diversity is observed in the center of the designed CDRs. The sequence logo below is the trastuzumab HCDR3 sequence. Figure 8D illustrates the pairwise edit distance between binders (min 1, max 15, median 8, mean 7.7, ±2.1 SD).

[0109] In addition to favorable affinity, the AI ​​model designs have high sequence diversity, both in terms of amino acid length and identity. The validated binders have HCDR3s ranging from 11 to 15 amino acids, compared to the Trastuzumab length of 13 (Figure 8F). The designed sequences are also distinct from the Trastuzumab antibody, with edit distances of 2 to 12 from the Trastuzumab sequence (Figure 8B). The average affinity decreases as the edit distance increases from the Trastuzumab sequence, but interestingly, the technique finds designs that still exhibit affinity below 10 nM across all edit distances. The technique observed one design with an edit distance of 9 that exhibited higher affinity than the Trastuzumab antibody. In addition, the technique found higher diversity in the center of the HCDR, which corresponds to the more diverse D germline gene, compared to the less diverse adjacent J and V germline genes (Figure 8A). The designs are also sequence diverse from each other, with an average edit distance of 7.7 ± 2.1 SD (Figure 8D, Figure 8E). The inter-design diversity is noteworthy because it indicates that the model-generated binders have not converged on a shared sequence motif, which is often seen with traditional antibody screening methods such as phage display.

[0110] The designed binders display sequence novelty Despite the high sequence diversity of the 421 designed binders, the technique compares the model output to the training set to ensure that the model generation is a novel sequence and not simply a recapitulated training example. This phenomenon has been observed in machine learning models, and past methods have been criticized for generating molecules similar to previously known molecules. The technique calculated the minimum distance between the designed binder and all HCDR3s in the model training and validation sets, and found that the designed binder was different from those observed during training (Figure 9A). The technique then calculated the distance to all HCDR3s in the Structural Antibody Database (SAbDab), a database of antibody-antigen complexes, and found that the binder was distant in sequence space from all antibodies in the database (Figure 9B).

[0111] FIG. 9A illustrates the minimum edit distance of binders to training data HCDR3s (min 2, max 8, median 5, mean 4.68±1.34SD). FIG. 9C illustrates the minimum edit distance of binders to OAS HCDR3s. 9.3% (38 out of 421) of HCDR3 designs are included in OAS (min 0, max 5, median 2, mean 1.91±1.08SD). FIG. 9D illustrates the naturalness scores of designed binders versus controls. De novo binders are those identified in our study. OAS refers to HCDR3s randomly selected from OAS. Frequency baseline samples amino acids at each position based on position frequency observed in OAS for HCDR3. Phage display baseline is a set of HCDR3s sampled from binding and non-binding antibodies of Liu et al. Scrambled OAS is a randomly permuted version of the OAS sequence set. The AI ​​designs have significantly higher naturalness scores on average than the control population (p<10e-50), but lower naturalness scores on average than trastuzumab and sequences randomly sampled from OAS (p<10e-15). The red dashed line is the naturalness score of trastuzumab. Figure 9E illustrates the naturalness scores of the designed binders versus the edit distance to trastuzumab. The red dashed line is the naturalness score of trastuzumab.

[0112] The technique examined sequence similarity of the model's output to sequences in the observed antibody space (OAS), a database of immune repertoire sequencing studies. The technique found that some of the model's designs already existed in the OAS, while other designs were unique with a minimum edit distance of 1-5 (Figure 9B). This indicates that the model is capable of generating sequences that achieve high naturalness scores and that are distinct from antibodies in the observed immune repertoire.

[0113] Zero-shot designs have high naturalness scores Successful therapeutic lead antibody candidates in the drug development process are typically developable with high affinity and low immunogenicity. In previous work, we described a language model that can assign a score to antibody sequences that indicates the likelihood of finding the sequence in a typical immune repertoire. This metric is called naturalness. High naturalness scores are associated with favorable developability and immunogenicity development outcomes. Using our design naturalness scoring model, we find that we can generate sequences with high affinity and high naturalness scores in a zero-shot fashion, even though they have not been trained or sampled based on these qualities (Figure 9E). Many designs exhibit naturalness scores.

[0114] FIG. 9D illustrates the naturalness measures of the designed binders of the present invention versus their diversity relative to trastuzumab. Importantly, many of the designs illustrated in FIG. 9D are more natural than the wild type, despite the fact that the wild type (illustrated by a horizontal line) corresponds to an FDA-approved molecule. Similar to binding affinity, the naturalness of the binders decreases as the edit distance to trastuzumab increases. However, empirical testing using the present technology still identifies binders with high edit distances (5-9 mutations) that have higher naturalness than trastuzumab. Table S5 shows the average naturalness scores across the different groups, and p-values ​​for the associated statistical comparisons to de novo binders. These results highlight the potential of zero-shot designs to bypass parts of the traditional lead optimization process, potentially saving time and resources in drug development.

[0115] 3D structural representation of de novo designed HCDR3 and their comparison with the trastuzumab-HER2 complex Table 1. Characteristics of various HCDR3 candidates selected for 3D structural modeling. HCDR3 candidates were selected based on affinity to trastuzumab, length, and edit distance. RMSD values ​​were calculated across all main-chain and side-chain atoms from the alignment of HCDR3 residues. All other atoms were excluded from the calculation. The range average of hydropathy values ​​of HCDR3 residues was calculated by averaging the hydropathy value of each residue and dividing by the sequence length. 58 [Table 4]

[0116] The technology then predicts the binding mechanism of the de novo designed HCDR3 variants of the present invention to better understand the structural basis of the highly sequence diverse variants. To this end, the technology constructed structural models of eight diverse HCDR3 candidates bound to HER2 in Fab format. These eight variants are selected based on their edit distance to the trastuzumab HCDR3, their length diversity (ranging from 12 to 15 amino acids), and their affinity range spanning three orders of magnitude (Table 1). The technology uses the trastuzumab Fab-HER2 complex (PDB:1N8Z) as a starting template for structural modeling. The technology performs locally constrained backbone geometry and side chain rotamer optimization, followed by relaxing the Fab-HER2 complex to correct for global conformational ambiguities, steric clashes, and suboptimal loop geometry. As a control, an experimental trastuzumab-Fab complex is optimized using the same protocol and used as a reference to compare the final HCDR3 structural model. The technique uses the lowest free energy pose of a de novo HCDR3 model for structural analysis and comparison.

[0117] The eight de novo Fab-HER2 structural models are overall similar to trastuzumab with all-atom HCDR3 RMSDs ranging from 1.9 to 2.4 Å, despite sequence differences. Minimal structural rearrangements are observed in the unmodified regions of the antigen heavy, light, and epitope residues (Figure 10A). In selected cases, side chains forming contacts with HCDR3 show slight rotamer differences to account for the presence of longer loops or steric clashes from residues with larger side chains (Figure 10B). With respect to the designed HCDR3 regions, alignment with the trastuzumab Fab-HER2 complex reveals a dynamic ensemble of conformations adopted by each HCDR3 (Figure 10C). The differences in HCDR3 loop structures are wide, with RMSDs ranging from 1.1 to 6.7 Å when aligned across all main chain and side chain atoms (Table 2). In FIG. 10C, despite sequence and length diversity, there are key residues that are conserved in space, corresponding to trastuzumab residues W107, G109, and Y113 (IMGT numbering scheme). Table 2. Multi-stage AI designed Trastuzumab variant binders against HER2 with all three HCDRs designed. ED indicates the edit distance from the HCDR to the Trastuzumab value. Note that the model occasionally reproduces the native Trastuzumab HCDR1 and HCDR2. We present nine variants here and have open-sourced the entire set of 23 designs in the accompanying sequence data. [Table 5]

[0118] Although the de novo HCDR3 adopts distinct conformations, there are important positional similarities between all structures. Closer analysis of the spatial orientation of the side chain conformations reveals conservation of identical side chains in five separate spatial locations, two of which correspond to IMGT residue positions R103 and Y117 in trastuzumab, which are highly conserved in most antibodies. However, there is physiochemical conservation in all structures corresponding to the spatial locations of IMGT residues 109, 113, and 117 in trastuzumab, which contribute to the paratope of the trastuzumab-Her2 complex. Although spatially conserved, these side chains may be required to orient key paratope side chains to make the same important protein-protein interactions with the epitope of HER2, stemming from the diversity of residue positions that highlights the observed conformational flexibility.

[0119] Although the overall binding region is identical, each designed HCDR3 exhibits a different binding mode between the HCDR3 and the epitope. In most cases, novel interactions that are not observed in the trastuzumab-HER2 complex are formed between the designed HCDR3 and domain IV HER2 (Figure 10B, Figure 10D). These interactions are diverse and consist of novel hydrogen bond interactions, non-polar interactions, aromatic interactions, and electrostatic interactions formed with each HCDR3 and the Her2 epitope. To further decipher the determinants of binding, the present technology calculated the surface area buried by each HCDR3 variant Fab when bound to HER2. The surface area buried when the antibody Fab binds to Her2 is defined as the binding surface area between the paratope and the epitope (Table 2).

[0120] In some cases, the de novo HCDR3 variants exhibited a larger binding surface area than trastuzumab, which may suggest a novel interaction with the HER2 epitope. Interestingly, no correlation was observed between the binding surface area and the binding affinity. This finding would suggest that hydrophobic contribution and surface area burial are not the major determinants of binding. Furthermore, the specific contacts formed between each designed HCDR3 and the epitope are important for the binding stability of the Fab-Her2 complex. Furthermore, the present technique calculated the grand average of the hydropathy values ​​(GRAVY) of each HCDR3 variant, which defines the total hydrophobic character summed over each residue, and compared it with the binding affinity. The present technique did not observe a correlation between affinity and hydrophobicity, which further confirms that the hydrophobic effect is not the major determinant of binding to the de novo designed HCDR3. (Table 2). Taken together, these results suggest that the binding affinity of the designed HCDR3 is intrinsic to the sequence design and is not driven by a common binding mechanism. The high dependence of binding on sequence attributes is consistent with the low probability of designing binders by chance.

[0121] Validation against additional targets The technology will next conduct pilot studies to demonstrate the applicability of our approach to a broader set of antigens. For these additional targets, the technology will not pre-screen by ACE assay. Rather, the technology will sample a small number of sequences and verify binding by SPR. The technology has successfully designed for the first time a therapeutic HCDR3 variant of ranibizumab that binds to human vascular endothelial growth factor A (VEGF-A), as shown in Figure 11A. The binder has an affinity of 48.2 nM as measured by SPR, compared to the sub-nanomolar binding of ranibizumab (0.37 nM). In addition, the designed HCDR3 is significantly different from ranibizumab, with an edit distance of 13, is novel, and is separated from any HCDR3 in OAS by a minimum of 4 mutations.

[0122] The technology then designs a set of HCDR3 variants of casirivimab tuned for Omicron spike protein binding. Casirivimab binds to multiple COVID spike protein variants, and in particular, binds weakly to Omicron. The technology measures casirivimab affinity for Omicron via SPR at 240.0 nM (Table S1). The technology identifies one AI-designed variant that binds with similar affinity to Omicron at 179.7 nM (Figure 11B, Table S1). Interestingly, the technology does not observe binding to other spike protein variants due to the AI ​​design, suggesting potential controllability of target specificity between homologous antigens. The designed variant has a distinct HCDR3 sequence compared to casirivimab, with an edit distance of 6 from any HCDR3 in OAS and a minimum edit distance of 2. In addition, the HCDR3 has an edit distance of at least 4 from any HCDR3 in CoV-AbDab. These binding designs for two additional antigens highlight the scalability of our zero-shot design approach and demonstrate the potential for selective antigen controllability through generative AI. Table S1. Measured binding affinities of casirivimab and AI-designed variants to COVID antigens. N / A indicates lack of binding. In this setting, the model was instructed to generate HCDR3 variants targeting Omicron that fit into the casirivimab framework. These results show that the model successfully designed HCDR3 variants that lost activity against COVID Spike Wildtype / Beta / Delta, yet maintained activity against Omicron. These results are a first step toward the controllability of antibody design against specific protein variants. [Table 6]

[0123] Expansion to multiple CDRs Expanding antibody design beyond HCDR3 allows for increased sequence diversity and controllability and represents the next step towards full de novo design. To this end, we applied an alternative multi-step AI design method that differs from our described zero-shot approach to generate variants of all heavy chain variable regions (HCDR1, HCDR2, HCDR3) simultaneously. The technology reports multiple binding designs against HER2 identified within a library of less than 500 multi-step design variants (Table 2). The technology again found that these binders were distinct from the model training data and antibody examples in the SAbDab and OAS databases (Figure 9C, Figure 9I).

[0124] Exemplary Materials and Methods Library Design For de novo design of HCDR3 trastuzumab variants, the technology may begin by designing, for example, two high diversity libraries (HDLs), denoted HDL1, consisting of 223,046 and 217308 designs, respectively. HDL1 may be screened using the ACE and dnACE assays described herein, and the results are used to design low diversity libraries (LDLs), consisting of 1,000 or less sequences, each of which is screened using SPR. HDL2 is screened using ACE, and the results are used again to design LDLs for SPR screening. In total, 199 binders are confirmed from HDL1 based on ACE scores, revealing a putative additional 3765 binders in HDL2 (Table S5). The ACE assay may not be used to screen variants of HCDR3 ranibizumab, HCDR3 casirivimab, or HCDR123 trastuzumab. Instead, small LDLs may be screened directly with SPR. Table S4. Estimated number of binders by ACE "bin" in HDL2. In addition to the 224 confirmed binders, an additional 3,780 designs are estimated to bind from HDL2 based on the distribution of sequences within the ACE bins. Note that the total number of identified designs (41,942) is less than the total number of designs (217,308). The difference corresponds to designs that were not sorted by ACE, indicating that they are less likely to bind. [Table 7]

[0125] Naturalness Score The naturalness score used in this study can be calculated using the pre-trained antibody language model discussed herein and introduced previously. This model is based on the pseudo-perplexity (e.g., defined by the integration of IMGT and Martin definition) over the extended CDRs of the antibody heavy chain under the language model. This index can be predictive of desirable antibody therapeutic properties, such as developability or lack of immunogenicity. The technology can place a variant HCDR3 on a wild-type scaffold by replacing HCDR3. In addition to calculating the naturalness of the de novo binder of the present invention, the technology can further include some controls.

[0126] 1. OAS: It consists of 1,000 HCDR3s randomly sampled from OAS. Naturality score is calculated over the grafting of HCDR3 to trastuzumab scaffold. Considering that the naturality model is pre-trained on OAS, we expect the sequence from OAS to have high naturality score, therefore we treat it as positive control. 2. Frequency baseline: These are 1000 sequences generated by randomly sampling from the length-adjusted frequency distribution of the OAS. The technology can calculate PL(l), the probability that the HCDR3 in the OAS has length l, and then the probability of sampling a particular sequence with length l using an independent factorization based on amino acid frequency defined as follows:

number

number

[0127] The naturalness score of the de novo design of the present invention is calculated using a two-sample t-test (H0: μ1 = μ2, H a :μ1!=μ2) and compared to these controls using a one-sample t-test with wild-type naturalness as the population mean (H0:μ1=μ, H a : μ1 != μ) A table of the mean naturalness scores is shown below along with p-values ​​for relevant statistical comparisons across the different groups and against de novo binders. [Table 8]

[0128] Table 2 illustrates the mean naturalness scores across the different groups and p-values ​​using the trastuzumab scaffold.

[0129] Furthermore, the present technology has already pointed out, in some embodiments, the diversity of designs for OAS, the pre-training data used for the model used to calculate the naturalness herein. Adopting samples far from the training set may lead to a decrease in naturalness. Indeed, as shown in Figure 9B, empirical evidence shows that as the distance from the OAS increases, the naturalness of the design tends to decrease. Of particular interest is the fact that the average naturalness of the designs of the present invention contained in the 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). Indeed, the designs of the present invention that are only one mutation away from any HCDR3 in the OAS have significantly higher naturalness than the OAS control sequence (p=9.83e-5).

[0130] Exemplary Binder Data Consensus As discussed, the technology includes a training model that uses training data. For example, the technology may use the following binder data for training in some embodiments. The following Tables E1-E4 include de novo trastuzumab HCDR3 variant HER2 binder, de novo nibizumab HCDR3 variant VEGF-A binder, de novo casirivimab HCDR3 variant COVID-Omicron binder, and de novo trastuzumab HCDR123 variant HER2 binder, respectively. The following consensus sequences are applied to the binder data:

[0131] HER2 binders For HER2 binders, the following information is relevant but does not vary for each binder.

[0132] Parent antibody: trastuzumab Antigen: HER2 Parent PDB:1N8Z Parent heavy chain sequence (SEQ ID NO:76):EVQLVESGGGLVQPGGSLRLSCAASGFNIKDTYIHWVRQAPGKGLEWVARIYPTNGYTRYADSVKGRFTISADTSKNTAYLQMNSLRAEDTAVYYCSRWGGDGFYAMDYWGQGTLVTVSSASTKGPSVFPLAPSSKSTSGGTAALGCLVKDYFPEPVTVSWNSGALTSGVHTFPAVLQSSGLYSLSSVVTVPSSSLGTQTYICNVNHKPSNTKVDKKVEP Parent HCDR1 (SEQ ID NO:54): GFNIKDTY Parent HCDR2 (SEQ ID NO:55): IYPTNGYT Parent HCDR3 (SEQ ID NO: 20): SRWGGDGFYAMDY Parent light chain sequence (SEQ ID NO:77):DIQMTQSPSSLSASVGDRVTITCRASQDVNTAVAWYQQKPGKAPKLLIYSASFLYSGVPSRFSGSRSGTDFTLTISSLQPEDFATYYCQQHYTTPPTFGQGTKVEIKRTVAAPSVFIFPPSDEQLKSGTASVVCLLNNFYPREAKVQWKVDNALQSGNSQESVTEQDSKDSTYSLSSTLTLSKADYEKHKVYACEVTHQGLSSPVTKSFNRGEC Antigen chain sequence (SEQ ID NO: 78): TQVCTGTDMKLRLPASPETHLDMLRHLYQGCQVVQGNLELTYLPTNASLSFLQDIQEVQGYVLIAHNQVRQVPLQRLRIVRGTQLFEDNYALAVLDNGDPLNNTTPVTGASPGGLRELQLRSLTEILKGGVLIQRNPQLCYQDTILWKDIFHKNNQLALTLIDTNRSRACHPCSPMCKGSRCWGESSEDCQSLTRTVCAGGCARCKGPLPTDCCHEQCAAGCTGPKHSDCLACLHFNHSGICELHCPALVTYNTDTFESMPNPEGRYTFGASCVTACPYNYLSTDVGSCTLVCPLH NQEVTAEDGTQRCEKCSKPCARVCYGLGMEHLREVRAVTSANIQEFAGCKKIFGSLAFLPESFDGDPASNTAPLQPEQLQVFETLEEITGYLYISAWPDSLPDLSVFQNLQVIRGRILHNGAYSLTLQGLGISWLGLRSLRELGSGLALIHHNTH LCFVHTVPWDQLFRNPHQALLHTANRPEDECVGEGLACHQLCARGHCWGPGPTQCVNCSQFLRGQECVEECRVLQGLPREYVNARHCLPCHPECQPQNGSVTCFGPEADQCVACAHYKDPPFCVARCPSGVKPDLSYMPIWKFPDEEGACQPCPIN

[0133] COVID, VEGF, and HER2 Common / invariant information for the COVID, VEGF, and HER2 affinity tables is as follows:

[0134] COVID: Parent antibody: casirivimab Parent antibody PDB:6XDG Antigens for which binding affinity data was collected via SPR: SARS-CoV2 spike RBD (wild type, beta, delta, omicron) Parent heavy chain sequence (SEQ ID NO: 79): QVQLVESGGGLVKPGGSLRLSCAASGFTFSDYYMSWIRQAPGKGLEWVSYITYSGSTIYYADSVKGRFTISRDNAKSSLYLQMNSLRAEDTAVYYCARDRGTTMVPFDYWGQGTLVTVSSASTKGPSVFPLAPSSKSTSGGTAALGCLVKDYFPEPVTVSWNSGALTSGVHTFPAVLQSSGLYSLSSVVTVPSSSLGTQTYICNVNHKPSNTKVDKKVEPKSCDKT Parent HCDR3 (SEQ ID NO:51):ARDRGTTMVPFDY Parent light chain sequence (SEQ ID NO: 80):DIQMTQSPSSLSASVGDRVTITCQASQDITNYLNWYQQKPGKAPKLLIYAASNLETGVPSRFSGSGSGTDFTFTISGLQPEDIATYYCQQYDNLPLTFGGGTKVEIKRTVAAPSVFIFPPSDEQLKSGTASVVCLLNNFYPREAKVQWKVDNALQSGNSQESVTEQDSKDSTYSLSSTLTLSKADYEKHKVYACEVTHQGLSSPVTKSFNRGEC Model-tuned antigen: SARS-CoV-2 omicron spike RBD PDB:7QO9

[0135] VEGF: Parent antibody: Ranibizumab · Patent antibody PDB:1CZ8 Antigen: VEGF-A Parent heavy chain sequence (SEQ ID NO: 82): EVQLVESGGGLVQPGGSLRLSCAASGYDFTHYGMNWVRQAPGKGLEWVGWINTYTGEPTYAADFKRRFTFSLDTSKSTAYLQMNSLRAEDTAVYYCAKYPYYYGTSHWYFDVWGQGTLVTVSSASTKGPSVFPLAPSGTAALGCLVKDYFPEPVTVSWNSGALTSGVHTFPAVLQSSGLYSLSSVVTVPSSSLGTQTYICNVNHKPSNTKVDKKVEPK Parent HCDR3 (SEQ ID NO: 49): AKYPYYYGTSHWYFDV Parent light chain sequence (SEQ ID NO: 83):DIQLTQSPSSLSASVGDRVTITCSASQDISNYLNWYQQKPGKAPKVLIYFTSSLHSGVPSRFSGSGSGTDFTLTISSLQPEDFATYYCQQYSTVPWTFGQGTKVEIKRTVAAPSVFIFPPSDEQLKSGTASVVCLLNNFYPREAKVQWKVDNALQSGNSQESVTEQDSKDSTYSLSSTLTLSKADYEKHKVYACEVTHQGLSSPVTKSFNRGE Antigen chain sequence (SEQ ID NO: 84): VVKFMDVYQRSYCHPIETLVDIFQEYPDEIEYIFKPSCVPLMRCGGCCNDEGLECVPTEESNITMQIMRIKPHQGQHIGEMSFLQHNKCECRPK

[0136] 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)). [Table 9-1] [Table 9-2] [Table 9-3] [Table 9-4] [Table 9-5] [Table 9-6] [Table 9-7] [Table 9-8] [Table 9-9] [Table 9-10] [Table 10] [Table 11] [Table 12-1] [Table 12-2] [Table 12-3]

[0137] Cloning In some embodiments, antibody variants may be cloned and expressed in Fab format. For example, to generate ACE assay and SPR data sets, the technology may include synthesizing DNA variants spanning HCDR3, HCDR1 to HCDR3 across different libraries in oligonucleotide format using ssDNA oligo pools (Twist Bioscience) and a double-stranded DNA format using eBlocks (IDT). Codons may be randomly selected from the two most common in E. coli B strains for each variant. Amplification of the ssDNA oligo pools may be performed by PCR following TwistBioscience's recommendations, except that Q5 high fidelity DNA polymerase (NEB) may be used instead of KAPA polymerase. Briefly, a 25 μL reaction may consist of 1× Q5 Mastermix, 0.3 M each of the forward and reverse primers, and 10 ng of the oligo pool. Reactions may be initially denatured at 95° C. for 3 min, followed by 13 cycles of 95° C. for 20 s, 66° C. for 20 s, 72° C. for 15 s, and a final extension at 72° C. for 1 min. DNA amplification may be confirmed by agarose gel electrophoresis, and the amplified DNA was subsequently purified (DNA Clean and Concentrate Kit, Zymo Research).

[0138] To construct libraries for SPR validation of the model designs in independent experiments, oligonucleotides spanning the appropriate CDRs and the immediately upstream / downstream adjacent nucleotides were synthesized by Integrated DNA Technologies (IDT).

[0139] To generate a linearized vector, the technique may involve performing a two-step PCR to split the proprietary plasmid vector carrying Fab format trastuzumab into two fragments in a manner that provides approximately 25 nucleotides (nt) on the 5' and 3' ends of the amplified ssDNA oligo pool library or 40 nt cloning overlap on the 5' and 3' ends of the IDT eBlocks. The vector linearization reaction may be digested with DpnI (New England Bioloabs) and purified from a 0.8% agarose gel (Gel DNA Recovery Kit, Zymo Research) to eliminate carryover of the parent vector.

[0140] Cloning reactions may consist of 50 fmol of each purified vector fragment, 100 fmol of either purified library (Twist Bioscience) or 10 pmol of gBlock insert (IDT), and 1× final concentration of NEBuilder HiFi DNA Assembly (New England Biolabs). Reactions may be incubated at 50° C. for 2 hours (Twist Bioscience libraries) or 25 minutes (IDT libraries) and then purified (DNA Clean and Concentrate Kit, Zymo Research).

[0141] For HDL, Transformax EPI300 (Lucigen) E. coli can be transformed by electroporation (BioRad MicroPulser) with the purified assembly reaction and grown overnight at 30°C in 20 mL Teknova LB Broth with 50 μg / mL kanamycin, shaking at 270 rpm for 18 hours at 30°C, 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 overnight at 30°C in 20 mL Teknova LB Broth with 50 μg / mL kanamycin, shaking at 270 rpm for 18 hours at 30°C, 80% humidity.

[0142] For LDL, Absci's SoluPro™ host strain can be transformed with the purified assembly reaction and grown overnight at 30° C. on agar plates containing 50 g / ml kanamycin and 1% glucose. Colonies can be selected for QC analysis before culturing for induction. The experimental parameters and process flow described above can be modified in some embodiments.

[0143] Exemplary QC Analysis In this technique, the quality of high diversity variant libraries can be assessed by deep sequencing. Briefly, library plasmid pools can be amplified by PCR across the HCDR3 region and sequenced, for example, with 2x150nt reads using an Illumina MiSeq platform with 20% PhiX. PCR reactions can be 10 nM primer concentration, Q5 2x master mix (NEB), and 1 ng of input DNA diluted in H2O. Reactions can be first denatured at 98°C for 3 minutes, followed by 30 cycles of 98°C for 10 seconds, 59°C for 30 seconds, and 72°C for 15 seconds, with a final extension at 72°C for 2 minutes. Sequencing results can be analyzed for mutation distribution, variant representation, library complexity, and expected sequence recovery. Indicators can include the coefficient of variation of sequence representation, the read share of the top 1% most common sequences, and the percentage of designed library sequences observed within the library. The quality of the low diversity variant library can be assessed by performing rolling circle amplification (Equiphi29, Thermo Fisher Scientific) on 24 colonies 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 matches to library sequences (Geneious Prime).

[0144] Exemplary antibody expression in SoluPro™ E. coli B strain After transformation and 8 hours of recovery, HDL can be grown in 50 mL of Teknova LB Broth with 50 μg / mL kanamycin at 30° C., 80% humidity, shaking at 270 rpm for 24 hours. At the end of the 24 hours, the preculture can be OD600 normalized to 0.1 in induction medium (IBM) (4.5 g / L potassium phosphate monobasic, 13.8 g / L ammonium sulfate, 20.5 g / L yeast extract, 20.5 g / L glycerol, 1.95 g / L citric acid) containing inducer and supplements (250 μM arabinose, 50 μg / mL kanamycin, 8 mM magnesium sulfate, 1 mM propionic acid, 1× Korz trace metals) and grown for 16 hours in a 500 mL baffled flask at 26° C., 80% humidity, shaking at 270 rpm. At the end of the 16 hours, a 500 mL aliquot can be adjusted to 20% v / v glycerol and stored at -80°C.

[0145] After transformation and QC of LDL, individual colonies may be picked into deep well plates containing 400 μL of Teknova LB Broth 50 μg / mL kanamycin and incubated at 30° C. and 80% humidity with shaking at 1000 rpm for 24 hours. At the end of 24 hours, a 150 μL sample may be centrifuged (3300 g, 7 min) and the supernatant decanted from the pre-culture plate for sequence analysis. 80 μL of the pre-culture was transferred to 400 μL of IBM containing inducers and supplements as above. Cultures may be grown for 16 hours at 26° C., 80% humidity with shaking at 270 rpm. At the end of 16 hours, a 150 μL sample may be taken and centrifuged (3300 g, 7 min) into a pellet from which the supernatant was decanted before being stored at −80° C.

[0146] Activity-specific cell enrichment (ACE) assay Cell preparation High-throughput quantitative selection of antigen-specific Fab-expressing cells can be adapted from conventional approaches in some aspects. For staining, thawed glycerol stocks with OD600=2 from induced cultures were transferred to 0.7 ml matrix tubes, centrifuged (4000 g, 5 min), and the resulting pelleted cells were washed three times with PBS (pH 7.4, 1 mM EDTA). The washed cells can be thoroughly resuspended in 250 μL of phosphate buffer (32 mM, pH 7.4) by pipetting before fixation by addition of 250 μL of 0.6% paraformaldehyde and 0.04% glutaraldehyde in phosphate buffer (32 mM, pH 7.4). After 40 min incubation on ice, samples may be centrifuged (4000 g, 5 min), pellets washed three times with PBS (pH 7.4, 1 mM EDTA), resuspended in permeabilization buffer (20 mM Tris, 50 mM glucose, 10 mM EDTA, 5 μg / mL rLysozyme) and incubated on ice for 8 min. Fixed and permeabilized cells may then be centrifuged (4000 g, 5 min) and washed several 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).

[0147] staining Prior to 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 of staining buffer containing 100 nM His / Avi tagged human HER2 (Acro Biosystems) conjugated to either 50 nM streptavidin-AF647 (Invitrogen) directly conjugated to AF647 via the free amine or untagged human Her2 (Acro Biosystems). The library can be incubated with the probe overnight (16 hours) at 4° C. with end-to-end rotation, centrifuged (4000 g, 5 min), and the pellet washed three times with PBS. The pellet can be resuspended in 500 uL staining buffer containing 26.5 nM anti-kappa light chain: BV421 (BioLegend) and incubated for 2 hours with end-to-end rotation at 4° C. before centrifugation (4000 g, 5 min), washing 3 times with PBS, and resuspended in 200 uL PBS for sorting.

[0148] Selection The library can be sorted by one of two methods based on binding in the quantitative ACE assay or the binary version of the ACE assay described herein. In either method, the library can be sorted on a FACSymphony S6 (BD Biosciences) instrument. Immediately before sorting, 50 μL of stained sample was transferred to a flow tube containing 1 mL PBS + 3 μL propidium iodide.

[0149] Aggregates, debris, and impermeable cells can be removed with singlet, size, and PI+ parent gating, respectively. Cells can then be gated to include only cells with kappa light chain expression (BV421). For the quantitative ACE assay, collection gates can be drawn on the sample across the log range of binding signal. The rightmost gate can be set to collect the brightest 0.1% of the library, and the leftmost gate can be set to collect at the low end of the positive binding signal based on the stained control line. Four additional gates of equal width can be distributed in between, each set to be approximately half the gMFI of the gate to the right.

[0150] For the binary version of the ACE assay, a total of three collection gates can be set to sample the upper end of the binding range (top 0.1%), the remaining positive binding signal events, and a negative gate containing events with no binding signal. Libraries can be sorted simultaneously on two instruments with the light multiplier adjusted to normalize fluorescence intensity, and collected events were processed independently as technical replicates.

[0151] Next-generation sequencing 2. Preparation of Selected Material Samples Sample preparation for sequencing may follow the same protocol for both the quantitative ACE assay and the binary version of the ACE assay described above. Specifically, cell material from the sorted gate may be collected in a diluted PBS mixture (VWR) in a 1.5 mL tube (Eppendorf). A sample of unsorted library material may be processed for QC and ACE index calculation. The post-sorted sample may be centrifuged (3,800 x g) and the tube volume may be normalized to 20 μl. Amplicons encompassing the HCDR3 region may be generated by PCR. The reaction may use 10 nM primer concentration, Q5 2x master mix (NEB), and 20 μl of sorted cell material input suspended in diluted PBS (VWR). The reaction may be first denatured at 98°C for 3 min, followed by 30 cycles of 98°C for 10 s, 59°C for 30 s, and 72°C for 15 s, with a final extension 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, normalized, and pooled by Qubitluorometer (Invitrogen) ready for sequencing using ThruPLEX DNA-Seq Kit (Takara Bio). Pool size can be verified via Tapestation 1000 HS and sequenced using 20% ​​PhiX on an Illumina NextSeq 1000 P2 (2 × 150 nt).

[0152] In silico structural modeling The FastRelax application of Rosetta is run with flexible backbone and side-chain degrees of freedom. Before the relaxation procedure, we first idealize all candidate structures using Rosetta's Idealize protocol to avoid steric clashes and improper geometries. We relax using the maximum number of rotamers by passing the -EX1, -EX2, -EX3, and -EX4 flags at initialization. We also include the flag -packing:repack_only to disable the design, and the flags -no_his_his_pairE and -multi_cool_annealer 10 to set the number of annealing iterations. For FastRelax conformation ranking, we use Rosetta's ref2015 energy function. Since it is known that performing relaxation on a structure often moves the backbone by several Angstroms1, we include an additional term that includes harmonic distance constraints for all pairs of C / 3 atoms that are not part of the CDR loop or within a distance of 10 to any atom in the CDR loop, based on the conformation of the initial structure. These constraints are given with a weight of 10-4. The protocol is run 10 times for each target, and the decoy with the lowest energy in the HCDR3 loop region is finally selected.

[0153] antibody As described herein, the technology may include a computer system, a computer-implemented method, and / or a computer-readable medium that predicts or generates structural information of a biomolecule such as an antibody. As used herein, the term "antibody" refers to a whole 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 that includes at least two heavy (H) chains and two light (L) chains interconnected by disulfide bonds. Each heavy chain is composed of a heavy chain variable region (abbreviated herein as VH) and a heavy chain constant region. The heavy chain constant region is composed of three domains: CH1, CH2, and CH3. Each light chain is composed of a light chain variable region (abbreviated herein as VL) and a light chain constant region. The light chain constant region is composed of one domain: CL. The VH and VL regions can be further subdivided into regions of hypervariability called complementarity determining regions (CDRs), interspersed with more conserved regions called framework regions (FRs). Each VH and VL is composed of three CDRs and four FRs arranged from amino-terminus to carboxy-terminus in the following order: 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 the antibody can mediate the binding of the 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 (scFvs), disulfide-linked Fvs (sdFvs), Fab fragments, F(ab') fragments, and anti-idiotypic (anti-Id) antibodies (including, for example, anti-Id antibodies to an antibody of the invention), as well as epitope-binding fragments of any of the above. Antibodies 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. Antibodies or epitope-binding fragments may be or be components of multispecific molecules.

[0154] Both the light and heavy chains are divided into regions of structural and functional homology. The terms "constant" and "variable" are used functionally. In this regard, it will be understood that the variable domains of both the light (VL) and heavy (VH) chain portions determine antigen recognition and specificity. Conversely, the constant domains of the light (CL) and heavy (CH1, CH2, or CH3) chains confer important biological properties such as secretion, transplacental transfer, Fc receptor binding, and complement binding. By convention, the numbering of the constant region domains increases as they become more distal from the antigen-binding site or amino-terminus of the antibody. The N-terminus is the variable region and the constant region is at the C-terminus. The CH3 and CL domains actually comprise the carboxy termini of the heavy and light chains, respectively.

[0155] The phrase "antibody fragment" as used herein 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, Fab fragments, which are monovalent fragments consisting of VL, VH, CL, and CH1; F(ab)2 fragments, which are bivalent fragments containing two Fab fragments linked by a disulfide bridge at the hinge region; Fd fragments consisting of VH and CH1 domains; Fv fragments consisting of the VL and VH domains of a single arm of an antibody; dAb fragments consisting of the VH domain (Ward et al., (1989) Nature 341:544-546); and isolated complementarity determining regions (CDRs). Furthermore, although the two domains of the Fv fragment, VL and VH, are encoded by separate genes, they can be joined using recombinant methods by a synthetic linker that allows the VL and VH regions to pair into a single protein chain forming a monovalent molecule (known as single-chain Fv (scFv), see, e.g., 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 of skill in the art, and the fragments are screened for utility in the same manner as intact antibodies.

[0156] As described herein, an antibody may include a biologically active derivative, or variant, or fragment. As used herein, a "biologically active derivative" or a "biologically active variant" includes any derivative or variant of an antibody having substantially the same functional and / or biological properties, such as binding properties, as the antibody (e.g., a WT antibody), and / or the same structural basis, such as a peptide backbone or basic polymer unit including framework regions.

[0157] An "analog", such as a "variant" or "derivative", is an antibody that is substantially similar in structure, and in certain cases has the same biological activity, albeit to a different extent, as a native or WT antibody or another reference antibody, 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 and has the same biological activity as a reference antibody. Variants or analogs differ in their amino acid sequence composition compared to the reference antibody from which the analog is derived, based on one or more mutations including (i) deletion of one or more amino acid residues at one or more termini of the antibody and / or at one or more internal regions (e.g., fragments) of the antibody, (ii) insertion or addition of one or more amino acids at one or more termini of the antibody (typically "additions" or "fusions") and / or at one or more internal regions (typically "insertions") of the antibody sequence, or (iii) substitution of one or more amino acids in the antibody sequence for other amino acids. As an example, a "derivative" refers to an antibody that is a type of analog, and that shares the same or substantially similar structure as a reference antibody that has been chemically modified, for example.

[0158] In some embodiments, the variants or sequence variants are mutants in which one, two, three, four, five, six or more amino acids in one or more CDRs are mutated compared to the reference antibody. In some embodiments, the CDRs on the light chain, the heavy chain, or both the heavy and light chains are mutated. In some embodiments, one or more framework amino acid residues are mutated compared to the reference antibody.

[0159] In substitution variants, one or more amino acid residues of an antibody, for example, in CDR region, are removed and replaced with alternative residues.In one aspect, these substitutions are essentially conservative, and this type of conservative substitution is well known in the art.Alternatively, the present 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].

[0160] Antibodies contemplated 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. Thus, antibodies include those that (1) have an amino acid sequence that has 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 amino acid sequence identity with a reference antibody (e.g., encoded by a reference nucleic acid or amino acid sequence described herein) over a region of at least about 25, about 50, about 100, about 200, about 300, about 400, or more amino acids. In accordance with the present disclosure, the term "recombinant protein" or "recombinant antibody" includes any protein obtained via recombinant DNA technology. In certain aspects, the term encompasses the antibodies described herein.

[0161] In some embodiments, the antibodies or antibody variants described herein are expressed from one or more expression constructs and / or in a cell or line described herein.

[0162] Exemplary wild-type or reference antibodies include commercially available antibodies or other known antibodies, including therapeutic monoclonal antibodies. Reference antibodies according to the present disclosure may include any antibody currently known or later developed, including those that are not clinically and / or commercially available.

[0163] Cells and expression constructs cell The antibodies of the present disclosure include wild-type (WT) and variant antibodies, and in some embodiments are produced in cells. Cells comprising one or more of the expression constructs described herein are contemplated in various embodiments of the present disclosure.

[0164] Prokaryotic host cells. In some embodiments of the present disclosure, expression constructs designed for expression of gene products, including the fusion proteins described herein, are provided in host cells, such as prokaryotic host cells. Prokaryotic host cells can be selected from the group consisting of archaea (e.g., Haloferax volcanii, Sulfolobus solfataricus), gram-positive bacteria (e.g., Bacillus subtilis, Bacillus licheniformis, Brevibacillus choshinensis, Lactobacillus brevis, Lactobacillus buchneri, Lactococcus lactis, and Streptomyces lividans), or alphaproteobacteria (e.g., Agrobacterium tumefaciens). tumefaciens, Caulobacter crescentus, Rhodobacter 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 E. coli), Klebsiella, Proteus, Salmonella (including Salmonella typhimurium), Serratia (including Serratia marcescans), and Shigella.

[0165] 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 species, Pichia pastoris, Saccharomyces cerevisiae, Saccharomyces pastorianus, also known as Saccharomyces carlsbergensis, Schizosaccharomyces pombe, and the like). pombe, Dekkera / Brettanomyces, and Yarrowia lipolytica; other fungi (Aspergillus nidulans, Aspergillus niger, Neurospora crassa, Penicillium, Tolypocladium, Trichoderma lesia, etc.) reesia); insect cell lines (Drosophila melanogaster Schneider 2 cells and Spodoptera frugiperda Sf9 cells); and mammalian cell lines, 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 are available from the American Type Culture Collection.

[0166] As explained in WO / 2017 / 106583, which is incorporated herein by reference in its entirety, the production of gene products such as therapeutic proteins at commercial scale and in soluble form is addressed by providing suitable host cells that can grow at high cell densities in fermentation cultures and produce soluble gene products in an oxidizing host cell cytoplasm through highly controlled inducible gene expression. Host cells of the present disclosure with these qualities are produced by combining some or all of the following features: (1) The host cell is genetically modified to have an oxidizing cytoplasm by increasing the expression or function of an oxidizing polypeptide in the cytoplasm and / or by decreasing the expression or function of a reducing polypeptide in the cytoplasm. Specific examples of such genetic modifications are provided herein. Optionally, the host cell can also be genetically modified to express chaperones and / or cofactors that assist in the production of the desired gene product and / or to glycosylate the polypeptide gene product. (2) The host cell comprises one or more expression constructs designed for expression of one or more gene products of interest, and in certain embodiments, at least one expression construct comprises an inducible promoter and a polynucleotide encoding a gene product expressed from the inducible promoter, and (3) the host cell comprises additional genetic modifications designed to improve particular aspects of gene product expression from the expression construct.In a particular embodiment, the host cell has (A) a modification of gene function of at least one gene encoding a transporter protein for an inducer of at least one inducible promoter, and as another example, the gene encoding the transporter protein is selected from the group consisting of araE, araE, araG, araH, rhaT, xylF, xylG, and xylH, or in particular araE, or more specifically, the modification of gene function is expression of araE from a constitutive promoter, and / or (B) a modification of gene function of at least one gene encoding a protein that metabolizes an inducer of at least one inducible promoter. and / or (C) having reduced levels of gene function of at least one gene encoding a protein involved in the biosynthesis of an inducer of at least one inducible promoter, which in further embodiments is selected from the group consisting of scpA / sbm, argK / ygfD, scpB / ygfG, scpC / ygfH, rmlA, rmlB, rmlC, and rmlD.

[0167] A host cell having an oxidized cytoplasm. The expression system of the present disclosure is designed to express a gene product, and in certain embodiments of the present disclosure, the gene product is expressed in the host cell. Examples of host cells are provided that allow efficient and cost-effective expression of gene products, including components of multimeric products. Host cells can include cells that are part of a multicellular organism, or cells grown in different organisms or systems of organisms, in addition to isolated cells in culture. In certain embodiments of the present disclosure, the host cell is a microbial cell, such as a yeast (such as Saccharomyces, Schizosaccharomyces, etc.), or a bacterial cell, or a Gram-positive or Gram-negative bacterium, or an E. coli, or an E. coli B strain, or an E. coli (B strain) EB0001 cell (also called an E. coli ASE (DGH) cell), or an E. coli (B strain) EB0002 cell. Growth experiments with E. coli host cells with oxidizing cytoplasm, specifically the E. coli B strains SHuffle® Express (NEB Catalog No. C3028H) and SHuffle® T7 Express (NEB Catalog No. C3029H), and the E. coli K strain SHuffle® T7 (NEB Catalog No. C3026H), have shown that these E. coli B strains with oxidizing cytoplasm can grow to much higher cell densities than their closest corresponding E. coli K strain (WO / 2017 / 106583).

[0168] Modifications to host cell gene function. Certain modifications can be made to gene function of host cells containing an inducible expression construct to facilitate efficient and uniform induction of a host cell population by an inducer. Preferably, the combination of expression construct, host cell genotype, and induction conditions results in at least 75% (more preferably at least 85%, and most preferably at least 95%) of the cells in culture expressing a gene product from each of the induced promoters, as measured by the method of Khlebnikov et al., described in Example 9 of WO / 2017 / 106583. For host cells other than E. coli, these modifications can involve genes whose functions are structurally similar to E. coli genes, or genes that perform a function in the host cell similar to that of E. coli genes. Modifications to host cell gene function include removing or reducing gene function by deleting the gene protein coding sequence in its entirety or deleting a sufficiently large portion of the gene, inserting a sequence into the gene, or otherwise modifying the gene sequence such that a reduced level of functional gene product is produced from that gene. Modifications to host cell gene function also include increasing gene function, for example, by modifying a native promoter to create a stronger promoter that drives higher levels of transcription of the gene, or by introducing a missense mutation in a protein coding sequence that results in a more highly active gene product. Modifications to host cell gene function include modifying gene function in any way, including, for example, modifying a native inducible promoter to create a constitutively activated promoter. In addition to modifying gene function for inducer transport and metabolism, as described herein with respect to modifying the expression of inducible promoters and / or chaperone proteins, it is also possible to modify the reduced-oxidative environment of the host cell.

[0169] Host cell reducing-oxidizing environment. In bacterial cells such as E. coli, proteins requiring disulfide bonds are typically transported into the periplasm, where disulfide bond formation and isomerization are catalyzed by the Dsb system, including DsbABCD and DsbG. Increased expression of the cysteine ​​oxidase DsbA, the disulfide isomerase DsbC, or a combination of Dsb proteins, all of which are normally transported into the periplasm, has been utilized in the expression of heterologous proteins that require 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 versions of DsbA and / or DsbC ('cDsbA or 'cDsbC'), which lack the signal peptide and therefore are not transported into the periplasm. Cytoplasmic Dsb proteins, such as cDsbA and / or cDsbC, are useful for making the host cell cytoplasm more oxidized, thus contributing to the formation of disulfide bonds in heterologous proteins produced in the cytoplasm. The host cell cytoplasm can also be made less reducing and thus more oxidized by directly modifying the thioredoxin and glutaredoxin / glutathione enzyme systems: mutants defective in glutathione reductase (gor) or glutathione synthetase (gshB), together with thioredoxin reductase (trxB), confer cytoplasmic oxidation. These strains are unable to degrade ribonucleotides and therefore 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., Microb Cell Fact 2012 May 8;11:56;doi:10.1186 / 1475-2859-11-56) convert it into a disulfide reductase that generates reduced glutathione, allowing channeling of electrons onto the enzyme ribonucleotide reductase, allowing cells defective in gor and trxB, or cells defective in gshB and trxB, to grow in the absence of DTT. Different classes of mutant forms of AhpC can allow strains defective in gamma-glutamylcysteine ​​synthetase (gshA) activity and defective 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 readily oxidized, leading to the formation of disulfide bonds, in a process catalyzed by thioredoxin, reversing their physiological function. Other proteins that may be useful for reducing the effects of oxidative stress in oxidized cytoplasmic host cells are HPI (hydroperoxidase I) catalase peroxidase, encoded by E. coli katG, and HPII (hydroperoxidase II) catalase peroxidase, encoded by E. coli katE, which disproportionate peroxide into water and oxygen (Farr and Kogoma, Microbiol Rev. 1991 Dec;55(4):561-585; Review). Increasing the levels of KatG and / or KatE proteins in host cells through induced co-expression or through increased levels of constitutive expression is an aspect of some embodiments of the present disclosure.

[0170] 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 cell cytoplasm, which has been shown to increase production of a variety of complex disulfide-bonded 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).

[0171] Host cells containing the expression constructs preferably also express cDsbA and / or cDsbC and / or Ervlp, are defective in trxB gene function, are defective in any of gor, gshB or gshA gene function, optionally have increased levels of katG and / or katE gene function, and express an appropriate mutant form of AhpC such that the host cells can grow in the absence of DTT.

[0172] Chaperones. In some embodiments, the desired gene product is co-expressed with other gene products, e.g., chaperones, that are beneficial to the production of the desired gene product. Chaperones are proteins that assist in the folding or unfolding and / or assembly or disassembly of other gene products, but do not occur in the resulting monomeric or multimeric gene product structures when the structures are performing their normal biological function (completing the folding and / or assembly process). Chaperones can be expressed from an inducible or constitutive promoter in an expression construct, or expressed from a host cell chromosome, and preferably, expression of the chaperone protein in the host cell is at a high enough level to produce the co-expressed gene products that are properly folded and / or assembled into the desired product. Examples of chaperones present in E. 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 of cytoplasmic or periplasmic proteins. Expression of these chaperones in combination has been shown to be beneficial for protein expression, as DnaK / DnaJ / GrpE, GroEL / GroES, and ClpB can function synergistically in supporting protein folding (Makino et al., Microb Cell Fact 2011 May 14;10:32). When expressing eukaryotic proteins in prokaryotic host cells, eukaryotic chaperone proteins such as protein disulfide isomerase (PDI) from the same or related eukaryotic species are expressed or inducibly co-expressed with the desired gene product in certain embodiments.

[0173] One chaperone that can be expressed in a host cell is a protein disulfide isomerase from the soil nematode (soft-rot fungus) Humicola insolens. The amino acid sequence of Humicola insolens PDI is set forth as SEQ ID NO:1 in WO / 2017 / 106583 and lacks the signal peptide of the native protein, thereby remaining in the host cytoplasm. The nucleotide sequence encoding PDI was optimized for expression in E. coli and an expression construct for PDI is set forth as SEQ ID NO:2 in WO / 2017 / 106583. SEQ ID NO:2 contains a GCTAGC Nhel restriction site at its 5' end, an AGGAGG ribosome binding site at nucleotides 7-12, the PDI coding sequence at nucleotides 21-1478, and a GTCGAC Sail restriction site at its 3' end. The nucleotide sequence of SEQ ID NO:2 was designed to be inserted immediately downstream of a promoter, such as an inducible promoter. The Nhel and Sail restriction sites of SEQ ID NO:2 can be used to insert this into a vector multiple cloning site, such as that of the pSOL expression vector (SEQ ID NO:3 of WO / 2017 / 106583) described in published U.S. Patent Application No. 2015 / 353940A1, which is incorporated by reference in its entirety. Other PDI polypeptides can also be expressed in host cells, including PDI polypeptides from a variety of species (Saccharomyces cerevisiae (UniProtKB PI 7967), Homo sapiens (UniProtKB P07237), Mus musculus (UniProtKB P09103), Caenorhabditis elegans (UniProtKB Q17770 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 a host cell of the present disclosure shares 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, where the amino acid sequence identity is determined according to Example 10 of WO / 2017 / 106583.

[0174] Cellular transport of cofactors. When using the expression system of the present disclosure to produce enzymes that require cofactors for function, it is helpful to use host cells that can synthesize the cofactor from available precursors or take it up from the environment. Common cofactors include ATP, coenzyme A, flavin adenine dinucleotide (FAD), NAD+ / NADH, and heme. Polynucleotides encoding cofactor transporting and / or cofactor synthesising polypeptides can be introduced into the host cell, and such polypeptides can be expressed constitutively or inducibly co-expressed with the gene products produced by the methods of the present disclosure.

[0175] Glycosylation of polypeptide gene products. Host cells may have modifications in their ability to glycosylate polypeptides. For example, eukaryotic host cells may have eliminated or reduced gene function in glycosyltransferase and / or oligo-saccharyltransferase genes, thereby impairing normal eukaryotic glycosylation of polypeptides to form glycoproteins. Prokaryotic host cells, such as E. coli, that do not normally glycosylate polypeptides, can be modified to express a set of eukaryotic and prokaryotic genes that provide glycosylation functions (DeLisa et al., WO2009089154A2, 2009 Jul 16).

[0176] Available host cell lines with altered genetic function. To generate preferred strains of host cells for use in the expression systems and methods of the present disclosure, it is useful to start with a strain that already contains the desired genetic alterations (Table A, WO / 2017 / 106583). [Table 13]

[0177] Expression constructs In some embodiments of the present disclosure, inducible promoters are contemplated for use with the expression constructs. Exemplary promoters are described herein and also in WO / 2016 / 205570, which is incorporated by reference in its entirety. As described herein, cells comprising one or more expression constructs can optionally comprise one or more inducible promoters for expressing the antibodies of the present disclosure, including wild-type and variant antibodies.

[0178] Expression constructs. An expression construct is a polynucleotide designed for the expression of one or more antibodies, and thus is not a natural molecule. An expression construct can be integrated into a host cell chromosome, or can be maintained in a host cell as a polynucleotide molecule that replicates independently of the host cell chromosome, such as a plasmid or artificial chromosome. An example of an expression construct is a polynucleotide resulting from the insertion of one or more polynucleotide sequences into a host cell chromosome, where the inserted polynucleotide sequence alters the expression of a chromosomal coding sequence. An expression vector is a plasmid expression construct that is specifically used for the expression of one or more gene products, such as one or more antibodies. One or more expression constructs can be integrated into a host cell chromosome, or maintained on an extrachromosomal polynucleotide, such as a plasmid or artificial chromosome. Below is a description of certain types of polynucleotide sequences that can be used in an expression construct for the expression or co-expression of an antibody.

[0179] Origin of replication. Expression constructs must contain an origin of replication, also called a replicon, to be maintained in a host cell as an independently replicating polynucleotide. Different replicons that use the same replication mechanism cannot be maintained together in a single host cell through repeated cell divisions. As a result, plasmids can be classified into incompatibility groups depending on the origin of replication they contain, as shown in Table 2 of WO / 2016 / 205570. Origins of replication can be selected for use in expression constructs based on incompatibility group, copy number, and / or host range, among other criteria. As explained above, when two or more different expression constructs are used in the same host cell for co-expression of multiple antibodies or antibody components (including, for example, heavy and light chains, fragments described herein), in one embodiment, the different expression constructs contain origins of replication from different incompatibility groups: for example, a pMB1 replicon in one expression construct and a pl5A replicon in another expression construct. The average copy number of an expression construct in a cell relative to the number of host chromosomal molecules is determined by the origin of replication contained in the expression construct. The copy number can vary from a few copies to several hundred per cell (Table 2 in WO / 2016 / 205570). In one embodiment of the present disclosure, different expression constructs are used that contain an inducible promoter activated by the same inducer, but with different origins of replication. By selecting an origin of replication that maintains each different expression construct at a certain approximate copy number in the cell, it is possible to adjust the overall production level of an antibody component or fragment (e.g., heavy or light chain) expressed from one expression construct compared to another antibody component or fragment (e.g., heavy or light chain) expressed from a different expression construct. As an example, to co-express subunits A and B (e.g., including heavy and light chains) of a multimeric protein, an expression construct is made that contains the colEl replicon, the am promoter, and the coding sequence for subunit A expressed from the am promoter: colEl-Para-A.

[0180] Another expression construct is made, containing the pl 5A replicon, the am promoter, and the coding sequence for subunit B: 'pl5A-Para-B'. These two expression constructs can be maintained together in the same host cell, and expression of both subunit A and subunit B is induced by adding one inducer, arabinose, to the growth medium. If the expression level of subunit A needs to be significantly increased relative to that of subunit B, a new expression construct for subunit A, pUC9ori-Para-A, can be created, for example, with a modified pMB1 as found in the origin of replication of the pUC9 plasmid ("pUC9ori"), to bring the stoichiometric ratio of the expression amounts of the two subunits closer to the desired ratio. Expressing 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 pl5A-Para-B. Similarly, using an origin of replication that maintains an expression construct at a lower copy number, such as pSOOl (WO / 2016 / 205570), may reduce the overall level of gene product expressed from the construct. The choice of origin of replication may determine which host cells are able to maintain an expression construct containing the replicon.For example, an expression construct containing the colEl origin of replication has a relatively narrow range of available hosts, species within the family Enterobacteriaceae, while an expression construct containing the RK2 replicon can be maintained in E. coli, Pseudomonas aeruginosa, Pseudomonas putida, Azotobacter vinelandii, and Alcaligenes eutrophus, but when an expression construct contains the RK2 replicon and some regulator genes from the RK2 plasmid, it can be maintained in Sinorhizobium meliloti, Agrobacterium tumefaciens, Caulobacter crescentus, Acinetobacter calcoaceticus, and Escherichia coli. It can be maintained in a variety of host cells, such as Bacillus subtilis, Bacillus calcoaceticus, and Rhodobacter sphaeroides (Kiies and Stahl, Microbiol Rev 1989 Dec;53(4):491-516).

[0181] Similar considerations can be used to create expression constructs for inducible or co-expression in eukaryotic cells. For example, the 2 micron circle plasmid of Saccharomyces cerevisiae is compatible with plasmids from other yeast strains, such as pSRl (ATCC Accession Nos. 48233 and 66069; Araki et al., J Mol Biol 1985 Mar 20;182(2):191-203) and pKDl (ATCC Accession No. 37519; Chen et al., Nucleic Acids Res 1986 Jun 11;14(11):4471-4481).

[0182] Selection markers. Expression constructs usually contain a selection gene, also called a selection marker, which encodes a protein necessary for the survival or growth of a host cell in a selective culture medium. Host cells that do not contain an expression construct containing a selection gene will not survive in the culture medium. Typical selection genes encode proteins that confer resistance to antibiotics or other toxins or complement auxotrophic defects of the host cell. One example of a selection scheme utilizes a drug, such as an antibiotic, to stop the growth of the host cell. Those cells that contain an expression construct containing a selection marker produce a protein that confers drug resistance and survive the selection regime. Some examples of antibiotics (and abbreviations that indicate genes that provide an antibiotic resistance phenotype) that are commonly used for selecting selectable markers are ampicillin (AmpR), chloramphenicol (CmlR or CmR), kanamycin (KanR), spectinomycin (SpcR), streptomycin (StrR), and tetracycline (TetR). Many of the plasmids in Table 2 of WO / 2016 / 205570 contain a selectable marker, 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 selectable marker portion of the expression construct, along with the coding sequence for its gene product. Alternatively, the coding sequence for the selection gene can be expressed from a constitutive promoter.

[0183] In various embodiments, suitable selectable markers include, but are not limited to, neomycin phosphotransferase (npt II), hygromycin phosphotransferase (hpt), dihydrofolate reductase (dhfr), zeocin, phleomycin, bleomycin resistance gene (ble), gentamicin acetyltransferase, streptomycin phosphotransferase, a mutant form 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), beta-lactamase (BLA), puromycin N-acetyltransferase (pac), blasticidin acetyltransferase (bls), blasticidin deaminase (bsr), histidinol dehydrogenase (HDH), N-succinyl-5-aminoimidazole-4-carboxamide ribotide (SAICAR) synthetase (ade1), argininosuccinate lyase (arg4), beta-isopropylmalate dehydrogenase (leu2), invertase (suc2), orotidine-5'-phosphate (OMP) decarboxylase (ura3), and orthologs of any of the above.

[0184] Inducible promoter. As described herein, there are several different inducible promoters that can be included in the expression construct as part of the inducible co-expression system of the present disclosure. A preferred inducible promoter shares at least 80% polynucleotide sequence identity (more preferably at least 90% identity, most preferably at least 95% identity) for at least 30 (more preferably at least 40, most preferably at least 50) consecutive bases of the promoter polynucleotide sequence as defined in Table 1 of WO / 2016 / 205570 with reference to the E. coli K-12 substrain MG1655 genome sequence, where the percentage of polynucleotide sequence identity is determined using the method of Example 11 of WO / 2016 / 205570. Under "standard" induction conditions (see Example 5 of WO / 2016 / 205570), preferred inducible promoters have a strength of at least 75% (more preferably at least 100%, most preferably at least 110%) of the corresponding "wild-type" inducible promoter of E. coli K-12 substrain MG1655, 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 positioned 5' (or "upstream") of the coding sequence for a gene product (e.g., an antibody or antibody fragment) to be inducibly expressed, such that the presence of the inducible promoter directs transcription of the gene product coding sequence in the 5' to 3' direction relative to the coding strand of a polynucleotide encoding the gene product.

[0185] Ribosome Binding Site. For some antibodies or antibody fragments, the nucleotide sequence of the region between the transcription initiation site and the start codon of the coding sequence of the inducibly expressed gene product corresponds to the 5' untranslated region ("UTR") of the mRNA for the polypeptide gene product. Preferably, the region of the expression construct corresponding to the 5'UT contains a nucleotide sequence similar to a consensus ribosome binding site (RBS, also called 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 such as E. coli, the RBS consensus sequence is AGGAGG or AGGAGGU. The RBS is typically separated from the start codon by 5-10 intervening nucleotides. In the expression construct, the RBS sequence is preferably at least 55% identical, more preferably at least 70% identical, and most preferably at least 85% identical to the AGGAGGU consensus sequence and is separated from the start codon by 5-10 intervening nucleotides, more preferably 6-9 intervening nucleotides, and most preferably 6 or 7 intervening nucleotides. The ability of a given RBS to generate a desired translation initiation rate can be calculated using the RBS Calculator at the website salis.psu.edu / software / RBSLibraryCalculatorSearchMode, and the same tool can be used to optimize synthetic RBSs for translation rate over a range of 100,000 fold or more (Salis, Methods Enzymol 2011;498:19-42).

[0186] Multiple Cloning Site. A multiple cloning site (MCS), also called a polylinker, is a polynucleotide that contains multiple restriction sites adjacent to or overlapping each other. Restriction sites in an MCS typically occur once within the MCS sequence and preferably do not occur within the remainder of the plasmid or other polynucleotide construct, allowing a restriction enzyme to cut the plasmid or other polynucleotide construct only within the MCS. Examples of MCS sequences are those contained 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 in the pPRO series of expression vectors derived from the pBAD vectors, such as pPR018, pPR018-Cm, pPR018-Kan, pPR024, pPRO30, and pPR033 (U.S. Patent No. 8,178,338 B2; May 15, 2012; Keasling, Jay). A multiple cloning site can be used in making an expression construct by placing the multiple cloning site 3' (or downstream) of the promoter sequence, but using the MCS to insert into the construct a coding sequence for a gene product to be expressed or co-expressed at the appropriate location relative to the promoter so that transcription of the coding sequence occurs. Depending on which restriction enzyme is used to cut within the MCS, there may be some portion of the MCS sequence that remains in the expression construct after a coding sequence or other polynucleotide sequence is inserted into the expression construct. Any remaining MCS sequence can be upstream, downstream, or on either side of the inserted sequence. A ribosome binding site can be located 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 is 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 to cut within the MCS will determine whether the RBS is retained and in what relationship to the inserted sequence. A further alternative is to include an RBS within the polynucleotide sequence inserted into the expression construct at the MCS, preferably in appropriate relationship to any coding sequence to stimulate initiation of translation from the transcribed messenger RNA.

[0187] Expression from a constitutive promoter. The expression constructs of the present disclosure can also include a coding sequence expressed from a constitutive promoter. Unlike an inducible promoter, a constitutive promoter initiates continuous production of a gene product under most growth conditions. One example of a constitutive promoter is that of the Tn3 bla gene, which encodes beta-lactamase and is responsible for the ampicillin resistance (AmpR) phenotype conferred on host cells by many plasmids, including pBR322 (ATCC 31344), pACYClW (ATCC 37031), and pBAD24 (ATCC 87399). Another constitutive promoter that can be used in the expression construct is the promoter for the E. coli lipoprotein gene, which maps to 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). A further example of a constitutive promoter that has been used for heterologous gene expression in E. coli is the trpLEDCBA promoter, which maps to positions 1321169-1321133 (minus strand) in E. coli K-12 substrain MG1655 (Windass et al., Nucleic Acids Res 1982 Nov 11;10(21):6639-6657). Constitutive promoters can be used in expression constructs for the expression of selectable markers, as described herein, and also for the constitutive expression of other gene products useful for co-expression of desired products. For example, inducible promoters, such as transcriptional regulators AraC, PrpR, RhaR, and XylR, if not expressed from a bidirectional inducible promoter, can alternatively be expressed from a constitutive promoter, either in the same expression construct as the inducible promoter they regulate, or in a different expression construct. Similarly, gene products useful for the production or transport of inducers, such as PrpEC, AraE, or Rha, or proteins that modify the redox environment of the cell, as a few examples, can be expressed from constitutive promoters in expression constructs.Co-expressed gene products, and the resulting gene products useful for producing a desired product, also include chaperone proteins, cofactor transporters, and the like.

[0188] Signal peptides. Antibodies or antibody fragments expressed or co-expressed by the disclosed methods can contain or lack a signal peptide, depending on whether it is desired that such gene products be transported from the cytoplasm into the periplasm or retained in the cytoplasm, respectively, of the host cell. Signal peptides (also called signal sequences, leader sequences, or leader peptides) are structurally characterized by a stretch of hydrophobic amino acids, about 5-20 amino acids long, often about 10-15 amino acids long, that has a tendency to form a single alpha-helix. This hydrophobic stretch is often immediately preceded by a shorter stretch rich in positively charged amino acids, especially lysines. Signal peptides that are cleaved from mature polypeptides typically end with a stretch of amino acids that is recognized and cleaved by a signal peptidase. Signal peptides are functionally characterized by their ability to direct the transport of a polypeptide, either co-translationally or post-translationally, through the cell membrane of prokaryotes (or the inner membrane of gram-negative bacteria such as E. coli) or into the endoplasmic reticulum of eukaryotic cells. The extent to which a signal peptide enables a polypeptide to be transported into the periplasmic space of a host cell, such as E. coli, can be determined by separating periplasmic proteins from proteins retained in the cytoplasm, e.g., using methods such as those described in Example 12 of WO / 2016 / 205570.

[0189] Below are some descriptions of inducible promoters that can be used in expression constructs for expression or co-expression of gene products, as well as genetic modifications that can be made to host cells containing such expression constructs. These examples of inducible promoters and associated genes are derived from Escherichia coli (E. coli) strain MG1655 (American Type Culture Collection Accession No. ATCC 700926), which is a substrain of E. coli K-12 (American Type Culture Collection Accession No. ATCC 10798), unless otherwise specified. Table 1 of WO / 2016 / 205570 lists the genomic location in E. coli MG1655 of the nucleotide sequences of these examples of inducible promoters and associated genes. The nucleotide and other gene sequences are referenced by genomic location in Table 1 of WO / 2016 / 205570 and are expressly incorporated herein by reference. Additional information regarding the E. coli promoters, genes, and strains described herein can be found in many public sources, including the online EcoliWiki resource, located at ecoliwiki.net.

[0190] Arabinose promoter. (As used herein, "arabinose" means L-arabinose.) Several E. coli operons involved in arabinose utilization are inducible by arabinose-araBAD, araC, arciE, and araFGH-, however, the terms "arabinose promoter" and "ara promoter" are typically used to designate the araBAD promoter. Several additional terms have been used to refer to the E. coli araBAD promoter, such as Para, ParaB, ParaBAD, and PBAD-. Use herein of "ara promoter" or any of the alternative terms given above means the E. coli araBAD promoter. As can be seen from the use of the alternative term "araC-araBAD promoter," the araBAD promoter is considered to be part of a bidirectional promoter, where the araBAD promoter controls expression of the araBAD operon in one direction and the araC promoter, adjacent to and on the opposite strand from the araBAD promoter, controls expression of the araC coding sequence in the other direction. The AraC protein is both a positive and negative transcriptional regulator of the araBAD promoter. In the absence of arabinose, the AraC protein represses transcription from the PBAD, but in the presence of arabinose, the AraC protein becomes a positive regulatory element that changes its conformation upon binding to arabinose and allows transcription from the PBAD. The araBAD operon encodes proteins that metabolize L-arabinose by converting it to D-xylulose-5-phosphate through the intermediates L-ribulose and L-ribulose phosphate. For the purpose of maximizing induction of expression from an arabinose-inducible promoter, it is useful to remove or reduce the function of AraA, which catalyzes the conversion of L-arabinose to L-ribulose, and optionally also remove or reduce the function of at least one of AraB and AraD.By eliminating or reducing the host cell's ability to reduce the effective concentration of arabinose in the cell, it is possible to eliminate or reduce the cell's ability to convert arabinose into other sugars, making more arabinose available for induction of arabinose-inducible promoters. Genes encoding transporters that move arabinose into the host cell are the araE operon, which encodes the low-affinity L-arabinose proton symporter, and the araFGH operon, which encodes subunits of the ABC superfamily high-affinity L-arabinose transporter. Other proteins that can transport L-arabinose into cells are specific variants of the LacY lactose permease: the LacY(AlWC) and LacY(AlWV) proteins, which have a cysteine ​​or valine amino acid instead of alanine at position 177, respectively (Morgan-Kiss et al., Proc Natl Acad Sci USA 2002 May 28;99(11):7373-7377). To achieve uniform induction of an arabinose-inducible promoter, it is useful to make the transport of arabinose into the cell independent of regulation by arabinose. This can be achieved by removing or reducing the activity of the AraFGH transporter protein and modifying the expression of araE so that it is only transcribed from a constitutive promoter. Constitutive expression of araE can be achieved by removing or reducing the function of the native araE gene and introducing into the cell an expression construct containing the coding sequence for the AraE protein expressed from a constitutive promoter. In a similar manner, as an additional alternative for uniform induction of an arabinose-inducible promoter, a host cell lacking AraE function can have any functional AraFGH coding sequence present in the cell expressed from a constitutive promoter. As another alternative, it is possible to express both the araE gene and the araFGH operon from a constitutive promoter by replacing the native araE and araFGH promoters with constitutive promoters in the host chromosome.It is also possible to use a mutation in the LacY lactose permease that eliminates or reduces the activity of both the AraE and AraFGH arabinose transporters, allowing the protein to transport arabinose in that context. Since expression of the lacY gene is not normally regulated by arabinose, the use of LacY mutants such as LacY(A177C) or LacY(A177V) does not lead to an "all or nothing" induction phenomenon when an arabinose-inducible promoter is induced by the presence of arabinose. The use of a polynucleotide encoding the LacY(A177C) protein is preferred over the use of a polynucleotide encoding the LacY(A177V) protein, since the LacY(A177C) protein appears to be more effective at transporting arabinose into the cell.

[0191] Propionic acid promoter. The "propionic acid promoter" or "prp promoter" is the promoter for the E. coli prpBCDE operon, also referred to as PPΦB. Like the ara promoter, the prp promoter is part of a bidirectional promoter that controls expression of the prpBCDE operon in one direction, and the prpR promoter controls expression of the prpR coding sequence in the other direction. The PrpR protein is a transcriptional regulator of the prp promoter, and activates transcription from the prp promoter when the PrpR protein binds 2-methylcitric acid ("2-MC"). Propionic acid (also called propanoic acid) is the ion of propionic acid (or propanoic acid), CH3CH2COO-, the smallest of the "fatty" acids with the general formula H(CH2),,COOH that share certain properties of this class of molecules: they produce an oily layer when salinizing water, and have a soapy potassium salt. Commercially available propionic acid is generally sold as a monovalent cation salt of propionic acid, such as sodium propionate (CH3CH2COONa), or as a divalent cation salt, such as calcium propionate (Ca(CH3CH2COO)2). Propionic acid is membrane permeable and is metabolized to 2-MC by conversion of propionate to propionyl-CoA by PrpE (propionyl-CoA synthetase) and then conversion of propionyl-CoA to 2-MC by PrpC (2-methylcitrate synthase). Other proteins encoded by the prpBCDE operon, PrpD (2-methylcitrate dehydratase) and PrpB (2-methylisocitrate lyase), are involved in further catabolism to smaller products such as pyruvate and succinate. To maximize induction of a propionate-inducible promoter by propionate added to the cell growth medium, it is therefore desirable to have a host cell with PrpC and PrpE activity to convert propionate to 2-MC, but also with eliminated or reduced PrpD activity, and optionally eliminated or reduced PrpB activity, to prevent 2-MC from being metabolized.Another operon encoding proteins involved in 2-MC biosynthesis is the scpA-argK-scpBC operon, also called the sbm-yg / DGH operon. These genes encode proteins required for the conversion of succinate to propionyl-CoA, which can then be converted to 2-MC by PrpC. Removal or reduction of the function of these proteins may eliminate a parallel pathway for the production of 2-MC inducer, thus reducing the background level of expression of the propionate-inducible promoter and increasing the sensitivity of the propionate-inducible promoter to exogenously supplied propionate. The deletion of sbm-ygfD-ygfG-ygfH-ygfl was introduced into E. coli BL21(DE3) to generate the JSB strain (Lee and Keasling, Appl Environ Microbiol 2005 Nov;71(11):6856-6862) and was found to be useful in reducing background expression in the absence of exogenously supplied inducer, but this deletion also reduced overall expression from the prp promoter in the JSB strain. However, it should be noted that the deletion sbm-ygfD-ygfG-ygfH-ygfl also apparently affects ygfl, which encodes a putative LysR family transcriptional regulator of unknown function. The genes sbm-yg / DGH are transcribed as one operon and ygfl is transcribed from the opposite strand. The 3' ends of the ygfti and ygfl coding sequences overlap by a few base pairs, so a deletion that removed all of the sbm-yg / DGH operon also apparently removed the ygfl coding function.Removing or reducing the function of a subset of the sbm-ygfDGH gene products, such as YgfG (also called 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 ygfl is unaffected, may be sufficient to reduce background expression from a propionate-inducible promoter without reducing the maximum level of induced expression.

[0192] Rhamnose promoter. (As used herein, "rhamnose" means L-rhamnose.) The "rhamnose promoter" or "rha promoter", or PrhaSR, is the promoter for the E. coli rhaSR operon. Like the ara and prp promoters, the rha promoter is part of a bidirectional promoter, with the rhaBAD promoter controlling expression of the rhaSR operon in one direction and the rhaBAD operon in the other direction. The rha promoter, however, has two transcriptional regulators involved in regulating expression: RhaR and RhaS. The RhaR protein activates expression of the rhaSR operon in the presence of rhamnose, while the RhaS protein activates expression of rhaBAD and rhaT, the L-rhamnose catabolic and transport operons, respectively (Wickstrum et al, J Bacteriol 2010 Jan;192(1):225-232). The RhaS protein can also activate expression of the rhaSR operon, but in fact RhaS negatively autoregulates this expression by interfering with the ability of RhaR and the cyclic AMP receptor protein (CRP) to coactivate expression to much higher levels. The rhaBAD operon encodes the rhamnose catabolic proteins RhaA (L-rhamnose isomerase), which converts L-rhamnulose to L-rhamnulose; RhaB (rhamnulokinase), which phosphorylates L-rhamnulose to form L-rhamnulose-1-P; and RhaD (rhamnulose-1-phosphate aldolase), which converts L-rhamnulose-1-P to L-lactaldehyde and DHAP (dihydroxyacetone phosphate). To maximize the amount of rhamnose in the cell available for induction of expression from a rhamnose-inducible promoter, it is desirable to reduce the amount of rhamnose that is catalytically degraded by removing or reducing the function of RhaA, or optionally of at least one of RhaA and RhaB and RhaD.E. coli cells can also synthesize L-rhamnose from alpha-D-glucose-1-P through the activity of the proteins RmlA, RmlB, RmlC, and RmlD (also called RfbA, RfbB, RfbC, and RfbD, respectively) encoded by the rmlBDACX (or rfbBDACX) operon. To reduce background expression from a rhamnose-inducible promoter and to enhance the sensitivity of induction of the rhamnose-inducible promoter by exogenously supplied rhamnose, it may be useful to remove or reduce the function of one or more of the RmlA, RmlB, RmlC, and RmlD proteins.

[0193] L-rhamnose is transported into the cell by RhaT, rhamnose permease or L-rhamnose:proton symporter. As described above, expression of RhaT is activated by the transcriptional regulator RhaS. To make expression of RhaT independent of induction by rhamnose (which induces expression of RhaS), the host cell can be modified so that all functional RhaT coding sequences in the cell are expressed from a constitutive promoter. In addition, the coding sequence for RhaS can be deleted or inactivated such that no functional RhaS is produced. By removing or reducing the function of RhaS in the cell, the level of expression from the rhaSR promoter increases due to the lack of negative autoregulation by RhaS, the level of expression of the rhamnose catalytic operon rhaBAD decreases, and the ability of rhamnose to induce expression from the rha promoter is further increased.

[0194] Xylose promoter. (As used herein, "xylose" means D-xylose.) Xylose promoter, or "xyl promoter", or PxyiA, refers to the promoter for the E. coli xylAB operon. The xylose promoter region is similar in organization to other inducible promoters in that both the xylAB operon and the xylFGHR operon are expressed in opposite orientations on the E. coli chromosome from adjacent xylose-inducible promoters (Song and Park, J Bacteriol. 1997 Nov;179(22):7025-7032). The transcriptional regulator of both the PxyiA and PxyiF promoters is XylR, which activates expression of these promoters in the presence of xylose. The xylR gene is expressed either as part of the xylFGHR operon or from its own weak promoter, not inducible by xylose, located between the xylH and xylR protein coding sequences. D-xylose is catabolized by XylA (D-xylose isomerase), which converts D-xylose to D-xylulose, which is then phosphorylated by XylB (xylulokinase) to form D-xylulose-5-P. To maximize the amount of xylose available in the cell for induction of expression from a xylose-inducible promoter, it is desirable to reduce the amount of xylose catalytically degraded by removing or reducing the function of at least XylA, or possibly both XylA and XylB. The xylFGHR operon encodes XylF, XylG, and XylH, subunits of the ABC superfamily high affinity D-xylose transporter. The xylE gene encodes the E. coli low affinity xylose-proton symporter and represents a separate operon, the expression of which is also inducible by xylose. To make expression of the xylose transporters independent of induction by xylose, the host cells can be engineered such that all functional xylose transporters are expressed from a constitutive promoter.For example, the xylFGHR operon can be modified such that the xylFGH coding sequence is deleted, leaving XylR as 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 a PxyiA or promoter in an expression construct, while the xylFGHR operon is deleted and xylE is constitutively expressed, or alternatively, the xylFGH operon (lacking the xylR coding sequence because it is present in the expression construct) is expressed from a constitutive promoter and the xylE coding sequence is deleted or modified so that it does not produce an active protein.

[0195] Lactose promoter. The term "lactose promoter" refers to the lactose-inducible promoter for the lacZYA operon, also known as lacZpl. The lactose promoter is located at approximately 365603-365568 in the genomic sequence of E. coli K-12 substrain MG1655 (NCBI Reference Sequence NC 000913.2, 11-JAN-2012) (minus strand, NA polymerase binding ("-35") site is at approximately 365603-365598, Pribnow box ("-10") is at 365579-365573, and transcription start site is at 365567). In some embodiments, the inducible expression system of the present disclosure may include a lactose-inducible promoter, such as the lacZYA promoter. In other embodiments, the inducible expression system of the present disclosure includes one or more inducible promoters that are not lactose-inducible promoters.

[0196] Alkaline phosphatase promoter. The terms "alkaline phosphatase promoter" and "phoA promoter" refer to the promoter for the phoApsiF operon, a promoter that is induced under conditions of phosphatase starvation. The phoA promoter region maps to about 401647-401746 (positive strand, Pribnow box ('-10') at 401695-401701 (Kikuchi et al., Nucleic Acids Res 1981 Nov 11;9(21):5671 -5678)) in the genome sequence of E. coli K-12 substrain MG1655 (NCBI Reference Sequence NC 000913.3, 16 Dec. 2014). The transcriptional activator for the phoA promoter is PhoB, a transcriptional regulator that, together with the sensor protein PhoR, forms a two-component signal transduction system in E. coli. PhoB and PhoR are transcribed from the phoBR operon located at about 417050-419300 (positive strand, with the PhoB coding sequence at 417,142-417,831 and the PhoR coding sequence at 417,889-419,184) in the genome sequence of E. coli K-12 substrain MG1655 (NCBI Reference Sequence NC 000913.3, December 16, 2014). The phoA promoter differs from the inducible promoters described above in that it is induced by the lack of a substance - intracellular phosphate - rather than by the addition of an inducer. For this reason, the phoA promoter is generally used to direct the transcription of gene products produced at stages when the host cell is starved for phosphate, e.g., in the later stages of fermentation. In some embodiments, the inducible co-expression system of the present disclosure can 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.

[0197] Affinity assays As described herein, the technology may include a computer system, a computer-implemented method, and / or a computer-readable medium that predicts or generates structural information of a biomolecule such as an antibody. In some aspects, the computer system predicts, for example, the binding affinity between an antibody and an antigen. Nevertheless, wet-lab techniques are contemplated to (i) confirm the predicted affinity or (ii) generate data to train the affinity model described herein. Antibody binding and antibody affinity determination assays are well known in the art.

[0198] In one embodiment, the activity specific cell enrichment method (ACE) can be used to identify host cells expressing "active" antibodies rather than "inactive material". Active antibodies can be distinguished from inactive antibodies by the ability of active antibodies to specifically bind to a binding partner molecule (e.g., an antigen or epitope). The ACE assay protocol is described in WO / 2021 / 146626, which is incorporated herein by reference. It will be understood by those skilled in the art that ACE can not only discriminate between active / inactive in a binary manner, but can also calculate a score proportional to affinity. In this way, ACE provides quantitative assay information rather than simply binary / Boolean information, thereby enabling the modeling techniques herein to perform regression techniques. This richer modeling represents an advantageous improvement over the limited binary classification of prior art techniques.

[0199] In another embodiment, the HiPR binding assay described in WO / 2021 / 163349, and incorporated herein by reference, is used in conjunction with the methods provided herein.

[0200] Binding assays, for example, assays for measuring protein-protein interactions, including antibody-antigen interactions and measuring binding affinity, are well known in the art. As examples, surface plasmon resonance (SPR), dual polarization interferometry (DPI), static light scattering (SLS), dynamic light scattering (DLS), flow-induced dispersion analysis (FIDA), fluorescence polarization / anisotropy, fluorescence resonance energy transfer (FRET), biolayer interferometry (BLI), isothermal titration calorimetry (ITC), microscale thermophoresis (MST), and single-color reflectometry (SCORE) are contemplated. In addition, bimolecular fluorescence complementation (BiFC), affinity electrophoresis, label transfer, phage display, tandem affinity purification (TAP), crosslinking, quantitative immunoprecipitation combined with knockdown (QUICK), and proximity ligation assay (PLA) are other well-known assays that provide protein-protein interaction information.

[0201] 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 were immobilized on a HC 30M chip at four different densities / antibody concentrations. Varying concentrations (0-500 nM) of antibody target are then bound to the captured antibodies. Kinetic analysis was performed using Carterra software to determine the association and dissociation rate constants (k a and k d) Extract the apparent affinity constant (K D ) is k d / k aIn some embodiments, the kinetics and affinity are determined using a Carterra LSA platform. In other embodiments, the binding affinity can be measured by surface plasmon resonance (e.g., BIAcore™) using, for example, an IBIS MX96 SPR system from IBIS Technologies or a Carterra LSA SPR platform, or by biolayer interferometry using, for example, an Octet™ system from ForteBio. In some embodiments, biosensor instruments such as Octet RED384, ProteOn XPR36, IBIS MX96, and Biacore T100 are used (Yang, D., et al., J. Vis. Exp., 2017, 122:55659).

[0202] K D is the equilibrium dissociation constant between an antibody and its antigen, k off / k on is the ratio of K D and affinity are inversely related. D The value is related to the concentration of the antibody, so K D The lower the value (lower concentration), thus the higher the affinity of the antibody. Antibodies, including reference antibodies and variant antibodies, can be used, for example, in accordance with various embodiments of the present disclosure, such as K D is in the micromolar range (10 -4 ~10 -6 ), nanomolar range (10 -7 ~10 -9 ), picomolar range (10 -10 ~10 -12 ), or in the femtomole range (10 -13 ~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 compared to the reference antibody. This improvement may also be in the range of a fold change (e.g., 2x, 4x, 6x, or a 2-, 3-, 4-, 5-, 6-, 7-, 8-, 9-, 10-fold or more improvement in binding activity) and / or an order of magnitude (e.g., 10 7 , 10 8, 10 9 It can be expressed in terms of multiples of 1000, ...20

[0203] Data generated from the antibodies and assays described herein are, in some embodiments, used to train one or more models, as described next.

[0204] Additional Considerations Before the present disclosure is further described, it is to be understood that this disclosure is not limited to particular embodiments described, as such may, of course, vary. It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only, and is not intended to be limiting, since the scope of the present disclosure will be limited only by the appended claims.

[0205] When a range of values ​​is provided, it is understood that each intervening value between the upper and lower limits of that range, to one-tenth of the unit of the lower limit unless otherwise clearly indicated by context, and any other stated or intervening value in that stated range, is encompassed within the disclosure. The upper and lower limits of these smaller ranges may independently be included in the smaller ranges, and are also encompassed within the disclosure, subject to any specifically excluded limit in the stated range. The upper and lower limits of these smaller ranges may independently be included in the smaller ranges, and are also encompassed within the disclosure, subject to any specifically excluded limit in the stated range.

[0206] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs. Although any methods and materials similar or equivalent to those described herein can also be used in the practice or testing of this disclosure, the preferred methods and materials are described herein. All publications mentioned herein are incorporated by reference to disclose and describe the relevant methods and / or materials for which the publications are cited.

[0207] Note that, as used herein, the singular forms "a," "an," and "the" include plural referents unless the context clearly indicates otherwise. Thus, for example, a reference to a "conformation-switching probe" includes a plurality of such conformation-switching probes, a reference to a "microfluidic device" includes a reference to one or more microfluidic devices and equivalents thereof known to those skilled in the art, and so forth. It is further noted that the claims may be drafted to exclude any element, e.g., any optional element. As such, this statement is intended to serve as a premise basis for the use of exclusive terminology, such as "solely," "only," and the like, in connection with the recitation of claim elements or the use of "negative" limitations.

[0208] As will be apparent to one of ordinary skill in the art upon reading this disclosure, each of the individual embodiments described and illustrated herein has separate components and features that may be readily separated from or combined with the features of any of the other several embodiments without departing from the scope or spirit of the disclosure. Any recited method may be carried out 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.

[0209] The various embodiments described above can be combined to provide further embodiments. All U.S. patents, U.S. patent application publications, U.S. patent applications, foreign patents, foreign patent applications, and non-patent publications mentioned herein and / or listed in the Application Data Sheet are incorporated herein by reference in their entirety. Aspects of the embodiments can be modified as necessary to use concepts from the various patents, applications, and publications to provide still further embodiments.

[0210] These and other changes can be made to the embodiments in light of the above detailed description. Generally, in the following claims, the terms used should not be construed to limit the claims to the specific embodiments disclosed in the specification and claims, but should be construed to include all possible embodiments, along with the full scope of equivalents to which such claims are entitled. Accordingly, the claims are not limited by this disclosure.

Claims

1. 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 storing instructions that, when executed by the one or more processors, cause the computing system to: receiving one or more training inputs, the 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; processing the one or more training inputs with a machine-learned biomolecule prediction model to generate predicted biomolecule structural information; evaluating a loss function that compares the predicted biomolecular structure information with ground truth values; A computing system that modifies one or more values ​​of one or more parameters of the machine-learned model 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 biomolecular structural information comprises three-dimensional coordinates of a primary sequence.

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

5. The training inputs each include, in the non-transitory computer readable medium:

2. The computing system of claim 1, wherein the data is represented as at least one of: (i) a Protein Data Bank (PDB) data format; (ii) a JSON data format; or (iii) an XML data format.

6. 2. The computing system of claim 1, wherein the training input includes at least one, at least five, at least ten, or at least one hundred of the binders of Table E1, Table E2, Table E3, and / or Table E4 above.

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

8. The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to: The computing system of claim 1 , further configured to receive the training input from a database comprising antibody and / or antigen structures.

9. The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to: The computing system of claim 1 , further comprising: a computing system configured to receive the training input from a structural antibody database (SAbDab).

10. 1. A computing system for generating structural information of a target biomolecule, comprising: one or more processors; One or more non-transitory computer-readable media, comprising: A machine-learned biomolecule prediction model trained to predict structural information of a biomolecule based on an input; When executed by the one or more processors, the computing system includes: (1) receiving a target input comprising one or more of 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; (2) one or more non-transitory computer-readable media storing instructions for predicting the structural information of the target biomolecule by processing the target input with the machine-learned biomolecule prediction model.

11. 11. The computing system of claim 10, wherein the target input comprises structures of CDRH1 and CDRH2 of trastuzumab-Bh1 and VEGF.

12. 11. The computing system of claim 10, wherein the target input comprises the structures of CDRH1 and CDRH2 of trastuzumab and HER2.

13. The structural information, in the computer-readable medium, comprises:

11. The computing system of claim 10, wherein the data is represented as at least one of: (i) a Protein Data Bank (PDB) data format; (ii) a JSON data format; or (iii) an XML data format.

14. The computing system of claim 1 or 10, wherein the machine-learned biomolecular prediction model is an artificial neural network.

15. 15. The computing system of claim 14, wherein the artificial neural network comprises at least one of a geometric neural network, a transformer network, or a geometric vector perceptron network.

16. The structural information of the target biomolecule is (i) an amino acid of the target biomolecule; (ii) a 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 11. The computing system of claim 10, further comprising: (viii) data representing one or more three-dimensional coordinates of a primary sequence of the target biomolecule.

17. The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to: (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 20. The computing system of claim 16, further comprising: (viii) outputting said data representing said one or more of said three-dimensional coordinates of a primary sequence of said target biomolecule.

18. 20. The computing system of claim 17, 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.

19. 20. The computing system of claim 17, wherein the polypeptide sequence of the target biomolecule corresponds to a complementarity determining region of an antibody.

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

21. The computing system of claim 10 , wherein the target biomolecule is an antibody and the target binding partner is an antigen.

22. The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to: The computing system of claim 10 , further comprising: a computing system configured to repeatedly perform step 2 until the structural information of the target biomolecule is complete.

23. The computing system of claim 22 , wherein the repeating is performed iteratively.

24. The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to: receiving input parameters specifying a desired length of amino acids of the target biomolecule; The computing system of claim 10 , wherein the computing system generates a target biomolecule having the desired length.

25. The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to: The computing system of claim 10 , further comprising: a computing system for updating the structural information of the target biomolecule by folding one or more proteins.

26. 11. The computing system of claim 10, wherein the target binding partner is at least one of: (i) a primary amino acid sequence, (ii) an antigen epitope, and / or (iii) a known three-dimensional structure of an antigen.

27. 11. The computing system of claim 10, wherein the structural information of the target biomolecule is predicted in at least one of: (i) N-terminal to C-terminal order, (ii) C-terminal to N-terminal order, (iii) via random sampling, or (iv) via non-random sampling.

28. The one or more non-transitory computer readable media, and further storing a machine-learned docking model trained to generate a docked complex from two or more input three-dimensional biomolecules; The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to: The computing system of claim 10 , further comprising: a docking model for processing the target input and the predicted structural information of the target biomolecule to generate a docking complex comprising the target input and the target biomolecule.

29. The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to:

30. The computing system of claim 28, wherein the docking complex is provided as a docking complex output.

30. The one or more non-transitory computer readable media, and further storing a machine-learned binding affinity prediction model trained to predict binding affinity from an input docked complex; The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to:

30. The computing system of claim 28, wherein the trained affinity prediction model is used to process the docked complex to generate a binding affinity value.

31. 31. The computing system of claim 30, wherein the machine-learned model is an artificial neural network.

32. 32. The computing system of claim 31, 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.

33. The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to:

32. The computing system of claim 31, wherein a pooling operation is used to summarize node information contained within the artificial neural network to generate the binding affinity value.

34. The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to:

32. The computing system of claim 31, causing the docking complex to be provided as a docking complex output.

35. The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to: The computing system of claim 10 , further comprising: training the machine-learned biomolecule prediction model by providing sequence data as input to supervise the biomolecule prediction model.

36. The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to:

30. The computing system of claim 28, further comprising: training said docking model by providing antibody structural data as input and subjecting said docking model to supervised constraints.

37. The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to:

31. The computing system of claim 30, wherein the affinity prediction model is trained by providing sequence data as input and subjecting the affinity prediction model to supervised training.

38. The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to:

31. The computing system of claim 30, wherein the machine-learned biomolecule prediction model is trained using at least one, at least five, at least ten, or at least one hundred of the binders of Tables E1, E2, E3, and / or E4.

39. 1. A computing system for predicting affinity of a target biomolecule, comprising: one or more processors; One or more non-transitory computer-readable media, comprising: A machine-learned affinity prediction artificial neural network, comprising: (i) one or more biomolecular prediction layers trained to predict biomolecular 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; (iii) one or more affinity prediction layers trained to predict affinity from an input docking complex; a machine-learned affinity prediction artificial neural network to which the one or more biomolecule prediction layers, the one or more docking layers, and the one or more affinity prediction layers are connected; 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 one or more non-transitory computer readable media comprising instructions for processing the target input using the affinity prediction artificial neural network to generate docking complexes corresponding to the target input and corresponding structure affinity values.

40. The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to: receiving a sequence affinity prediction value from an affinity sequencing prediction model; 40. The computing system of claim 39, wherein the structural affinity values ​​are averaged with the sequence affinity prediction values ​​to generate an ensemble affinity prediction value.

41. The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to:

40. The computing system of claim 39, wherein the one or more biomolecule prediction layers are pre-trained using bound or unbound structures.

42. The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to:

40. The computing system of claim 39, wherein a binding structure is used to pre-train the one or more biomolecular prediction layers and the one or more docking layers.

43. The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to:

40. The computing system of claim 39, wherein affinity training data is used to pre-train the machine-learned affinity prediction artificial neural network.

44. 40. The computing system of claim 39, wherein the affinity training data comprises an affinity score proportional to activity.

45. The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to: receiving an external computational docking complex corresponding to said target input; 40. The computing system of claim 39, wherein the external computational docking complex is compared to the generated docking complex.

46. The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to:

40. The computing system of claim 39, further comprising: controlling a gradient flow of the machine-learned affinity prediction artificial neural network by applying a stopping gradient function to at least one of the one or more biomolecular prediction layers, the one or more docking layers, or the one or more affinity prediction layers.

47. The one or more non-transitory computer readable media have further instructions stored thereon, the further instructions, when executed by the one or more processors, causing the computing system to:

40. The computing system of claim 39, further comprising: training a neural network using at least one, at least five, at least ten, or at least one hundred of the binders of Table E1, Table E2, Table E3, and / or Table E4.

Images