Machine learning techniques for predicting thermal stability

Abstract

Techniques for computationally screening a set of single chain variable fragments (scFvs). The techniques include determining a thermal stability indication for each scFv in a set of scFvs using a machine learning model to obtain a plurality of thermal stability indications, the set of scFvs comprising a first scFv having a first residue sequence, the determining comprises: obtaining an interaction energy metric for each of a plurality of pairs of residues in the first residue sequence using information indicative of a 3D structure of the first scFv; generating a first set of features using the interaction energy metrics; and providing the first set of features as input to the machine learning model to obtain a corresponding output indicative of a first thermal stability of the first scFv; identifying a subset of the set of scFvs for subsequent production based on the plurality of thermal stability indications; and producing at least one scFv of the identified scFvs.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims the benefit of priority under 35 U.S.C. §119(e) to U.S. Provisional Patent Application Serial No. 63 / 340,332 (Agent Docket No. A1350.70001US00), filed on May 10, 2022, and entitled “MACHINELEARNING TECHNIQUES FOR PREDICTING SINGLE-CHAIN ​​VARIABLE FRAGMENT (SCFV) THERMOSTABILITY,” which is incorporated herein by reference in its entirety. Background Art

[0003] Monoclonal antibodies (mAbs) represent a large class of therapeutic agents, with more than 100 FDA-approved products marketed in the United States. Multispecific biologics that engage more than one target or epitopes on the same target are increasingly important for gaining access to novel therapeutically relevant pathways and mechanisms of action. Several multispecific biologics have been approved for use, and many more are in clinical and preclinical development.

[0004] A common building block for constructing multispecific biologics is the single-chain variable fragment (scFv), which consists of a target-engaging antibody variable heavy chain (VH) and a variable light chain (VL) connected via a flexible linker. Multispecific format platforms such as BiTE, IgG-scFv, and XmAb incorporate scFv modules. Summary of the invention

[0005] Some embodiments provide a method for computationally screening a set of single-chain variable fragments (scFvs) based on the thermal stability of the scFvs predicted by a trained machine learning model, the set of scFvs including scFvs having different residue sequences, the method comprising: determining a thermal stability indication for each scFv in the set of scFvs using the trained machine learning model and at least one computer hardware processor to obtain a plurality of thermal stability indications, the set of scFvs including a first scFv having a first residue sequence, the determining comprising: obtaining interaction energy metrics for each of a plurality of pairs of residues, the residues being in the first residue sequence, using information indicative of a three-dimensional (3D) structure of the first scFv; generating a first feature set to provide as input to the trained machine learning model, the generating comprising including the interaction energy metrics in the first feature set; and providing the first feature set as input to the trained machine learning model to obtain a corresponding output indicative of a first thermal stability of the first scFv; identifying a subset of the set of scFvs for subsequent production based on the plurality of thermal stability indications; and producing at least one of the scFvs in the identified subset.

[0006] In some embodiments, the set of scFvs further includes a second scFv different from the first scFv, the second scFv having a second residue sequence. In some embodiments, determining the thermal stability indication of each scFv in the set of scFvs further includes: obtaining a second interaction energy metric for each of a second plurality of pairs of second residues, the second residues being in the second residue sequence; generating a second feature set to provide as input to the trained machine learning model, the generating including including the second interaction energy metrics in the second feature set; and providing the second feature set as input to the trained machine learning model to obtain a corresponding output indicating a second thermal stability of the second scFv.

[0007] In some embodiments, the output indicative of a first thermal stability of the first scFv indicates a first temperature at which the first scFv is thermally stable.

[0008] In some embodiments, the first temperature is an estimate of the temperature corresponding to half-maximal binding of the first scFv.

[0009] In some embodiments, the output indicative of the first thermal stability of the first scFv indicates a first temperature range that includes at least one temperature at which the first scFv is thermally stable.

[0010] In some embodiments, the first temperature range is an estimate of the temperature range that includes the temperature corresponding to half-maximal binding of the first scFv.

[0011] In some embodiments, providing the first feature set as input to the trained machine learning model to obtain an output indicating a first thermal stability of the first scFv includes: using the trained machine learning model, classifying the first scFv into one of a plurality of categories using the first feature set, wherein each of the plurality of categories corresponds to a respective temperature range.

[0012] In some embodiments, obtaining the interaction energy metrics comprises determining information indicative of the 3D structure of the first scFv by generating information indicative of the 3D structure from the first residue sequence using protein structure prediction software.

[0013] In some embodiments, obtaining the interaction energy metrics comprises determining the interaction energy metrics using molecular modeling software to generate the interaction energy metrics using information indicative of the 3D structure of the first scFv.

[0014] In some embodiments, generating the first feature set includes: for each specific energy metric among these interaction energy metrics, generating a corresponding two-dimensional (2D) matrix of the values ​​of the specific energy metric, wherein the rows and columns of the 2D matrix correspond to the corresponding residues in the first residue sequence, and wherein the entries in the i-th row and the j-th column of the 2D matrix correspond to the values ​​of the specific energy metric of the i-th residue in the first residue sequence and the j-th residue in the first residue sequence; and including the generated 2D matrix in the first feature set.

[0015] In some embodiments, the generated 2D matrix includes rows for at least 75% of the residues in the first residue sequence. In some embodiments, the generated 2D matrix includes rows for at least 90% of the residues in the first residue sequence. In some embodiments, the generated 2D matrix includes rows for at least 95% of the residues in the first residue sequence. In some embodiments, the generated 2D matrix includes rows for at least 99% of the residues in the first residue sequence.

[0016] In some embodiments, the generated 2D matrix includes a row for each residue in the first residue sequence.

[0017] In some embodiments, generating the first feature set further comprises: encoding the first residue sequence to obtain an encoding sequence; and including the encoding sequence in the first feature set.

[0018] In some embodiments, encoding the first residue sequence comprises: one-hot encoding the first residue sequence to obtain the encoded sequence, the encoded sequence comprising a one-hot encoded version of the first residue sequence.

[0019] In some embodiments, the trained machine learning model comprises a trained neural network model.

[0020] In some embodiments, the trained neural network model comprises a trained convolutional neural network (CNN) model having a plurality of 2D convolutional layers.

[0021] In some embodiments, the trained CNN model further includes a fully connected layer.

[0022] In some embodiments, the trained CNN model is configured to output a plurality of probabilities that the scFv is thermally stable within each of a plurality of temperature ranges.

[0023] In some embodiments, providing the first feature set as input to the trained machine learning model to obtain a corresponding output indicative of a first thermal stability of the first scFv comprises: providing the first feature set to the trained CNN model to obtain a first plurality of probabilities that the first scFv is thermally stable within each of the plurality of temperature ranges; and determining the first thermal stability as: (i) a temperature range in the plurality of temperature ranges associated with a highest probability in the first plurality of probabilities; or (ii) a temperature determined as a weighted linear combination of averages weighted by the temperature ranges in the first set of probabilities.

[0024] In some embodiments, identifying a subset of the set of scFvs for subsequent production based on the plurality of determined thermal stability indicators comprises: determining whether a first thermal stability of the first scFv satisfies at least one criterion; and after determining that the first thermal stability satisfies the at least one criterion, identifying the first scFv for subsequent production.

[0025] Some embodiments further comprise testing the thermal stability of at least one of the scFvs in an in vitro assay.

[0026] Some embodiments provide a method for predicting the thermal stability of a single-chain variable fragment (scFv) using a trained machine learning model, the method comprising: determining a first thermal stability indication of a first scFv using the trained machine learning model and at least one computer hardware processor, the first scFv having a first residue sequence: obtaining interaction energy metrics for each of a plurality of pairs of residues in the first residue sequence using information indicative of a three-dimensional (3D) structure of the first scFv; generating a first feature set to provide as input to the trained machine learning model, the generating comprising including the interaction energy metrics in the first feature set; and providing the first feature set as input to the trained machine learning model to obtain a corresponding output indicative of a first thermal stability of the first scFv.

[0027] Some embodiments provide a system comprising: at least one computer hardware processor; and at least one non-transitory computer-readable storage medium storing processor-executable instructions, which when executed by the at least one computer hardware processor cause the at least one computer hardware processor to perform a method for predicting the thermal stability of a single-chain variable fragment (scFv) using a trained machine learning model, the method comprising: determining a first thermal stability indication of a first scFv using the trained machine learning model and the at least one computer hardware processor, the first scFv having a first residue sequence: obtaining interaction energy metrics for each of a plurality of pairs of residues in the first residue sequence using information indicative of a three-dimensional (3D) structure of the first scFv; generating a first feature set to provide as input to the trained machine learning model, the generating comprising including the interaction energy metrics in the first feature set; and providing the first feature set as input to the trained machine learning model to obtain a corresponding output indicative of a first thermal stability of the first scFv.

[0028] Some embodiments provide at least one non-transitory computer-readable storage medium storing processor-executable instructions that, when executed by at least one computer hardware processor, cause the at least one computer hardware processor to perform a method for predicting the thermal stability of a single-chain variable fragment (scFv) using a trained machine learning model, the method comprising: determining a first thermal stability indication of a first scFv using the trained machine learning model and at least one computer hardware processor, the first scFv having a first residue sequence: obtaining interaction energy metrics for each of a plurality of pairs of residues in the first residue sequence using information indicative of a three-dimensional (3D) structure of the first scFv; generating a first feature set to provide as input to the trained machine learning model, the generating comprising including the interaction energy metrics in the first feature set; and providing the first feature set as input to the trained machine learning model to obtain a corresponding output indicative of a first thermal stability of the first scFv.

[0029] Some embodiments further comprise determining a thermal stability indication for each scFv in a set of scFvs including the first scFv using the trained machine learning model and the at least one computer hardware processor to obtain a plurality of thermal stability indications.

[0030] Some embodiments further comprise identifying a subset of the panel of scFvs for subsequent production based on the plurality of thermal stability indicators.

[0031] Some embodiments provide a method for computationally screening a set of monoclonal antibodies (mAbs) based on the thermal stability of mAbs predicted by a trained machine learning model, the set of mAbs including mAbs with different residue sequences, the method comprising: determining a thermal stability indication for each mAb in the set of mAbs using the trained machine learning model and at least one computer hardware processor to obtain a plurality of thermal stability indications, the set of mAbs including a first mAb having a first residue sequence, the determining comprising: obtaining interaction energy metrics for each of a plurality of pairs of residues, the residues being in the first residue sequence, using information indicative of a three-dimensional (3D) structure of the first mAb; generating a first feature set to provide as input to the trained machine learning model, the generating comprising including the interaction energy metrics in the first feature set; and providing the first feature set as input to the trained machine learning model to obtain a corresponding output indicative of a first thermal stability of the first mAb; identifying a subset of the set of mAbs for subsequent production based on the plurality of thermal stability indications; and producing at least one mAb among the mAbs in the identified subset.

[0032] Some embodiments further comprise testing the thermal stability of at least one of the mAbs in an in vitro assay.

[0033] Some embodiments provide a method for computationally screening a panel of antibodies based on the thermal stability of the antibodies predicted by a trained machine learning model, the panel of antibodies including antibodies with different residue sequences, the method comprising: determining a thermal stability indication for each antibody in the panel of antibodies using the trained machine learning model and at least one computer hardware processor to obtain a plurality of thermal stability indications, the panel of antibodies including a first antibody having a first residue sequence, the determining comprising: obtaining interaction energy metrics for each of a plurality of pairs of residues, the residues being in the first residue sequence, using information indicative of a three-dimensional (3D) structure of the first antibody; generating a first feature set to provide as input to the trained machine learning model, the generating comprising including the interaction energy metrics in the first feature set; and providing the first feature set as input to the trained machine learning model to obtain a corresponding output indicative of a first thermal stability of the first antibody; identifying a subset of the panel of antibodies for subsequent production based on the plurality of thermal stability indications; and producing at least one of the antibodies in the identified subset.

[0034] Some embodiments further comprise testing the thermal stability of at least one of the antibodies in an in vitro assay. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The drawings are not intended to be drawn to scale. In the drawings, each identical or nearly identical component shown in different figures is represented by the same numeral. For clarity, not every component may be labeled in every figure. In the drawings:

[0036] Figure 1A is a diagram depicting an illustrative technique 100 for determining the thermal stability of a single-chain variable fragment (scFv) according to some embodiments of the technology described herein.

[0037] Figure 1B is a block diagram of an example system 150 for determining thermal stability of a scFv according to some embodiments of the technology described herein.

[0038] Figure 2A is a flow chart of an illustrative process 200 for computationally screening a panel of scFvs to identify a subset of the panel of scFvs for production in accordance with some embodiments of the technology described herein.

[0039] Figure 2B is a flow chart of an illustrative process 250 for determining the thermal stability of a scFv according to some embodiments of the technology described herein.

[0040] Figure 3A is a diagram of an illustrative technique for computationally screening a set of scFvs using a trained machine learning model that is trained to generate an indication of the thermal stability of the scFv based on an input comprising interaction energy metrics for residue pairs in the scFv according to some embodiments of the technology described herein.

[0041] Figure 3B is a diagram of an illustrative technique for computationally screening a set of scFvs using a trained machine learning model according to some embodiments of the technology described herein, the trained machine learning model being trained to generate an indication of the thermal stability of the scFv based on inputs comprising interaction energy metrics for residue pairs in the scFv and features representing the sequence of the scFv.

[0042] Figure 3C is a diagram of an illustrative technique for computationally screening a set of scFvs using a trained machine learning model that is trained to generate an indication of the thermal stability of an scFv based on an input comprising features representing a sequence of the scFv, according to some embodiments of the technology described herein.

[0043] Figure 4A is a flow chart of an illustrative process 400 for training a machine learning model to predict thermal stability of a scFv according to some embodiments of the technology described herein.

[0044] Figure 4B is an illustration of an example technique 450 for experimentally generating data for training a model to predict thermal stability of scFvs according to some embodiments of the technology described herein.

[0045] Figure 5A is a flow chart of an illustrative process 500 for computationally screening a panel of antibodies to identify a subset of the panel for production in accordance with some embodiments of the technology described herein.

[0046] Figure 5B is a flow chart of an illustrative process 550 for determining thermal stability of an antibody according to some embodiments of the technology described herein.

[0047] Figure 6 is an illustration of an example of a scFv, a feature set generated for the scFv, and a machine learning model trained to predict the thermal stability of the scFv according to some embodiments of the technology described herein.

[0048] Figure 7 is an illustration of an example technique for training and using a trained machine learning model to predict thermal stability of a scFv according to some embodiments of the technology described herein.

[0049] Figure 8 is an illustration of an example technique 800 for predicting thermal stability of a scFv using a trained convolutional neural network (CNN) model according to some embodiments of the technology described herein.

[0050] Fig. 9 t-Stochastic Nearest Neighbor Embedding (t-SNE) according to some embodiments of the technology described herein are shown, which show the improved performance of the technology described herein for predicting the thermal stability of scFvs using a machine learning model trained using the interaction energy metric of scFvs.

[0051] Fig.10 Receiver operating characteristic (ROC) curves are shown showing the performance of a machine learning model trained to predict the thermal stability of scFvs using interaction energy metrics for residue pairs of scFv residue sequences according to some embodiments of the technology described herein.

[0052] Fig.11 Shown is a chart comparing the performance of various machine learning models described herein for predicting thermal stability of scFvs according to some embodiments of the technology described herein.

[0053] Fig. 12A Contains graphs showing thermal stability prediction performance using various neural network architectures according to some embodiments of the techniques described herein.

[0054] Fig. 12B is a table showing that prediction of thermal stability using an ensemble of neural network models is more accurate than prediction of thermal stability using a single neural network model according to some embodiments of the technology described herein.

[0055] Fig. 12C t-SNEs according to some embodiments of the technology described herein are shown, which show the improved performance of the technology described herein for predicting the thermal stability of scFv using an ensemble of machine learning models.

[0056] FIG. 13A to FIG. 13B are graphs showing that thermal stability prediction using machine learning techniques can be used to observe thermal stability trends of scFvs according to some embodiments of the technology described herein.

[0057] Fig.14 is a diagram depicting an example technique for predicting thermal stability of scFvs using a trained language model according to some embodiments of the technology described herein.

[0058] Fig.15Ais an illustration of an example pre-trained language model configured for zero-shot learning thermal stability prediction using sequence data according to some embodiments of the techniques described herein.

[0059] Fig. 15B is an illustration of an example fine-tuned pre-trained language model configured for thermal stability prediction using sequence data according to some embodiments of the techniques described herein.

[0060] FIG. 15C to FIG. 15F are graphs showing that fine-tuning predictions achieve improved correlation with thermal stability compared to zero-shot learning predictions according to some embodiments described herein.

[0061] Fig.16 is an illustration according to some embodiments of the technology described herein showing that thermal stability predictions determined using machine learning techniques are consistent with experimentally determined thermal stability.

[0062] FIG. 17A to FIG. 17B Some embodiments according to the technology described herein demonstrate that thermostability prediction using machine learning techniques can be used to identify thermostabilizing mutations compared to traditional germlining techniques.

[0063] FIG. 18A to FIG. 18B are graphs showing that thermal stability predictions output by a supervised convolutional neural network achieve improved correlation with thermal stability compared to thermal stability predictions made by an unsupervised pre-trained language model according to some embodiments of the technology described herein.

[0064] Fig.19 Contains a chart comparing the performance of various machine learning models described herein for predicting thermal stability of scFvs according to some embodiments of the technology described herein.

[0065] Fig. 20 Contains graphs showing that training and prediction based on residue pair interaction energy metrics are more accurate than training and prediction based on coding sequences and training and prediction based on both coding sequences and residue pair interaction energy metrics according to some embodiments of the technology described herein.

[0066] FIG. 21A to FIG. 21B are graphs showing that training and prediction based on residue pair interaction energy metrics are more accurate than training and prediction based on residue pair interaction energy metrics and coding sequences according to some embodiments of the technology described herein.

[0067] Fig. 22 Distributions of training, validation, and test datasets are shown according to some embodiments of the techniques described herein.

[0068] Fig.23 is a graph showing temperature distribution of TS50 measurements for an experimental data set according to some embodiments of the techniques described herein.

[0069] Fig.24 is a schematic diagram of an illustrative computing device with which aspects described herein may be implemented. DETAILED DESCRIPTION

[0070] The inventors have developed a machine learning technique for predicting the thermal stability of a single-chain variable fragment (scFv) or other multi-specific constructs. In some embodiments, predicting the thermal stability of an scFv includes using a trained machine learning model to process a feature set generated for the scFv. In some embodiments, the feature set is generated using the residue sequence of the scFv. For example, in some embodiments, the feature set includes interaction energy metrics for residue pairs in the residue sequence. Additionally or alternatively, the feature set includes the residue sequence of the scFv. In some embodiments, the feature set is provided as input to a trained machine learning model to obtain an output (e.g., a thermal stability indication) indicating the temperature at which the scFv is thermally stable.

[0071] In some embodiments, the technology described herein is used to screen a group of scFvs. For example, the group of scFvs may include one or more scFvs as candidates for subsequent production. In some embodiments, the technology for screening the group of scFvs includes determining the thermal stability indication of each scFv included in the group of scFvs, and identifying a subgroup of the group of scFvs based on the determined thermal stability indication. For example, the identified subgroup of scFvs may include scFvs that are thermally stable. Therefore, one or more of the scFvs included in the identified subgroup of scFvs may be produced (e.g., manufactured) subsequently.

[0072] As mentioned above, multi-specific biologics are increasingly important for obtaining novel therapeutically relevant pathways and mechanisms of action. One property used to measure the potential developability of multi-specific biologics, scFv modules or multi-specificity containing scFv is thermal stability. Thermal stability is a characteristic that can indicate its stability under certain environmental conditions. Various factors (such as the specific amino acid sequence and / or the resulting structure of scFv) can affect the thermal stability of scFv. For example, interactions between amino acid residues (such as hydrophobic and electrostatic interactions) can affect the thermal stability of scFv. Additionally or alternatively, the presence of specific bonds (such as disulfide bonds) can affect the thermal stability of scFv.

[0073] Thermal stability is relevant because a thermostable scFv can withstand certain environmental conditions and have corresponding adaptations to maintain scFv function under these conditions. For example, such a thermostable scFv can withstand exposure to temperatures in the temperature range that may occur in practice. In contrast, a scFv with poor thermostability properties may unfold and denature in the temperature range that may occur in practice, resulting in a loss of enzyme activity.

[0074] The traditional method of optimizing the thermal stability of scFv involves experimental screening of candidate scFvs for thermal stability. However, experimental screening is resource-intensive, time-consuming and expensive because it requires production and experimentation of each scFv in a large group of candidate scFvs to be screened to determine which scFvs have the desired thermal stability properties.

[0075] The inventors have realized that computational screening of the thermal stability of scFv will be particularly useful. However, the inventors recognize that traditional computational methods for predicting the thermal stability of scFv are unreliable and inaccurate. For example, some traditional computational techniques include processing amino acid sequences using machine learning models trained on other amino acid sequences and their corresponding thermal stabilities. The machine learning model is trained to predict the thermal stability of an amino acid sequence by identifying similar amino acid sequences used to train the model. However, although two amino acid sequences may have high similarity, this does not mean that they will have similar thermal stabilities. For example, two amino acids may be identical, but there is a mutation. However, this mutation may result in a significant change in thermal stability. Therefore, the predictions generated by such machine learning models may be inaccurate.

[0076] Other traditional computational techniques have been used to predict thermal stability based on the total energy of a protein or fragment. However, such techniques still lead to inaccurate predictions because they do not take into account the structure of the scFv, which strongly affects its thermal stability. Without taking into account the structure of the scFv, traditional techniques ignore some information that could drastically change the resulting thermal stability predictions. Therefore, these techniques are also unreliable and inaccurate.

[0077] The inventors have recognized the need for accurate computational methods to predict scFv thermal stability based on the primary amino acid sequence of a candidate scFv, and have realized that such methods will guide thermal stability engineering work and will be very valuable for multi-specific drug development. In particular, the inventors have recognized that, compared with conventional techniques, considering the structure of the scFv (e.g., 2D and / or 3D structure) leads to more accurate thermal stability predictions. Specifically, the inventors have developed a technique that considers per-residue interactions when predicting thermal stability. These interactions capture information about the scFv structure and allow prediction techniques (e.g., machine learning techniques) to consider the structure.

[0078] Therefore, the inventors have developed machine learning techniques for predicting the thermal stability of single-chain variable fragments (scFv). In order to determine the thermal stability of a particular scFv, these techniques involve obtaining features from the primary amino acid sequence of the particular scFv, and providing these features as input to a trained machine learning model to generate a corresponding output indicating the thermal stability of the particular scFv ("thermal stability indication"). The output can be a measure of thermal stability, such as a temperature at which the scFv is thermally stable (e.g., a temperature corresponding to the half-maximal binding of the scFv) or a temperature range including such a temperature. In some embodiments, the features provided as input to the trained machine learning model include only energy features, for example, interaction energy metrics between residue pairs in the residue sequence of the particular scFv. Additionally or alternatively, in some embodiments, features may include encodings of residue sequences.

[0079] An example application of machine learning techniques for predicting the thermal stability of scFvs is to computationally screen scFvs prior to production to identify those scFvs with favorable thermal stability properties. Thus, machine learning techniques can be used to determine the thermal stability indicator for each scFv included in a set of scFvs that are computationally screened, and to identify a subset of the set of scFvs based on the determined thermal stability indicator. At least some of the scFvs so identified can subsequently be produced.

[0080] Accordingly, some embodiments provide a method for computationally screening a set of single-chain variable fragments (scFvs) based on the thermal stability of the single-chain variable fragments (scFvs) predicted by a trained machine learning model, the set of scFvs comprising scFvs having different residue sequences, the method comprising: (A) determining a thermal stability indication for each scFv in the set of scFvs using the trained machine learning model and at least one computer hardware processor to obtain a plurality of thermal stability indications, the set of scFvs comprising a first scFv having a first residue sequence, the determining comprising: (i) obtaining an interaction energy metric for each of a plurality of pairs of residues using information indicative of a three-dimensional (3D) structure of the first scFv, the residues being in the first residue sequence; (ii) generating a first feature set to provide as input to the trained machine learning model, the generating comprising including the interaction energy metrics in the first feature set; and (iii) providing the first feature set as input to the trained machine learning model to obtain a corresponding output indicative of a first thermal stability of the first scFv; (B) identifying a subset of the set of scFvs for subsequent production based on the plurality of thermal stability indications; and (C) producing at least one scFv among the scFvs in the identified subset.

[0081] In some embodiments, the output indicating a first thermal stability of a first scFv indicates a first temperature at which the first scFv is thermally stable. The first temperature can be an estimate of the temperature corresponding to half-maximal binding of the first scFv (this can be referred to as the "TS50" temperature). In this manner, the machine learning model can be configured to operate as a regression model.

[0082] In some embodiments, the output indicating the first thermal stability of the first scFv indicates a first temperature range, the first temperature range including at least one temperature when the first scFv is thermally stable. The first temperature range can be an estimate of the temperature range including the temperature corresponding to the half-maximal binding of the first scFv. In some embodiments, providing the first feature set as input to the trained machine learning model to obtain an output indicating the first thermal stability of the first scFv includes: using the trained machine learning model, using the first feature set to classify the first scFv into one of a plurality of categories, wherein each of the plurality of categories corresponds to a corresponding temperature range. In this way, the machine learning model can be configured to operate as a classification model.

[0083] In some embodiments, residue pair interaction energy metrics can be obtained in a two-stage process in which: (1) the residue sequence of the scFv is used to determine information indicative of the 3D structure of the first scFv (e.g., using protein structure prediction software, examples of which are provided herein); and (2) the interaction energy metrics are determined using information indicative of the 3D structure of the first scFv (e.g., by using molecular modeling software, examples of which are provided herein).

[0084] The inventors have recognized that the manner in which interaction energy metrics are provided as input to a trained neural network model may affect the performance of the model in predicting thermal stability. In particular, the inventors have recognized that organizing the interaction energy metrics in a two-dimensional array or matrix generates a matrix with local spatial structure suitable for analysis by a convolutional neural network model, and that organizing the interaction energy metrics in this manner may produce improved performance in some embodiments (e.g., compared to providing a linear sequence of interaction energy metrics).

[0085] Thus, in some embodiments, generating the first feature set includes generating, for each specific energy metric, a corresponding two-dimensional matrix of values ​​of the specific energy metric, wherein the rows and columns of the 2D matrix correspond to corresponding residues in the first residue sequence. Thus, the entry in the i-th row and j-th column of the 2D matrix corresponds to the value of the specific energy metric between the i-th residue in the first residue sequence and the j-th residue in the first residue sequence.

[0086] Although in some embodiments, the interaction energy metrics between all residue pairs of the scFv are provided as input to the trained machine learning model, in other embodiments, only the interaction energy metrics between some residue pairs are provided. For example, the interaction energy metrics between at least 50%, at least 60%, at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, or at least 99% of the residue pairs can be provided as input to the trained machine learning model. In some such embodiments, the 2D matrix of interaction energy metrics includes rows for at least 50%, 60%, 75%, 85%, 90%, 95%, and 99% of the residues in the residue sequence of the first scFv.

[0087] In some embodiments, sequence data can be provided as an additional input (in addition to the interaction energy metric) to the trained machine learning model. Thus, in some such embodiments, generating the first feature set further comprises encoding (e.g., one-hot encoding) the first residue sequence to obtain an encoding sequence; and including the encoding sequence in the first feature set.

[0088] In some embodiments, the trained machine learning model includes a trained neural network model, such as a convolutional neural network (CNN) model with multiple 2D convolutional layers. The CNN model can have a fully connected layer. Other types of architectures are also possible (e.g., with one or more dropout layers, one or more pooling layers, etc.).

[0089] In some embodiments, the trained CNN model is configured to output a plurality of probabilities that the scFv is thermally stable within each of a plurality of temperature ranges. These probabilities can be used to identify a single most likely temperature range (e.g., the temperature range associated with the maximum probability value) or to determine an estimate of a particular temperature value, e.g., as a weighted linear combination of averages weighted by their respective probabilities.

[0090] As used herein, the term "single-chain Fv" or "scFv" refers to a single polypeptide chain antibody fragment comprising variable regions from heavy and light chains, but lacking a constant region. Generally, a single-chain Fv further comprises a peptide linker between the VH and VL domains that allows it to form a desired structure that will allow antigen binding. Such constructs are discussed in detail in Pluckthun, The Pharmacology of Monoclonal Antibodies, Vol. 113, Rosenburg and Moore, eds., Springer-Verlag, New York, pp. 269-315 (1994) and U.S. Pat. No. 7,112,324 entitled "CD 19xCD3 SPECIFIC POLYPEPTIDES AND USES THEREOF", each of which is incorporated herein by reference in its entirety. Structures that can be formed from scFv include structures in the bispecific T cell engager (BiTE) format described in U.S. Pat. No. 7,112,324, entitled "CD 19xCD3 SPECIFIC POLYPEPTIDES AND USES THEREOF" (which is incorporated herein by reference in its entirety), which is a fusion protein consisting of two single-chain variable fragments (scFv) connected by a peptide linker, starting from the N-terminus with a variable heavy chain (VH) domain followed by a variable light chain (VL) domain. The term "peptide linker" refers to an amino acid sequence by which the amino acid sequences of one (variable and / or binding) domain and another (variable and / or binding) domain of the scFv are connected to each other. Suitable peptide linkers include those described in U.S. Pat. No. 4,751,180, entitled "EXPRESSION USING FUSED GENES PROVIDING FOR PROTEIN PRODUCT," and U.S. Pat. No. 4,935,233, entitled "COVALENTLY LINKED POLYPEPTIDE CELL MODULATORS," or WO 88 / 09344, entitled "TARGETED MULTIFUNCTIONAL PROTEINS," each of which is incorporated herein by reference in its entirety.

[0091] It should be understood that the techniques described herein can be applied to predict the thermal stability of scFvs, as well as to any antibody construct having variable domains (VH and VL domains), including constructs with or without peptide linkers. In addition, the techniques described herein can be applied to multispecific constructs by utilizing training data sets of multispecific constructs having the same (or very similar, e.g., at least 80%, 85%, 90%, 95%, 99% similarity) sequence and / or structure. In addition, the techniques described herein can be applied to any type of protein by utilizing training data sets of proteins having the same (or very similar, e.g., at least 80%, 85%, 90%, 95%, 99% similarity) sequence and / or structure.

[0092] The term "multispecific construct" refers to a molecule in which the structure and / or function is based on that of an antibody (e.g., a full-length or complete immunoglobulin molecule) and / or is extracted from the variable heavy chain (VH) and / or variable light chain (VL) domains of an antibody or fragment thereof. The definition of the term "multispecific construct" includes monovalent, bivalent and multivalent (polyvalent / multivalent) constructs, and thus includes bispecific constructs that specifically bind to only two antigen structures, as well as multispecific constructs that specifically bind to more than two (e.g., three, four or more) antigen structures through different binding domains. In addition, the definition of the term "multispecific construct" includes molecules consisting of only one polypeptide chain (e.g., scFv) as well as molecules consisting of more than one polypeptide chain, which chains may be identical (homodimers, homotrimers or homooligomers) or different (heterodimers, heterotrimers or heterooligomers). Examples of the above-identified molecules and variants or derivatives thereof are described in, inter alia: Harlow and Lane, Antibodies a laboratory manual, CSHL Press (1988) and Using Antibodies: a laboratory manual, CSHL Press (1999), Kontermann and Dübel, Antibody Engineering, Springer, 2nd edition, 2010, and Little, Recombinant Antibodies for Immunotherapy, Cambridge University Press 2009, each of which is incorporated herein by reference in its entirety. Additional examples of multispecific construct formats include Fab fragments, a monovalent fragment having VL, VH, CL and CH1 domains; F(ab') 2Fragments include, a bivalent fragment having two Fab fragments connected by a disulfide bridge at the hinge region; an Fd fragment having two VH and CH1 domains; an Fv fragment having the VL and VH domains of a single arm of an antibody; a dAb fragment (Ward et al., (1989) Nature 341:544-546, which is incorporated herein by reference in its entirety), which has a VH domain; isolated complementarity determining regions (CDRs); and a single-chain Fv (scFv). Examples of multispecific constructs are described, for example, in WO 00 / 006605 entitled “HETEROMINIBODIES,” WO 2005 / 040220 entitled “MULTISPECIFIC DEIMMUNIZED CD3-BINDERS,” WO 2008 / 119567 entitled “CROSS-SPECIES-SPECIFIC CD3-EPSILON BINDING DOMAIN,” WO 2010 / 037838 entitled “CROSS-SPECIES-SPECIFIC SINGLE DOMAIN BISPECIFIC SINGLE CHAIN ​​ANTIBODY,” WO 2011 / 036964 entitled “BISPECIFIC T CELL ACTIVATING ANTIGEN BINDING MOLECULES,” WO 2013 / 061698 entitled “Cross-Species-SPECIFIC CD3-EPSILON BINDING DOMAIN,” WO 2014 / 061698 entitled “Cross-Species-SPECIFIC SINGLE DOMAIN BISPECIFIC SINGLE CHAIN ​​ANTIBODY,” WO 2015 / 061698 entitled “Cross-Species-SPECIFIC SINGLE DOMAIN BISPECIFIC SINGLE CHAIN ​​ANTIBODY,” WO 2016 / 061698 entitled “Cross-Species-SPECIFIC 2013 / 026837, WO 2013 / 026833 entitled “BISPECIFIC T CELL ACTIVATING ANTIGENBINDING MOLECULES”, US 2014 / 0308285 entitled “HETERODIMERIC BISPECIFIC ANTIBODIES”, US 2014 / 0302037 entitled “BISPECIFIC-FC MOLECULES”, WO 2014 / 144722 entitled “BISPECIFIC FC MOLECULES”, WO 2014 / 151910 entitled “HETERODIMERIC BISPECIFIC ANTIBODIES” and VC-FC-VC ANTIBODY [VC-FC-VC ANTIBODY]”, each of which is incorporated herein by reference in its entirety.Furthermore, the multispecific construct may be a fragment of a full-length antibody, such as VH, VHH, VL, (s)dAb, Fv, Fd, Fab, Fab', F(ab')2 or "rIgG" ("half antibody").

[0093] In addition, it will be appreciated that the techniques described herein can be applied to constructs comprising modified fragments of antibodies, such as scFv, di-scFv or bi(s)-scFv, scFv-Fc, scFv-zipper, scFab, Fab 2 ,Fab 3 , diabodies, single-chain diabodies, tandem diabodies (Tandab's), tandem di-scFv, tandem tri-scFv, "multibodies" such as triabodies or tetrabodies, single domain antibodies such as nanobodies, or single variable domain antibodies comprising only one variable domain (which may be VHH, VH or VL) (which specifically binds an antigen or epitope independently of other V regions or domains), and human heavy chain antibodies As described in WO 2020206330A1, entitled “HEAVY CHAIN ​​ANTIBODIES BINDING TO PSMA,” which is incorporated by reference in its entirety.

[0094] The term "binding domain" refers to a domain that (specifically) binds / interacts / recognizes a given target epitope or a given target side end on a target molecule (antigen) (e.g., CD33 and CD3, respectively). The structure and function of the first binding domain (recognizing, e.g., CD33) and preferably also the structure and / or function of the second binding domain (recognizing CD3) are based on the structure and / or function of an antibody, e.g., a full-length or complete immunoglobulin molecule, and / or are extracted from the variable heavy chain (VH) and / or variable light chain (VL) domains of an antibody or fragment thereof. Preferably, the binding domain is characterized by the presence of three heavy chain CDRs (i.e., CDR1, CDR2 and CDR3 in the VH region), followed by three light chain CDRs (i.e., CDR1, CDR2 and CDR3 in the VL region). The binding domains of the multispecific construct and the binding domains used in the training set should remain consistent.

[0095] Therefore, as should be understood from the foregoing, the technology described herein is not limited to being applied to single-chain variable fragments, and can also be applied to other constructs described herein. For example, the technology described herein can be applied to monoclonal antibodies (mAbs). For example, a mAb sequence can be input as a VH sequence, followed by a VL sequence (without an intervening joint sequence). As another example, a mAb sequence can be input as a VL sequence, followed by a VH sequence (without an intervening joint sequence). The input sequence can be used to predict the structure, and the structure can be used to calculate the energy metric. Then, the energy metric (and optionally, the encoding of the input sequence) can be provided as an input to a trained machine learning model to obtain a thermal stability indication of mAb. As another example, the technology described herein can be applied to one or more types of multispecific constructs, and examples of these constructs are provided herein.

[0096] In another example, the technology described herein can be applied to any type of antibody. For example, an antibody sequence can be used to predict a structure, and the structure can be used to calculate an energy metric. Then, the energy metric (and optionally, the encoding of the input sequence) can be provided as an input to a trained machine learning model to obtain an indication of the thermal stability of the antibody. In yet another example, the technology described herein can be applied to any type of protein. For example, a protein sequence can be used to predict a structure, and the structure can be used to calculate an energy metric. Then, the energy metric (and optionally, the encoding of the input sequence) can be provided as an input to a trained machine learning model to obtain an indication of the thermal stability of the protein.

[0097] The techniques described herein can be implemented in any of many ways, as the techniques are not limited to any particular implementation. Examples of the details of implementation are provided herein for illustrative purposes only. In addition, the techniques disclosed herein can be used alone or in any suitable combination, as the various aspects of the techniques described herein are not limited to use of any particular technique or combination.

[0098] Figure 1A is a diagram of an illustrative technique 100 for determining an indication 110 of thermal stability of a single-chain variable fragment (scFv) by providing features 106 generated from a residue sequence 102 of the scFv as input to a machine learning model 108 .

[0099] In certain embodiments, scFv sequence 102 specifies the residue sequence (e.g., primary amino acid sequence) of scFv. scFv sequence can specify the amino acid sequence of heavy chain and light chain and joint peptide of scFv. scFv sequence 102 can have any suitable length. For example, scFv sequence 102 can have 200 to 300 residues, 225 to 275 residues or 236 to 254 residues, or any other suitable range within these ranges. In certain embodiments, joint peptide can be composed of 10 to 25 amino acids, and scFv sequence 102 can include a subsequence of this length of amino acid in the joint representing scFv.

[0100] In some embodiments, the scFv sequence 102 can be specified by a user. For example, a user can communicate with a computing device (e.g., Figure 1B ) to specify the amino acid residue sequence of the scFv sequence 102. In some embodiments, the scFv sequence 102 can be automatically specified by the computing device (e.g., (multiple) computing devices 120). For example, the computing device can be programmed to generate the amino acid sequence of the scFv sequence 102 (e.g., using machine learning, by changing the amino acids iteratively, randomly, etc. in a predetermined order). This may be useful when the computer is programmed to automatically generate a list of scFvs for computational screening, for example, by automatically generating mutations by replacing one or more amino acids in the starting scFv sequence to obtain a candidate scFv sequence, which can be evaluated during screening (e.g., based on its predicted thermal stability). In some embodiments, the scFv sequence 102 can be generated at least partially automatically (e.g., by programming to change one or more amino acids in the sequence) and at least partially manually (e.g., based on user input of one or more amino acids in the specified sequence). The scFv sequence 102 can be specified in any suitable form (e.g., FASTA), because the various aspects of the technology described herein are not limited in this regard.

[0101] Next, as part of the technique 100, scFv data 104 is generated from the scFv sequence 102. The scFv data 104 may include any type of data generated from the scFv sequence 102. Figure 1A In the illustrated embodiment, the scFv data 104 includes sequence data 104a, structure data 104b, and energy data 104c, each of which will be described below. However, it should be understood that this example is illustrative, and in other embodiments, in addition to Figure 1A In addition to or in lieu of the types of data shown in , the scFv data 104 may include any other suitable data generated from the scFv sequence 102.

[0102] In some embodiments, the sequence data 104a includes the scFv sequence 102 itself or a subsequence thereof. In addition, the sequence data 104a may include information about the sequence (e.g., amino acid statistics, length, sequence identifier, and / or any other information associated with the sequence). In some embodiments, the sequence data 104a is stored in a text-based file (e.g., a FASTA file) and / or in any other suitable format, as the aspects of the technology described herein are not limited in this regard.

[0103] In some embodiments, the structural data 104b includes information indicating the three-dimensional structure of the scFv sequence 102. The information indicating the 3D structure may include a description and / or annotation of the protein structure, including atomic coordinates, secondary structure assignments, and / or atomic connectivity data. The structural data 104b may be in any suitable format describing the 3D structure of the scFv (e.g., a protein data bank (PDB) file format, a crystallographic information file (CIF) format, a macromolecular crystallographic information file (mmCIF) format).

[0104] In some embodiments, sequence data 104a is used to obtain structural data 104b. For example, obtaining structural data 104b may include using protein structure prediction software (e.g., Figure 1B The protein structure prediction module 160 shown in ) processes the sequence data 104a to predict the 3D structure of the scFv. Figure 2B Act 252 of process 250 shown in describes a technique for obtaining information indicative of the 3D structure of a scFv.

[0105] In some embodiments, energy data 104c includes information indicating the energy level of scFv residue sequence 102 in its 3D conformation. For example, energy data may include interaction energy metrics of one or more pairs of residues in scFv residue sequence 102. Interaction energy metrics may include the energy of several different types of interactions between a pair of residues. For example, interaction energy metrics may explain the interaction energy between non-bonded atomic pairs and the statistical potential for describing the torsional preference of the backbone and side chains in scFv. In some embodiments, energy data 104c is included in, for example, a delimited text file, such as a comma separated value (CSV) file.

[0106] In some embodiments, the structural data 104b is used to obtain the energy data 104c. For example, obtaining the energy data 104c may include using molecular modeling software (e.g., using Figure 1B The molecular modeling module 162 shown in ) processes the structural data 104b to predict the interaction energy metric for each of the plurality of pairs (e.g., some or all) of residues in the scFv sequence 102. Figure 2B Act 254 of process 250 shown in describes a technique for obtaining an interaction energy metric.

[0107] After obtaining the scFv data 104, the scFv data 104 is used to generate a feature set 106 for the scFv. Figure 1A In some embodiments, the feature set 106 includes an interaction energy metric 106a (e.g., included in the energy data 104c). In some embodiments, the interaction energy metric 106a includes the interaction energy metric of each of the multiple pairs of residues of the scFv sequence 102. The interaction energy metric 106a may include the interaction energy metric of at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 98%, at least 99% or all of the residue pairs of the scFv sequence 102. In some other embodiments, the interaction energy metric 106a may include the interaction energy metric of at most 99%, at most 98%, at most 95%, at most 90%, at most 85%, at most 80%, at most 75% or less of the residue pairs of the scFv sequence 102. Therefore, it should be understood that in some embodiments, the interaction energy metric 106a may include the interaction energy metric of a subset (i.e., not all) of the residue pairs of the scFv sequence 102.

[0108] like Figure 1A As shown, feature set 106 optionally includes a coding sequence 106b of scFv. Coding sequence 106b can be the coding of scFv sequence 102 (at least a subsequence or all thereof). As an example, coding sequence 106b can be obtained by the unique hot encoding of scFv sequence 102. Unique hot encoding is a technique for converting categorical data (each residue in a sequence is one of twenty possible amino acids) into numerical data (e.g., binary data, integer valued data, real valued data). scFv sequence 102 can be uniquely hot encoded by converting into a series of 20-dimensional vectors (one vector for each amino acid in the sequence). Each coordinate of the vector can correspond to a corresponding one of the 20 amino acids. Each amino acid can then be encoded as a single 20-dimensional vector having 1 at the coordinate of the amino acid and 0 elsewhere. However, it should be understood that aspects of the technology described herein are not limited to encoding scFv sequences using unique hot encoding, and other methods for encoding categorical data can be used.

[0109] In another embodiment, the feature set 106 may include only the encoded sequence data and does not include any interaction energy metrics. In fact, the machine learning techniques described herein can be used to predict the thermal stability of scFv using energy features alone (e.g., interaction energy metrics), sequence features alone (e.g., using the unique heat encoding of the scFv sequence), or using a combination of energy features and sequence features. As described herein, depending on which features are utilized, different machine learning models can be used to perform predictions (e.g., when the input includes an interaction energy metric, a 2D convolutional neural network is used, and when the input does not include an interaction energy metric, a language model is used).

[0110] In addition to sequence features and energy features, it should be understood that feature set 106 can include one or more additional or alternative features, as aspects of the technology described herein are not limited in this regard. For example, one or more additional or alternative features can be obtained from scFv data 104 and included in feature set 106.

[0111] After generating the feature 106, it is provided as input to a trained machine learning model 108 to obtain a corresponding thermal stability indication 110 of the scFv. The machine learning model can be of any suitable type. For example, the machine learning model can be a neural network model, such as a convolutional neural network (CNN) model. The CNN model can have one or more convolutional layers (e.g., one or more two-dimensional convolutional layers). The CNN model can have a fully connected layer. As described herein, in some embodiments, the convolutional neural network model can be configured to receive only energy features (e.g., as a 1D or 2D matrix interaction energy metric) or a combination of energy features and sequence features (e.g., a unique hot encoding of the scFv sequence) as input.

[0112] As another example, the machine learning model can be a pre-trained language model adapted to a protein setting. For example, a bidirectional encoder representation (BERT) model from transformers (such as an ESM-1b or ESM-1v language model) can be used. These types of models are described in: A Rives et al., Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences [Extending unsupervised learning to 250 million protein sequences, biological structure and function emerge] Proc. Natl. Acad. Sci. [Proceedings of the National Academy of Sciences of the United States] 118 (2021), and J Meier et al., Language models enable zero-shot prediction of the effects of mutations on protein function [Language models enable zero-shot learning to predict the effects of mutations on protein function] Adv. Neural Information Processing Systems [Advances in neural information processing systems] 35 (2021), each of which is incorporated herein by reference in its entirety. As another example, the UniRep language model can be used, which is described in Ma, Eric J. and Arkadij Kummer, “Reimplementing Unirepin JAX” bioRxiv (2020), which is incorporated herein by reference in its entirety.

[0113] In some embodiments, the machine learning model can be formed as an integration of multiple machine learning models. For example, the machine learning model can include an integration of neural networks. In some embodiments, implementing the integration of machine learning models includes training each of multiple (e.g., two or more, three or more, etc.) machine learning models on different training data sets, using each of the trained machine learning models to predict thermal stability, and averaging their predictions. In some embodiments, boosting can be used to combine multiple machine learning models.

[0114] Regardless of the specific type of machine learning model 108 used as part of the technique 100, the machine learning model outputs an indication of thermal stability indication 110. In some embodiments, the indication can be a temperature or temperature range at which the scFv is thermally stable. For example, the indication can be a temperature corresponding to half-maximal binding of the scFv (which can be referred to herein as the "TS50" temperature) or a temperature range that includes the TS50 temperature. As another example, the indication can be a thermal melting temperature (which can be referred to herein as the "T m ” temperature) or including T m Temperature range of temperature. m The temperature may correspond to a temperature at which the concentration of the scFv in its folded state is equal to the concentration of the scFv in its unfolded state.

[0115] In some embodiments, the machine learning model 108 is configured to output the probability that the scFv is thermally stable for each of a plurality of temperature ranges (e.g., below 50°C, 50°C to 60°C, 60°C to 70°C, and above 70°C). These ranges can be closed (e.g., 50°C to 60°C) or open (e.g., below 50°C or above 70°C). As a non-limiting example, the machine learning model 108 can output a first probability that the scFv is thermally stable within a first temperature range and a second probability that the scFv is thermally stable within a second temperature range.

[0116] Figure 1A The techniques shown in can be used to computationally screen a group of scFvs to identify a subset of scFvs for production. For example, the group of scFvs can be screened based on the thermal stability indication generated for the scFvs in the group of scFvs by the technique 100. As an example, when the thermal stability indication is an indication of temperature (e.g., TS50 temperature, Tm temperature), scFvs with predicted temperatures exceeding a specified threshold (e.g., greater than 50°C, greater than 55°C, greater than 60°C) can be selected for subsequent production.

[0117] In some embodiments, the thermal stability indicator 110 is used to identify scFvs for subsequent production. For example, a scFv with a predicted thermal stability indicator that meets one or more criteria can be identified for subsequent production. Figure 2A Techniques for screening scFvs for production are described.

[0118] Figure 1B1 is a block diagram of an example system 150 for predicting the thermal stability of scFvs and computationally screening scFvs based on such predictions according to some embodiments of the technology described herein. The system 150 includes a computing device(s) 120 configured to cause software 130 to execute thereon to perform various functions related to predicting the thermal stability of scFvs and computationally screening scFvs based on such predictions.

[0119] exist Figure 1B In the illustrated embodiment, computing device(s) 120 include software 130 configured to perform various functions with respect to scFv data (eg, scFv data 104).

[0120] The computing device(s) 120 may be one or more computing devices of any suitable type. For example, the computing device(s) 120 may be a portable computing device (e.g., a laptop, a smart phone) or a fixed computing device (e.g., a desktop computer, a server). When the computing device(s) 120 include multiple computing devices, the devices(s) may be physically co-located (e.g., in a single room) or distributed across multiple physical locations. In some embodiments, the computing device(s) 120 may be part of a cloud computing infrastructure.

[0121] In some embodiments, the computing device(s) 120 may be operated by one or more users 172, such as one or more researchers and / or other individual(s). For example, the user(s) 172 may provide the scFv sequences 102 and / or scFv data 104 as input to the computing device(s) 120 (e.g., by uploading one or more files), and / or may provide user input specifying processing or other methods to be performed on the scFv data.

[0122] like Figure 1B As shown, software 130 includes multiple modules. Each module may include processor executable instructions that, when executed by at least one computer hardware processor, cause the at least one computer hardware processor to perform the (multiple) functions of the module. Such modules are sometimes referred to as "software modules" in this article. Figure 1B The modules shown in the figure include processor-executable instructions that, when executed by a computing device, cause the computing device to perform one or more processes (as described herein at least with respect to FIG. 2A to FIG. 2B , Figure 4A and FIG. 5A to FIG. 5B It should be understood that Figure 1B The modules shown in are illustrative, and in other embodiments, except Figure 1BIn addition to or in lieu of the modules shown in FIG. 1 , one or more other software modules may be used to implement the software 130. In other words, the software 130 may be implemented with Figure 1B The methods shown in are organized differently internally.

[0123] like Figure 1B As shown, the software 130 includes multiple software modules for processing scFv data, such as a protein structure prediction module 160, a molecular modeling module 162, a feature generation module 164, a thermal stability prediction module 166, and a scFv screening module 168. Figure 1B In the embodiment of FIG. 1 , the software 130 further includes a user interface module 170 for obtaining user input.

[0124] In some embodiments, protein structure prediction module 160 obtains scFv data (e.g., scFv data 104) from scFv data storage 146 and / or (multiple) users 172 (e.g., by users uploading scFv data). In some embodiments, the obtained scFv data includes sequence data of scFv (e.g., sequence data 104a).

[0125] In some embodiments, the protein structure prediction module 160 can be configured to predict the 3D structure of the scFv using sequence data. For example, the protein structure prediction module 160 can be configured to generate structural data (e.g., structural data 104b) of the scFv from the sequence of the scFv. To this end, the protein structure prediction module 160 can use protein structure prediction software, such as DeepAb software, SAbPred software, or AlphaFold software. This article at least relates to Figure 2B Act 252 of process 250 shown in describes a technique for using structure prediction software to obtain information indicative of the 3D structure of a scFv.

[0126] In some embodiments, the molecular modeling module 162 obtains scFv data (e.g., scFv data 104) from the scFv data store 146, (multiple) users 172 (e.g., by users uploading scFv data), and / or the protein structure prediction module 160. In some embodiments, the obtained scFv data includes sequence data (e.g., sequence data 104a) and / or structural data (e.g., structural data 104b) of the scFv.

[0127] In some embodiments, the molecular modeling module 162 is configured to determine the interaction energy metrics of the residues of the scFv sequence (e.g., scFv sequence 102) in its 3D conformation. For example, the molecular modeling module 162 can be configured to generate the interaction energy metrics of the scFv (e.g., which can be part of the energy data 104c). To this end, the molecular modeling module 162 can use molecular modeling software, such as Rosetta software, Schrodinger or the Molecular Operating Environment (MOE) software from the Chemical Computing Group. Figure 2B Act 254 of process 250 shown in describes a technique for determining an interaction energy metric using molecular modeling software.

[0128] In some embodiments, feature generation module 164 obtains scFv data (e.g., scFv data 104) from scFv data storage 146, (multiple) users 172 (e.g., by users uploading scFv data), molecular modeling module 162, and / or protein structure prediction module 160, and uses the obtained scFv data to generate a feature set for the corresponding scFv. For example, feature generation module 164 can generate a feature set for an scFv having scFv sequence 102.

[0129] In some embodiments, the feature generation module 164 generates the feature set by including at least some of the acquired data (e.g., scFv data 104) in the feature set. For example, the feature generation module 164 can generate the feature set to include the interaction energy metric of the scFv. For example, the feature generation module 164 can generate the feature set to include a two-dimensional (2D) matrix for each specific energy metric in the interaction energy metric, and the two-dimensional matrix stores the value of the specific energy metric between the i-th residue and the j-th residue of the scFv at its (i, j) position. The 2D matrix generated in this way can be provided as an input to a trained neural network model. For example, Figure 8 As shown in the example of , if there are 20 interaction energy metrics, 20 such matrices can be generated and can be provided as input to the neural network model via 20 input channels. Additionally or alternatively, feature generation module 164 can generate a feature set including the coding sequence of the scFv. For example, the sequence can be one-hot encoded. However, it should be understood that feature generation module 164 can include additional or alternative features in the feature set, as the various aspects of the technology described herein are not limited in this regard. This article at least relates to Figure 2B Action 256 of process 250 shown in describes a technique for generating a feature set for a scFv.

[0130] In some embodiments, thermal stability prediction module 166 obtains one or more feature sets from feature generation module 164, obtains a trained machine learning model from machine learning model data store 152 (which can be any suitable type of data store), and uses the obtained machine learning model to process the obtained feature set(s) to obtain a thermal stability indication of one or more scFvs. For example, thermal stability prediction module 166 can use trained machine learning model 108 to process a feature set generated for an scFv having sequence 102 to obtain a thermal stability indication 110 for the scFv. Figure 2B Techniques for predicting the thermal stability of scFvs using machine learning are described.

[0131] In some embodiments, the scFv screening module 168 is used to calculate and screen a group of scFv to identify a subgroup for subsequent production. To this end, the scFv screening module 168 obtains a determined thermal stability indication from the thermal stability prediction module 166 (e.g., by calling module 166 to determine the thermal stability indication of the scFv in the group), and uses the thermal stability indication to identify the scFv for subsequent production. For example, in some embodiments, the scFv screening module 168 compares the thermal stability indication with one or more standards to determine whether the thermal stability prediction meets one or more standards. If the thermal stability indication meets the standard, the scFv screening module 168 can identify the scFv with determined thermal stability for subsequent production. For example, the thermal stability indication can indicate the corresponding temperature or temperature range when the scFv is in thermal stability for each scFv. The output can be compared with a threshold temperature, and those scFvs indicating that the temperature is higher than the threshold can pass the screening step and be selected for subsequent production.

[0132] In some embodiments, scFv screening module 168 can perform computational screening based on user input (e.g., user input provided by user(s) 172 via user interface module 170). User input can specify one or more criteria (e.g., threshold temperature) for passing an scFv through screening. Additionally or alternatively, a user can provide input for manually selecting one or more scFvs for subsequent production (e.g., based on their determined thermal stability or any other factor).

[0133] In some embodiments, protein structure prediction module 160, molecular modeling module 162, and / or feature generation module 164 obtain scFv data via user interface 170 and / or one or more other interface modules (not shown). The data may be provided by a communication network (not shown) such as the Internet or any other suitable network, as aspects of the technology described herein are not limited in this regard.

[0134] like Figure 1B As shown, the system 150 also includes a scFv data store 146 and a machine learning model data store 152. In some embodiments, the software 130 obtains data from the scFv data store 146, the machine learning model data store 152, and / or (multiple) users 172 (e.g., by uploading data). In some embodiments, the software 130 further includes a machine learning model training module 154, which is used to train one or more machine learning models (e.g., stored in the machine learning model data store 152).

[0135] In some embodiments, the scFv data is obtained from a scFv data store 146. The scFv data store 146 can be of any suitable type (e.g., a database system, multiple files, flat files, etc.) and can store the scFv data in any suitable manner and in any suitable format, as aspects of the technology described herein are not limited in this respect. The scFv data store 146 can be part of or external to the computing device(s) 120.

[0136] In some embodiments, scFv data store 146 stores scFv data obtained for scFvs (as described herein at least with respect to Figure 1A . In some embodiments, the stored scFv data may have been previously uploaded by a user (e.g., (multiple) users 172), and / or from one or more public data stores and / or studies. In some embodiments, a portion of the scFv data may be processed by a protein structure prediction module 160 to generate information indicating the structure of the scFv. In some embodiments, a portion of the scFv data may be processed by a molecular modeling module 162 to determine the interaction energy of the residue pairs of the scFv. In some embodiments, a portion of the scFv data may be processed by a feature generation module 164 to generate a feature set of the scFv, which will be provided as input to a machine learning model. In some embodiments, a portion of the scFv data may be used to train one or more machine learning models (e.g., using a machine learning model training module 154).

[0137] In some embodiments, the thermal stability prediction module 166 obtains (pulls or is provided with) a trained machine learning model from the machine learning model data store 152. The machine learning model may be provided via a communication network (not shown) such as the Internet or any other suitable network, as aspects of the technology described herein are not limited to any particular communication network.

[0138] In some embodiments, the machine learning model data store 152 includes any suitable data store, such as a flat file, a data store, multiple files, or any suitable type of data storage device, as aspects of the technology described herein are not limited to any particular type of data store. The machine learning model data store 152 may be part of the software 130 (not shown) or may not be included in the software 130, such as Figure 1B shown.

[0139] In some embodiments, machine learning model data store 152 stores one or more machine learning models for predicting the thermal stability of scFv. Data store 152 can be of any suitable type (e.g., a database system, multiple files, flat files, etc.), and can store trained machine learning models in any suitable manner and in any suitable format, as the various aspects of the technology described herein are not limited in this regard. Data store 152 can be part of or external to computing device(s) 120.

[0140] In some embodiments, the machine learning model training module 154 (referred to herein as training module 154) can be configured to train the one or more machine learning models to predict the thermal stability of scFv. In some embodiments, the training module 154 uses a training set of scFv data to train the machine learning model. For example, the training module 154 can obtain training data from the scFv data store 146. In some embodiments, the training module 154 can provide (multiple) trained machine learning models to the machine learning model data store 152. Figure 4A Techniques for training machine learning models are described.

[0141] In some embodiments, the predicted thermal stability may be output by the thermal stability prediction module 166. For example, the predicted thermal stability may be output to a user(s) 172 via the user interface 170. Additionally or alternatively, the predicted thermal stability may be stored in a memory and / or transmitted to one or more other computing devices.

[0142] In some embodiments, the identified scFv for subsequent production can be output by the scFv screening module 168. For example, the identified scFv can be output to (multiple) users 172 via the user interface 170. Additionally or alternatively, the identified scFv can be stored in a memory and / or transmitted to one or more other computing devices.

[0143] The user interface 170 can be a graphical user interface (GUI), a text-based user interface, and / or any other suitable type of interface through which a user can provide input and view information generated by the software 130. For example, in some embodiments, the user interface can be a web page or web application accessible through an Internet browser. In some embodiments, the user interface can be a graphical user interface (GUI) of an app executed on a user's mobile device. In some embodiments, the user interface can include a plurality of selectable elements through which a user can interact. For example, the user interface can include a drop-down list, a check box, a text field, or any other suitable element.

[0144] FIG. 2A to FIG. 2B is a flow chart depicting an illustrative process for computationally screening a panel of scFvs to identify a subset of the panel of scFvs for production according to some embodiments of the technology described herein.

[0145] Figure 2A 2 is a flow chart of an illustrative process 200 for computationally screening a panel of scFvs according to some embodiments of the technology described herein. One or more actions of process 200 can be automatically performed by any suitable computing device(s). For example, the actions(s) can be performed by a laptop computer, a desktop computer, one or more servers, or a computer system in a cloud computing environment, such as described herein. Fig.24 The computer system 2400 and / or any other suitable manner. For example, in some embodiments, action 202 can be automatically performed by any suitable (multiple) computing device. As another example, action 204 can be automatically performed by any suitable (multiple) computing device.

[0146] Process 200 begins at act 202, where a thermal stability indicator for each scFv in a set of scFvs is determined using a trained machine learning model. As described above, the thermal stability indicator can refer to a temperature at which the scFv is stable or a temperature range that includes at least one temperature at which the scFv is stable.

[0147] In some embodiments, determining the thermal stability indication of the scFv comprises generating a feature set of the scFv and processing the feature set using a trained machine learning model. The output of the machine learning model can indicate the thermal stability of the scFv (e.g., a thermal stability indication). For example, the output can indicate the temperature at which the scFv is thermally stable. The temperature can be a TS50 temperature, a Tm temperature, or any other type of temperature indicating that the scFv is thermally stable. Additionally or alternatively, the output can indicate a temperature range that includes one or more temperatures at which the scFv is thermally stable. This article at least relates to Figure 2BThe process 250 shown in describes a technique for determining a thermal stability indicator for a particular scFv using a trained neural network model.

[0148] The group of scFvs can include any suitable number of scFvs. For example, the group of scFvs can include at least 25 scFvs, at least 50 scFvs, at least 75 scFvs, at least 100 scFvs, at least 200 scFvs, at least 300 scFvs, at least 400 scFvs, at least 500 scFvs, at least 600 scFvs, at least 700 scFvs, at least 800 scFvs, at least 900 scFvs, at least 1000 scFvs, at least 5000 scFvs, at least 10000 scFvs, between 100 and 1000 scFvs, between 100 and 10000 scFvs, or any other suitable range within these ranges. Thus, determining the thermal stability indication at act 202 may include determining at least 25, at least 50, at least 75, at least 100, at least 200, at least 300, at least 400, at least 500, at least 600, at least 700, at least 800, at least 900, at least 1000, at least 5000, at least 10000, between 100 and 1000, or between 100 and 10000 thermal stability indications.

[0149] Process 200 then proceeds to act 204, where a subset of the set of scFvs is identified for subsequent production based on the thermal stability indicator determined at act 202. In some embodiments, identifying scFvs in the set of scFvs for subsequent production includes identifying scFvs that are thermally stable and manufacturable.

[0150] In some embodiments, identifying a subgroup of scFvs involves determining whether the thermostability indication meets one or more criteria. For example, scFvs whose thermostability indications include temperatures exceeding a particular threshold can be identified. As another example, scFvs whose thermostability indications include a particular temperature range can be included. When the thermostability indication determined for a particular scFv meets one or more criteria, the scFv can be included in the scFv subgroup for subsequent production. If the thermostability indication does not meet one or more criteria, the scFv can be excluded from the scFv subgroup for subsequent production.

[0151] As described above, in some embodiments, act 204 may be performed using a computing device (e.g., Figure 1B In some embodiments, the method of claim 1 , wherein the method comprises performing the step of: ...

[0152] In some embodiments, the identified subgroup does not include any, some, or all of the scFvs included in the original set of scFvs. For example, the identified subgroup may include 0%, less than 10%, less than 25%, less than 50%, less than 75%, less than 90%, or all of the scFvs included in the original set of scFvs.

[0153] Process 200 then proceeds to act 206, where at least some of the scFvs included in the identified scFv subset are produced. In some embodiments, scFvs are produced using techniques known in the art.

[0154] In some embodiments, producing at least one of the scFvs in the scFv subgroup includes producing one, some, or all of the scFvs in the subgroup. For example, in some embodiments, at least 10%, at least 25%, at least 50%, at least 75%, at least 90%, or all of the scFvs included in the identified subgroup are produced at act 206.

[0155] In some embodiments, implementation process 200 may include: Figure 2A In some embodiments, process 200 may include only a subset of the actions included in the example flowchart (eg, only action 202, only actions 202 and 204).

[0156] Figure 2B 2 is a flow chart of an illustrative process 250 for determining the thermal stability of a first scFv according to some embodiments of the technology described herein. In some embodiments, process 250 can be used to implement act 202 of process 200. Process 250 can be performed by any suitable computing device (e.g., Figure 1B Executed by (multiple) computing devices 120 shown in .

[0157] Process 250 begins at act 252, where information indicating the 3D structure of a first scFv is obtained. In some embodiments, the information of the first scFv is previously obtained. Thus, in some embodiments, obtaining information indicating the 3D structure of the first scFv may include accessing the information (e.g., from a memory, over a network, by providing a file via an appropriate interface, etc.).

[0158] In other embodiments, obtaining information indicating the 3D structure of the first scFv includes generating the information. Therefore, in some embodiments, obtaining information indicating the 3D structure of the first scFv includes generating the information by processing the residue sequence of the first scFv using protein structure prediction software. The protein structure prediction software can be configured to output information indicating the 3D structure of the first scFv. Any suitable protein structure prediction software can be used. For example, DeepAb software can be used, and various aspects of the software are described in Ruffolo, Jeffrey A., Jeremias Sulam and Jeffrey J. Gray, "Antibody structure prediction using interpretable deep learning [antibody structure prediction using interpretable deep learning]" Patterns [pattern] 3.2 (2022): 100406, which is incorporated herein by reference in its entirety. As another example, the SAbPred software can be used, aspects of which are described in Dunbar, James et al., "SAbPred: a structure-based antibody prediction server" Nucleic acids research 44.W1 (2016): W474-W478, which is incorporated herein by reference in its entirety. As yet another example, the AlphaFold software can be used, aspects of which are described in Jumper, Johns et al., "Highly accurate protein structure prediction with AlphaFold" Nature 596, 583-589 (2021), which is incorporated herein by reference in its entirety.

[0159] Process 250 then proceeds to action 254, where an interaction energy metric is obtained for each of the multiple pairs of amino acid residues of the first scFv. In some embodiments, obtaining the interaction energy metric includes generating the energy metric by processing information indicating the 3D structure of the first scFv using a molecular modeling software. The molecular modeling software can be configured to output the interaction energy metric of the first scFv. Any molecular modeling software capable of estimating residue interaction energy metrics can be used. For example, Rosetta molecular modeling software can be used. The Rosetta software and its techniques for estimating residue interaction energy metrics are described in: Alford et al., (The Rosetta All-Atom Energy Function for Macromolecular Modeling and Design 440 J. Chem. Theory Comput. 13, 3031-3048 (2017), Chaudhury et al., (PyRosetta: a script-based interface for implementing molecular modeling algorithms using Rosetta, Bioinformatics, 26(5), 689-691 (2010), and Leaver-Fay et al., (Rosetta 3: An object-oriented software suite for the simulation and design of macromolecules, Methods in Enzymology, 545-574), each of which is incorporated herein by reference in its entirety. As another example, Schrodinger's And / or Prime software (Schrodinger Suite 2020-2 version, Schrodinger Company, New York, NY). As another example, the Molecular Operating Environment (MOE) software of the Chemical Computing Group (Molecular Operating Environment (MOE), 2020.09 Chemical Computing Group ULC, 1010 Sherboke St. West, Suite #910, Montreal, QC, Canada, H3A 2R7, 2022) can be used.

[0160] In some embodiments, interaction energy metrics are obtained for some or all of the residue pairs of the residue sequence of the first scFv. For example, interaction energy metrics for at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 98%, at least 99% or all of the residue pairs of the amino acid sequence of the first scFv can be obtained.

[0161] For a pair of residues, any of a variety of types of interaction energy metrics may be obtained. For example, the interaction energy metric(s) for a pair of residues may include one or more of the Lennard-Jones attraction and / or repulsion energies and van der Waals energies (e.g., “fa_atr” – the attraction energy between two atoms on different residues separated by a distance d, “fa_rep” – the repulsion energy between two atoms on different residues separated by a distance d, and / or “fa_intra-rep” – the repulsion energy between two atoms on the same residue separated by a distance d). As another example, the interaction energy metric(s) for a pair of residues may include one or more solvation energies (e.g., “fa_sol” – the Gaussian exclusion implicit solvation energy between protein atoms in different residues, “fa_intra_sol” – the Gaussian exclusion implicit solvation energy between protein atoms in the same residue, and / or “lk_ball_wtd” – the orientation of polar atoms assuming an ideal water geometry). dependent solvation). As another example, the interaction energy metric(s) for a pair of residues are electrostatic energies (e.g., "fa_elec" - the interaction energy between two unbonded charged atoms separated by a distance d). As another example, the interaction energy metric(s) for a pair of residues may include one or more of hydrogen bonds and / or disulfide bridge energies (e.g., "hbond_lr_bb" - the energy of long-range hydrogen bonds; "hbond_sr_bb" - the energy of short-range hydrogen bonds; "hbond_bb_sc" - the energy of backbone-side chain hydrogen bonds; "hbond_sc" - the energy of side chain-side chain hydrogen bonds; "dslf_fa13" - the energy of disulfide bridges). As another example, the interaction energy metric(s) for a pair of residues may include one or more backbone statistics (e.g., "rama_prepro" - the backbone angles given an amino acid type). and ψ; “omega” – backbone-dependent penalty for cis-ω dihedral angles that deviate from 0° and trans-ω dihedral angles that deviate from 180°; “p_aa_pp” – probability of and ψ; “pro_close” – penalty for open proline rings and proline ω bonding energy; “yhh_planarity” – non-planar tyrosine χ 3Sine penalty for dihedral angles). As another example, (multiple) interaction energy metrics for a pair of residues may include knowledge-based rotamer energies (e.g., "fa_dun" - the probability that the selected rotamer is native-like given the backbone values). Thus, the interaction energy metric may include any (some or all) of the aforementioned examples of energy metrics. These aforementioned examples are further described in Alford et al., (The Rosetta All-Atom Energy Function for Macromolecular Modeling and Design [Rosetta All-Atom Energy Function for Macromolecular Modeling and Design] 440 J. Chem. Theory Comput. [Journal of Chemical Theory and Computation] 13, 3031-3048 (2017), which is incorporated herein by reference in its entirety. In addition to or in lieu of any one or more of the aforementioned example energy metrics, one or more other energy metrics may be used.

[0162] The process 250 then proceeds to act 256, where a first feature set is generated for the first scFv, such that the first feature set can be provided as an input to a trained machine learning model at act 258 to obtain a corresponding output indicative of a first thermal stability of the first scFv. In some embodiments, generating the first feature set involves including at least some of the data obtained at act 252 and / or act 254 of process 250 in the first feature set.

[0163] For example, at act 256a, the interaction energy metric is included in the first feature set. In some embodiments, the interaction energy metric includes some or all of the interaction energy metrics obtained at act 254, examples of which are provided herein.

[0164] In some embodiments, including the interaction energy metric in the first feature set at act 256a includes generating one or more matrices of the interaction energy metric. For example, for each particular energy metric, a corresponding 2D matrix may be generated having values ​​of the particular metric for the i-th residue and the j-th residue in its entries in the i-th row and the j-th column. When the interaction energy metric includes multiple metrics, a 2D matrix may be generated for each of them and thus included as part of the first feature set generated at act 256a of process 250. Thus, the 2D matrix may be taken as input (e.g., via Figure 8 The different channels shown in FIG. 1 and FIG. 2 are provided to a trained machine learning model. Alternatively, a 3D matrix can be generated instead of multiple 2D matrices, as aspects of the technology described herein are not limited in this regard.

[0165] Additionally or alternatively, at act 256b, the encoded first residue sequence is included in the first feature set. In some embodiments, encoding the first residue sequence to obtain the encoded first residue sequence may be performed using any suitable encoding technique, such as one-hot encoding. In some embodiments, the first residue sequence is encoded to form a particular input dimension. For example, the first residue sequence may be encoded to form an input dimension (V H +V L +3)×21, where V H and V L They correspond to the heavy chain sequence and light chain sequence of the scFv residue sequence, respectively.

[0166] Although not in Figure 2B , but it should be understood that one or more additional or alternative features may be included in the first feature set, as aspects of the technology described herein are not limited in this respect.

[0167] Next, process 250 then proceeds to act 258 where the first feature set is provided as input to the trained machine learning model to obtain an output indicative of the thermal stability of the first scFv.

[0168] In some embodiments, the machine learning model is of any suitable type. For example, the machine learning model can be a neural network, such as a convolutional neural network (CNN). The CNN can have one or more two-dimensional convolutional layers, one or more three-dimensional convolutional layers, and / or fully connected layers.

[0169] As a specific example, a CNN may have Figure 8 The architecture shown in , which includes one or more max pooling layers ( Figure 8 ), one or more 2D convolutional layers, one or more nonlinear layers (e.g., Figure 8 ), one or more batch normalization layers, and a dense layer. Figure 8 The illustrated architecture shows that both the energy metric and the one-hot encoded sequence are provided as input. However, as described herein, in some embodiments, only the energy metric or only the one-hot encoded sequence may be provided as input. Thus, in some embodiments, Figure 8 The upper branch or the lower branch or both branches of the architecture shown.

[0170] As another example, the machine learning model can be a pre-trained language model (for performing zero-shot learning predictions, or for tuning by adding a supervision head using transfer learning with a small amount of training data), such as the Bidirectional Encoder Representations from Transformers (BERT) model (e.g., the ESM-1b language model or the ESM-1v language model) or the UniRep language model, which are described herein. Fig.14 As shown), such a model can be used.

[0171] In some embodiments, the machine learning model includes multiple machine learning models (e.g., an integration of machine learning models). For example, the machine learning model may include an integration of neural networks. In some embodiments, implementing the integration of machine learning models includes training each of multiple (e.g., two or more, three or more, etc.) machine learning models on different training data sets, predicting thermal stability using each of the trained machine learning models, and averaging their predictions (e.g., weighted average). The outputs of multiple machine learning models can be combined in any other way, as the various aspects of the technology described herein are not limited in this regard. The integration of machine learning models can be obtained in part by using any suitable bagging or boosting techniques.

[0172] In some embodiments, a machine learning model is trained to process a first feature set to obtain a probability that the thermal stability of a first scFv belongs to each of a plurality of thermal stability categories. For example, this may include determining the probability that the first scFv is thermally stable within each of a plurality of temperature ranges. As a non-limiting example, the thermal stability categories may include the following four categories: below 50°C, 50°C to 60°C, 60°C to 70°C, and above 70°C.

[0173] In some embodiments, based on the determined probabilities, the machine learning model is configured to predict the category associated with the highest probability in the plurality of categories. For example, this may include identifying the temperature range associated with the highest probability in the plurality of temperature ranges. The identified temperature range includes one or more temperatures at which the first scFv is thermally stable.

[0174] Additionally or alternatively, based on the determined probability, the machine learning model can be configured to determine the temperature at which the first scFv is thermally stable. For example, the temperature can be the TS50 temperature. As another example, the temperature can correspond to the thermal melting temperature (Tm.). In some embodiments, determining the temperature at which the first scFv is thermally stable includes determining the temperature as a weighted linear combination of the average values ​​of the temperature range obtained by weighting the probability determined using the machine learning model.

[0175] At act 260, process 250 includes determining whether there is another scFv in the set of scFvs for which thermal stability can be determined. When it is determined at act 260 that there is another scFv for which thermal stability is to be determined, acts 252 to 258 are repeated for the other scFv. For example, for a second scFv, this would include determining a second feature set and providing the second feature set as input to the trained machine learning model to determine a thermal stability indicator for the second scFv.

[0176] Figure 3A is a diagram of an illustrative technique for computationally screening a set of scFvs using a trained machine learning model that is trained to generate an indication of the thermal stability of the scFv based on an input comprising interaction energy metrics for residue pairs in the scFv according to some embodiments of the technology described herein.

[0177] like Figure 3A As shown in the embodiment of the present invention, the exemplary technology 300 starts with a group of scFvs 302. In some embodiments, the group of scFvs 302 are candidates for production. For example, the scFvs included in the group 302 can have one or more desired properties (e.g., affinity, specificity, etc.). The group of scFvs can have any suitable size M, examples of which are provided herein, including reference Figure 2A .exist Figure 3A In the example of , the group of scFvs 302 includes a first scFv 302-1, a second scFv 302-2, and a second scFv 302-M.

[0178] Each scFv in group 302 has a corresponding residue sequence. For example, the first scFv 302-1 has a first residue sequence, the second scFv 302-2 has a second residue sequence, and the Mth scFv 302-M has an Mth residue sequence. In some embodiments, the first, second, and Mth residue sequences are different from each other. For example, these sequences can differ from each other by one or more residues.

[0179] like Figure 3A As shown, the example technique 300 involves generating a feature set for each scFv included in the set of scFvs 302. This includes, for example, generating a first feature set 304-1 for a first scFv 302-1, generating a second feature set 304-2 for a second scFv 302-2, and generating an Mth feature set 304-M for an Mth scFv 302-M. Figure 2B Act 256 of process 250 shown in FIG. 2 describes a technique for generating a feature set.

[0180] In some embodiments, the feature set generated for the scFv includes interaction energy metrics for each of a plurality of pairs of residues in the scFv. Figure 3A As shown, the first feature set 304-1 includes an interaction energy metric 322-1 for each of the multiple pairs of residues in the first residue sequence, the second feature set 304-2 includes an interaction energy metric 322-2 for each of the multiple pairs of residues in the second residue sequence, and the Mth feature set 304-M includes an interaction energy metric 322-M for each of the multiple pairs of residues in the Mth residue sequence. Figure 2B Techniques for obtaining interaction energy measures for residue pairs are described.

[0181] exist Figure 3A In the illustrated embodiment, after generating a feature set for each scFv in the group of scFvs 302, the generated feature sets 304-1, 304-2, ... 304-M are provided as input to a trained machine learning model 306.

[0182] In some embodiments, the machine learning model 306 is trained to predict the thermal stability indication of the scFv based on the feature set provided as input to the machine learning model 306. For example, the first feature set 304-1 can be provided to the machine learning model 306 to obtain an output 308-1 indicating the thermal stability of the first scFv 302-1. The second feature set 304-2 can be provided as input to the machine learning model 306 to obtain an output 308-2 indicating the thermal stability of the second scFv 302-2. The Mth feature set 304-M can be provided as input to the machine learning model 306 to obtain an output 308-M indicating the thermal stability of the Mth scFv 302-M. At least one embodiment of the present invention relates to Figure 2B Techniques for predicting the thermal stability of scFvs using machine learning models are described.

[0183] In some embodiments, the example technique 300 includes identifying a subset 310 of the set of scFvs 302 based on the determined thermal stability indications (e.g., first thermal stability 308-1, second thermal stability 308-2, and Mth thermal stability 308-M). In some embodiments, identifying scFvs to be included in the subset 310 includes identifying scFvs having thermal stability that meets one or more criteria. This can include, for example, comparing thermal stability to a threshold temperature and identifying those scFvs having thermal stability that exceeds the threshold. The identified scFvs can have thermal stability that makes them suitable for production. This article at least relates to Figure 2A Act 204 of process 200 shown in FIG. 2 describes a technique for identifying scFvs for subsequent production.

[0184] In some embodiments, the subset of scFvs includes one or more of the scFvs included in the set of original scFvs 302. Figure 3A In the embodiment shown in FIG. 3 , the identified subgroup 310 includes N scFvs. For example, subgroup 310 includes a first scFv 302-1, a second scFv 302-2, and an Nth scFv 302-N. In some embodiments, N is less than 10%, less than 20%, less than 30%, less than 40%, less than 50%, less than 60%, less than 70%, less than 80%, or less than 90%, or less than 100% of M. In some embodiments, N is equal to M.

[0185] In some embodiments, at least one of the scFvs included in the scFv subset is subsequently produced. For example, one, some, or all of the scFvs included in subset 310 can be subsequently produced.

[0186] Figure 3B is a diagram of an illustrative technique for computationally screening a set of scFvs using a trained machine learning model according to some embodiments of the technology described herein, the trained machine learning model being trained to generate an indication of the thermal stability of the scFv based on inputs comprising interaction energy metrics for residue pairs in the scFv and features representing the sequence of the scFv.

[0187] like Figure 3B As shown in the embodiment of FIG. 3 , the exemplary technique 340 starts with a set of scFv 302. Figure 3A A set of exemplary scFvs are described.

[0188] In some embodiments, the example technique 340 includes generating a feature set for each scFv included in the set of scFvs 302. This includes, for example, generating a first feature set 344-1 for the first scFv 302-1, generating a second feature set 344-2 for the second scFv 302-2, and generating an Mth feature set 344-M for the Mth scFv 302-M. Figure 2B Act 256 of process 250 shown in FIG. 2 describes a technique for generating a feature set.

[0189] In some embodiments, the feature set includes interaction energy metrics for each of a plurality of pairs of residues of the corresponding scFv. Figure 3BAs shown, the first feature set 344-1 includes an interaction energy metric 352-1 for each of the multiple pairs of residues in the first residue sequence, the second feature set 344-2 includes an interaction energy metric 352-2 for each of the multiple pairs of residues in the second residue sequence, and the Mth feature set 344-M includes an interaction energy metric 352-M for each of the multiple pairs of residues in the Mth residue sequence. Figure 2B Techniques for obtaining interaction energy measures for residue pairs are described.

[0190] In addition, Figure 3B In the illustrated embodiment, the feature set includes the coding residue sequence of the corresponding scFv. For example, the first feature set 344-1 includes the coding sequence 354-1 of the first scFv 302-1, the second feature set 344-2 includes the coding sequence 354-2 of the second scFv 302-2, and the Mth feature set 344-M includes the coding sequence 354-M of the Mth scFv 302-M. Figure 2B Techniques used to obtain coding residue sequences are described.

[0191] exist Figure 3B In the illustrated embodiment, after generating a feature set for each scFv in the set of scFvs 302, the example technique 340 includes providing the generated feature set as input to a trained machine learning model 346.

[0192] In some embodiments, the machine learning model 346 is trained to predict the thermal stability indication of the scFv based on the feature set provided as input to the machine learning model 346. For example, a first feature set 344-1 can be provided to the machine learning model 346 to obtain an output 348-1 indicating the thermal stability of the first scFv 302-1. A second feature set 344-2 can be provided as input to the machine learning model 346 to obtain an output 348-2 indicating the thermal stability of the second scFv 302-2. An Mth feature set 344-M can be provided as input to the machine learning model 346 to obtain an output 348-M indicating the thermal stability of the Mth scFv 302-M. At least one embodiment of the present invention relates to Figure 2B Techniques for predicting the thermal stability of scFvs using machine learning models are described.

[0193] In some embodiments, the example technique 340 includes identifying a subset 350 of the set of scFvs 302 based on the determined thermal stability indicators (e.g., the first thermal stability 348-1, the second thermal stability 348-2, and the Mth thermal stability 348-M). Figure 3A Exemplary techniques for identifying subsets of scFvs are described.

[0194] Figure 3C is a diagram of an illustrative technique for computationally screening a set of scFvs using a trained machine learning model that is trained to generate an indication of the thermal stability of an scFv based on an input comprising features representing a sequence of the scFv, according to some embodiments of the technology described herein.

[0195] like Figure 3C As shown in the embodiment of FIG. 3 , the exemplary technique 360 ​​starts with a set of scFv 302. Figure 3A A set of exemplary scFvs are described.

[0196] In some embodiments, the example technique 360 ​​includes generating a feature set for each scFv included in the set of scFvs 302. This includes, for example, generating a first feature set 364-1 for the first scFv 302-1, generating a second feature set 364-2 for the second scFv 302-2, and generating an Mth feature set 364-M for the Mth scFv 302-M. Figure 2B Act 256 of process 250 shown in FIG. 2 describes a technique for generating a feature set.

[0197] In some embodiments, the feature set includes the coding residue sequence of the corresponding scFv. For example, the first feature set 364-1 includes the coding sequence 372-1 of the first scFv 302-1, the second feature set 364-2 includes the coding sequence 372-2 of the second scFv 302-2, and the Mth feature set 364-M includes the coding sequence 372-M of the Mth scFv 302-M. Figure 2B Techniques used to obtain coding residue sequences are described.

[0198] exist Figure 3C In the illustrated embodiment, after generating a feature set for each scFv in the set of scFvs 302, the example technique 360 ​​includes providing the generated feature set as input to a trained machine learning model 366.

[0199] In some embodiments, the machine learning model 366 is trained to predict the thermal stability indication of the scFv based on the feature set provided as input to the machine learning model 366. For example, a first feature set 364-1 can be provided to the machine learning model 366 to obtain an output 368-1 indicating the thermal stability of the first scFv 302-1. A second feature set 364-2 can be provided as input to the machine learning model 366 to obtain an output 368-2 indicating the thermal stability of the second scFv 302-2. An Mth feature set 364-M can be provided as input to the machine learning model 366 to obtain an output 368-M indicating the thermal stability of the Mth scFv 302-M. At least one embodiment of the present invention relates to Figure 2B Techniques for predicting the thermal stability of scFvs using machine learning models are described.

[0200] In some embodiments, the example technique 360 ​​includes identifying a subset 370 of the set of scFvs 302 based on the determined thermal stability indicators (e.g., the first thermal stability 368-1, the second thermal stability 368-2, and the Mth thermal stability 368-M). Figure 3A Exemplary techniques for identifying subsets of scFvs are described.

[0201] Figure 4A 4 is a flow chart of an illustrative process 400 for training a machine learning model to generate an indication of the thermal stability of a scFv according to some embodiments of the technology described herein. The process 400 can be performed by any suitable computing device(s). For example, the processes can be performed by a laptop computer, a desktop computer, one or more servers, or a computer system in a cloud computing environment, as described herein. Fig.24 The computer system 2400 described above may be executed in any other suitable manner. In some embodiments, the software module (such as the one described herein) Figure 1B The machine learning model training module 154) includes processor executable instructions that, when executed by a computing device, cause the computing device to perform process 400.

[0202] Process 400 begins at act 402, where an experimental determination of a thermal stability indicator of the scFv is performed. In some embodiments, the experimental determination of the thermal stability indicator comprises producing the scFv and analyzing the produced scFv to determine the experimental thermal stability indicator. This may include, for example, experimentally determining a temperature corresponding to half-maximal binding of the scFv (TS50) and / or a temperature corresponding to the thermal melting (Tm) temperature of the scFv as an indicator of thermal stability. Figure 4B Techniques used to produce scFvs and experimentally determine indicators of thermal stability are described.

[0203] Process 400 then proceeds to act 404, where information indicative of the 3D structure of the scFv is obtained using the residue sequence of the scFv. Figure 2B Act 252 of process 250 shown in describes a technique for obtaining information indicative of the 3D structure of a scFv.

[0204] Process 400 then proceeds to act 406, where information about the 3D structure of the scFv is used to obtain interaction energy metrics for residue pairs of the residue sequence of the scFv. Figure 2B Act 254 of process 250 shown in describes a technique for obtaining an interaction energy metric.

[0205] When the machine learning model employs a particular type of feature, then process 400 includes generating features of that type and using them for training. The type of feature may include, for example, only energy features, only sequence features, or a combination of energy features and sequence features. For example, at action 408, process 400 includes generating a feature set for a scFv. Generating the feature set may include, at action 408a, including in the feature set an interaction energy metric for a residue pair of a scFv residue sequence. Additionally or alternatively, generating the feature set may include, at action 408b, including in the feature set an encoded residue sequence of the scFv. This article at least relates to Figure 2B Action 256 of process 250 shown in describes a technique for generating a feature set for a scFv.

[0206] In some embodiments, when the machine learning model employs sequence features, training the machine learning model may include providing the sequence features to the machine learning model in a particular order. For example, as described herein, the sequence features may include a sequence of encoded residues. In the case of scFv, the encoded scFv sequence may be provided to the machine learning model as an encoded VH sequence followed by an encoded VL sequence ("VH-VL"), or as an encoded VL sequence followed by an encoded VH sequence ("VL-VH").

[0207] In some embodiments, the machine learning model can be trained using training data that includes sequences that are all in the same order. For example, each sequence in the training data can consist of an encoded VH sequence followed by an encoded VL sequence. As another example, each sequence in the training data can consist of an encoded VL sequence followed by an encoded VH sequence.

[0208] If during training, all sequence data is provided to the machine learning model in the same order (e.g., all VH-VL or all VL-VH), new data may also be provided in the same order. If new sequence data is provided in a different order, the machine learning model may not perform well when processing the new data because it is not trained on the data provided in that order. For example, if the machine learning model is only trained on scFV sequences specified in VH-VL order, it may not perform well when processing new sequences provided in VL-VH order. In fact, in the case of scFv, VH-VL molecules may be physically different from VL-VH molecules. If the machine learning model is trained only on scFv sequences specified in one order, the machine learning model may not have the opportunity to understand the basic physical differences of scFv molecules specified in another order. Therefore, the machine learning model may not perform well when processing scFv sequences provided in other orders. In contrast, for antibodies, since VH and VL can be separate chains, whether the model sees VH-VL or VL-VH can be purely a matter of notation, the underlying physical entities can be the same, and so the performance differences may be simply due to the order of input. Therefore, when dealing with antibodies, performance differences due to input order may not be as significant as when dealing with scFVs.

[0209] In some embodiments, training the machine learning model includes providing the machine learning model with sequence data in two orders. For example, the sequence data may be provided in the order of an encoded VL sequence followed by an encoded VH sequence and an encoded VH sequence followed by an encoded VL sequence. If the sequence data is provided to the machine learning model in two orders (e.g., VH-VL and VL-VH) during training, new data may be provided to the machine learning model in two orders. Since the machine learning model is trained using sequence data provided in two orders, its performance may be consistent regardless of the order in which the new sequence data is provided.

[0210] Process 400 then proceeds to act 410 where a machine learning model is trained using the experimentally determined thermal stability indication determined at act 402 and the feature set generated at act 408 .

[0211] In some embodiments, training the machine learning model at act 410 includes estimating parameters of the machine learning model based on the training data. In some embodiments, the estimation can be done iteratively (e.g., using an iterative gradient descent technique). In some embodiments, the estimation can be done using optimization software to adjust the parameters. For example, in some embodiments, an ADAM optimizer can be used, aspects of which are described in Kingma, Diederik P. and Jimmy Ba, "Adam: A Method for Stochastic Optimization" (published in Proceedings of the 3 rd International Conference on Learning Representations [3rd International Learning Representations Conference Proceedings], ICLR (2015), which is incorporated herein by reference in its entirety. In some embodiments, estimating the parameters includes estimating the weights of the connections.

[0212] In some embodiments, any suitable hyperparameters may be used during training at act 410. Hyperparameters may be set manually or using any other suitable method. Non-limiting examples of hyperparameters include the number of layers, batch size, number of filters, kernel size, number of rounds, pooling size, and learning rate. However, it should be understood that any other suitable hyperparameters may be used during training at act 410.

[0213] Figure 4B 450 is a diagram of an example technique 450 for experimentally generating data for training a model to predict the thermal stability of a scFv 452 according to some embodiments of the technology described herein. As shown, at 454, the scFv is produced in, for example, an E. coli culture. Then at 456, the cells are lysed (e.g., using a freeze / thaw cycle), and the scFv is extracted and incubated. After incubation, the lysate can be incubated with target transfected cells (e.g., CHO cells), and the bound scFv can be detected and analyzed by flow cytometry or any other suitable technique. In some embodiments, the results of the analysis are used to estimate thermal stability at 458. For example, this can include estimating the TS50 value and / or T value of the scFv. m Example techniques for producing scFvs and experimentally estimating the thermal stability of scFvs are described herein at least in part with respect to "Generation of scFvs," "TS50 screening assay," and "nanoDSFT method."

[0214] The technology described herein is not limited to being applied to single-chain variable fragments, and can also be applied to other constructs described herein. For example, as mentioned above, the technology described herein can be applied to any type of antibody. The antibody sequence can be used to predict the structure, and the structure can be used to calculate the energy metric. Then, the energy metric (and optionally, the encoding of the input sequence) can be provided as input to the trained machine learning model to obtain the thermal stability indication of the antibody.

[0215] Figure 5A 5 is a flow chart of an illustrative process 500 for computationally screening a panel of antibodies according to some embodiments of the technology described herein. One or more actions of process 500 can be automatically performed by any suitable computing device(s). For example, the actions(s) can be performed by a laptop computer, a desktop computer, one or more servers, or a computer system in a cloud computing environment, such as described herein. Fig.24 The computer system 2400 and / or any other suitable manner. For example, in some embodiments, action 502 can be automatically performed by any suitable (multiple) computing devices. As another example, action 504 can be automatically performed by any suitable (multiple) computing devices.

[0216] Process 500 begins at act 502 where a thermal stability indicator for each antibody in a set of antibodies is determined using a trained machine learning model. As described above, the thermal stability indicator may refer to a temperature at which the antibody is stable or a temperature range that includes at least one temperature at which the antibody is stable.

[0217] In some embodiments, determining the thermal stability indication of an antibody comprises generating a feature set of the antibody and processing the feature set using a trained machine learning model. The output of the machine learning model may indicate the thermal stability of the antibody (e.g., a thermal stability indication). For example, the output may indicate the temperature at which the antibody is thermally stable. The temperature may be a TS50 temperature, a Tm temperature, or any other type of temperature indicating that the antibody is thermally stable. Additionally or alternatively, the output may indicate a temperature range that includes one or more temperatures at which the antibody is thermally stable. This article at least relates to Figure 5B The process 550 shown in , describes a technique for determining a thermal stability indicator for a particular antibody using a trained neural network model.

[0218] The group of antibodies can include any suitable number of antibodies. For example, the group of scFvs can include at least 25 antibodies, at least 50 antibodies, at least 75 antibodies, at least 100 antibodies, at least 200 antibodies, at least 300 antibodies, at least 400 antibodies, at least 500 antibodies, at least 600 antibodies, at least 700 antibodies, at least 800 antibodies, at least 900 antibodies, at least 1000 antibodies, at least 5000 antibodies, at least 10000 antibodies, between 100 and 1000 antibodies, between 100 and 10000 antibodies, or any other suitable range within these ranges. Thus, determining the thermal stability indication at act 502 may include determining at least 25, at least 50, at least 75, at least 100, at least 200, at least 300, at least 400, at least 500, at least 600, at least 700, at least 800, at least 900, at least 1000, at least 5000, at least 10000, between 100 and 1000, or between 100 and 10000 thermal stability indications.

[0219] Process 500 then proceeds to act 504, where a subset of the panel of antibodies is identified for subsequent production based on the thermal stability indication determined at act 502. In some embodiments, identifying antibodies in the panel of antibodies for subsequent production includes identifying antibodies that are thermally stable and manufacturable.

[0220] In some embodiments, identifying a subset of antibodies involves determining whether the thermal stability indication meets one or more criteria. For example, antibodies whose thermal stability indications include temperatures exceeding a particular threshold can be identified. As another example, antibodies whose thermal stability indications include a particular temperature range can be included. When the thermal stability indication determined for a particular antibody meets one or more criteria, the antibody can be included in the subset of antibodies for subsequent production. If the thermal stability indication does not meet one or more criteria, the antibody can be excluded from the subset of antibodies for subsequent production.

[0221] As described above, in some embodiments, act 504 may be performed using a computing device (e.g., Figure 1B In some embodiments, the method of claim 1 , wherein the method comprises performing the following steps: ...

[0222] In some embodiments, the identified subgroup does not include any, includes some, or includes all of the antibodies included in the group of original antibodies. For example, the identified subgroup may include 0%, less than 10%, less than 25%, less than 50%, less than 75%, less than 90%, or all of the antibodies included in the group of original antibodies.

[0223] Process 500 then proceeds to act 506, where at least some of the antibodies included in the identified subset of antibodies are produced. In some embodiments, antibodies are produced using techniques known in the art.

[0224] In some embodiments, producing at least one of the antibodies in the subset of antibodies includes producing one, some, or all of the antibodies in the subset. For example, in some embodiments, at least 10%, at least 25%, at least 50%, at least 75%, at least 90%, or all of the antibodies included in the identified subset are produced at action 506.

[0225] In some embodiments, implementation process 500 may include: Figure 5A In some embodiments, process 500 may include only a subset of the actions included in the example flowchart (eg, only action 502, only actions 502 and 504).

[0226] Figure 5B 5 is a flow chart of an illustrative process 550 for determining the thermal stability of a first antibody according to some embodiments of the technology described herein. In some embodiments, act 502 of process 500 can be implemented using process 550. Process 550 can be performed by any suitable computing device(s) (e.g., Figure 1B Executed by (multiple) computing devices 120 shown in .

[0227] Process 550 begins at action 552, where information indicating a 3D structure of a first antibody is obtained. In some embodiments, the information of the first antibody is previously obtained. Thus, in some embodiments, obtaining information indicating the 3D structure of the first antibody may include accessing the information (e.g., from a memory, over a network, by providing a file via an appropriate interface, etc.).

[0228] In other embodiments, obtaining information indicating the 3D structure of the first antibody includes generating the information. Thus, in some embodiments, obtaining information indicating the 3D structure of the first antibody includes generating the information by processing the residue sequence of the first antibody using protein structure prediction software. The protein structure prediction software can be configured to output information indicating the 3D structure of the first antibody. Any suitable protein structure prediction software can be used. Figure 2BAction 252 of describes an example of protein structure prediction software.

[0229] Process 550 then proceeds to act 554, where an interaction energy metric is obtained for each of the plurality of pairs of amino acid residues of the first antibody. In some embodiments, obtaining the interaction energy metric comprises generating the energy metric by processing information indicating the 3D structure of the first antibody using molecular modeling software. The molecular modeling software can be configured to output the interaction energy metric for the first antibody. Any molecular modeling software capable of estimating residue interaction energy metrics can be used. Figure 2B Action 254 of describes an example of molecular modeling software.

[0230] In some embodiments, interaction energy metrics are obtained for some or all of the residue pairs of the residue sequence of the first antibody. For example, interaction energy metrics for at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 98%, at least 99% or all of the residue pairs of the amino acid sequence of the first antibody can be obtained.

[0231] For a pair of residues, any of a number of types of interaction energy measures can be obtained. Figure 2B An example of an interaction energy metric is described in action 254 of FIG.

[0232] Process 550 then proceeds to act 556, where a first feature set is generated for the first antibody, such that the first feature set can be provided as an input to a trained machine learning model at act 558 to obtain a corresponding output indicative of a first thermal stability of the first antibody. In some embodiments, generating the first feature set involves including at least some of the data obtained at act 552 and / or act 554 of process 550 in the first feature set.

[0233] For example, at act 556a, the interaction energy metric is included in the first feature set. In some embodiments, the interaction energy metric includes some or all of the interaction energy metrics obtained at act 554, examples of which are provided herein.

[0234] In some embodiments, including the interaction energy metric in the first feature set at act 556a includes generating one or more matrices of the interaction energy metric. For example, for each particular energy metric, a corresponding 2D matrix may be generated having values ​​of the particular metric for the i-th residue and the j-th residue in its entries in the i-th row and the j-th column. When the interaction energy metric includes multiple metrics, a 2D matrix may be generated for each of them and thus included as part of the first feature set generated at act 556a of process 550. Thus, the 2D matrix may be taken as input (e.g., via Figure 8 The different channels shown in FIG. 1 and FIG. 2 are provided to a trained machine learning model. Alternatively, a 3D matrix can be generated instead of multiple 2D matrices, as aspects of the technology described herein are not limited in this regard.

[0235] Additionally or alternatively, at act 556b, the encoded first residue sequence is included in the first feature set. In some embodiments, encoding the first residue sequence to obtain the encoded first residue sequence may be performed using any suitable encoding technique, such as one-hot encoding. In some embodiments, the first residue sequence is encoded to form a particular input dimension. For example, the first residue sequence may be encoded to form an input dimension (V H +V L +3)×21, where V H and V L They correspond to the heavy and light chain sequences of the antibody residue sequences, respectively.

[0236] Although not in Figure 5B , but it should be understood that one or more additional or alternative features may be included in the first feature set, as aspects of the technology described herein are not limited in this respect.

[0237] Next, process 550 proceeds to act 558, where the first feature set is provided as input to the trained machine learning model to obtain an output indicative of the thermal stability of the first antibody.

[0238] In some embodiments, the machine learning model is of any suitable type. For example, the machine learning model can be a neural network, such as a convolutional neural network (CNN). The CNN can have one or more two-dimensional convolutional layers, one or more three-dimensional convolutional layers, and / or fully connected layers.

[0239] As a specific example, a CNN may have Figure 8 The architecture shown in , which includes one or more max pooling layers ( Figure 8Adaptive MaxPool 1D and Adaptive Max Pool2D layers in ), one or more 2D convolutional layers, one or more non-linear layers (e.g., Figure 8 ), one or more batch normalization layers, and a dense layer. Figure 8 The illustrated architecture shows that both the energy metric and the one-hot encoded sequence are provided as input. However, as described herein, in some embodiments, only the energy metric or only the one-hot encoded sequence may be provided as input. Thus, in some embodiments, Figure 8 The upper branch or the lower branch or both branches of the architecture shown.

[0240] As another example, the machine learning model can be a pre-trained language model (for performing zero-shot learning predictions, or for tuning by adding a supervision head using transfer learning with a small amount of training data), such as the Bidirectional Encoder Representations from Transformers (BERT) model (e.g., the ESM-1b language model or the ESM-1v language model) or the UniRep language model, which are described herein. Fig.14 As shown), such a model can be used.

[0241] In some embodiments, the machine learning model includes multiple machine learning models (e.g., an integration of machine learning models). For example, the machine learning model may include an integration of neural networks. In some embodiments, implementing the integration of machine learning models includes training each of multiple (e.g., two or more, three or more, etc.) machine learning models on different training data sets, predicting thermal stability using each of the trained machine learning models, and averaging their predictions (e.g., weighted average). The outputs of multiple machine learning models can be combined in any other way, as the various aspects of the technology described herein are not limited in this regard. The integration of machine learning models can be obtained in part by using any suitable bagging or boosting techniques.

[0242] In some embodiments, a machine learning model is trained to process the first feature set to obtain a probability that the thermal stability of the first antibody belongs to each of a plurality of thermal stability categories. For example, this may include determining the probability that the first antibody is thermally stable within each of a plurality of temperature ranges. As a non-limiting example, the thermal stability categories may include the following four categories: below 50°C, 50°C to 60°C, 60°C to 70°C, and above 70°C.

[0243] In some embodiments, based on the determined probabilities, the machine learning model is configured to predict the category associated with the highest probability in the plurality of categories. For example, this may include identifying the temperature range associated with the highest probability in the plurality of temperature ranges. The identified temperature range includes one or more temperatures at which the first antibody is thermally stable.

[0244] Additionally or alternatively, based on the determined probability, the machine learning model can be configured to determine the temperature at which the first antibody is thermally stable. For example, the temperature can be the TS50 temperature. As another example, the temperature can correspond to the thermal melting temperature (Tm.). In some embodiments, determining the temperature at which the first antibody is thermally stable includes determining the temperature as a weighted linear combination of the average values ​​of the temperature range obtained by weighting the probability determined using the machine learning model.

[0245] At act 560, process 550 includes determining whether there is another antibody in the set of antibodies for which thermal stability can be determined. When it is determined at act 560 that there is another antibody for which thermal stability is to be determined, acts 552 to 558 are repeated for the other antibody. For example, for a second antibody, this would include determining a second feature set and providing the second feature set as input to a trained machine learning model to determine a thermal stability indication for the second antibody.

[0246] Example

[0247] Machine learning techniques are developed for predicting the thermal stability of scFvs. In some embodiments, the developed machine learning techniques utilize different features generated for scFvs to predict their thermal stability. Figure 6 is a diagram of an example of a scFv 602, a feature set 604 that can be generated for the scFv 602, and a machine learning model 606 trained to predict the thermal stability of the scFv 602.

[0248] Figure 7 It is a diagram of an example technique for training and using various machine learning models to predict the thermal stability of scFv on a data set (e.g., a labeled TS50 data set) 702. Branch 704 shows transfer learning using an unsupervised network (such as a pre-trained language model (PTLM)), which can be used to predict thermal stability by zero-shot learning prediction and fine-tuning prediction. Branch 706 shows a supervised model, such as a neural network architecture (e.g., a supervised convolutional neural network (CNN)), which is trained to use features obtained from scFv to predict the thermal stability of the scFv. Both types of models can be used to predict thermal stability 708, calculate verification experimental design 710, or for any other suitable purpose, because the various aspects of the technology described herein are not limited in this regard.

[0249] A. Supervised Neural Networks

[0250] Figure 8 FIG. 5 is a diagram showing an example architecture of a supervised convolutional neural network (CNN) trained to predict the thermal stability of an scFv using sequence features and / or energy features generated for the scFv. The ADAM optimizer was used with a categorical cross entropy (CCE) loss and 10 -3 The parameters of the example model are estimated using a learning rate of . The model is trained using at least the datasets described in the Datasets section of this paper.

[0251] like Figure 8 As shown, the input scFv sequence 802 is processed using protein structure prediction software 804 (e.g., DeepAb, AlphaFold, etc.) to generate information indicating the 3D structure of the scFv. Using the Rosetta ref2015 energy function, the structural information is used to evaluate the thermodynamic characteristics 806 of each scFv (the total energy is divided into monomer iI residue energy and double body ij residue energy). The contribution of the i-th residue to each j-th residue (where j∈N, so that N=total number of residues) can be listed and classified in the ij matrix that constitutes the energy feature.

[0252] Additionally or alternatively, Figure 8 As shown, the input scFv sequence 802 is encoded (e.g., one-hot encoded) to obtain sequence features 810.

[0253] Energy feature 806 and sequence feature 810 are converted into fixed-length embeddings of size L x L and L, respectively, where L represents the maximum sequence length in the dataset, such that V H and V L are the maximum lengths of the heavy and light chains, respectively. Input scFv sequences 802 that are less than L are padded with zeros.

[0254] The energy feature 806 and the sequence feature 810 are provided as input to two parallel branches of the model. In particular, the energy feature 806 is provided as input to the 2D convolution layer 808, and the sequence feature 810 is provided as input to the 1D convolution layer 812. For example, the sequence feature and the energy feature can pass through the corresponding convolution layer with batch normalization and ReLU activated.

[0255] like Figure 8 As shown, the sequence input is transformed and concatenated with the energy input. The concatenated matrix is ​​then passed through another 2D convolutional layer 814, flattened, and provided to a dense layer 816 to output the logical value of each category. The class probability 818 can be obtained by performing a normalized exponential function on the logical value.

[0256] In some embodiments, class probability 818 represents the thermal stability (e.g., TS50 measurement or T m The probability that a measured value) corresponds to a specific temperature range.

[0257] To estimate the predicted thermostability (eg, TS50 value or Tm value) value, the probability is weighted by the mean thermostability (eg, mean TS50 or mean Tm) value for each category.

[0258]

[0259] Additionally or alternatively, these techniques can include classifying the scFv into a category corresponding to the highest probability output by the machine learning model. The identified category can correspond to a temperature range that includes at least one temperature at which the scFv is thermally stable.

[0260] In some embodiments, Figure 8 The architecture shown in can be used to predict thermal stability using only one of the energy feature 806 and the sequence feature 810 as input. In this case, the architecture of the model may not be changed. Instead, a zero tensor may be provided as input in place of one of these features. For example, to predict thermal stability using only the energy feature 806, a zero tensor may be provided instead of the sequence feature 810 via Figure 8 Similarly, to predict thermal stability using only sequence features 810, a zero tensor instead of energy features 806 can be passed Figure 8 The lower branch of the architecture is shown.

[0261] Experiments were conducted to evaluate various machine learning models (e.g., Figure 8 The performance of the "energetics only" model in predicting thermal stability according to the data set described in the "Data Sets" section of this document is shown in Figure 1. The "energetics only" model is configured to use only energy features (e.g., energy features 806) as input to predict thermal stability. The "sequence only" model is configured to use only sequence features (e.g., sequence features 810) as input to predict thermal stability. The "energetics + sequence" model is configured to use sequence features and energy features (e.g., energy features 806 and sequence features 810) to predict thermal stability.

[0262] The output of the model is used to assess whether the set of experiments from which the scFv is derived has an impact on prediction accuracy. The representations learned by the sequence-only model and the energetics-only model can be analyzed by projecting the dense layer embedding of each sequence to two dimensions via t-distributed stochastic nearest neighbor embedding (t-SNE). Fig. 9 The t-SNE generated for each model is shown. The embeddings of the sequence-only model are clustered by their experimental set, as Fig. 9The noisy embeddings of energetics demonstrate that the embeddings of the energetics-only model are independent of any clustering based on the experimental set. Thus, despite having a sequence-diverse dataset, fine-tuned and supervised models trained only on sequence features are able to infer the underlying experimental origin of the sequences and skew the thermal stability predictions, making them less generalizable to the newer blind dataset. In contrast, the energetics-only model is more generalizable to the newer blind dataset.

[0263] The performance of the energetics-only model was evaluated by constructing a receiver operating characteristic (ROC) curve for predictions from the above 70°C category. Fig.10 The constructed ROC curve is shown. ROC was evaluated for four test data sets: two reserved data sets (set P and set Q) and two blind data sets representing test antibodies (test Ab) and separated scFvs (separated scFv). The area under ROC exceeded 0.7, indicating high classification accuracy.

[0264] Fig.11 A graph comparing the performance of various machine learning models for predicting the thermal stability of scFvs is shown. In particular, Fig.11 The correlation coefficients for the energy-only model, the sequence-only model, and the energetics + sequence model for all four test datasets are shown separately. On the held-out dataset, the energetics-only model has a coefficient over 0.5, and the energetics + sequence model shows equally improved performance. However, on the blind dataset, the performance of energetics + sequence and sequence-only (coefficients below 0.1) drops. The energetics-only model still shows relatively high correlations for the blind dataset (0.2 and 0.4, respectively).

[0265] As a control, the weights were randomly initialized in the S-CNN for the classification task. The results showed that it could not distinguish between sequences based on thermal stability. Further, on the test set, weighted random predictions were performed, i.e., predicting class labels with weighted random selection, with the sample size of each class as weight. In both tests, only the energetics S-CNN was able to decipher some relationship between the energetics and thermal stability of the scFv. The randomly initialized model was unable to demonstrate any discernible relationship, demonstrating the importance of representations learned from supervised data.

[0266] As described in this article, Figure 8 An example supervised CNN architecture that can be used to predict thermal stability based on sequence features and / or energy features is shown. Figure 8 In the example shown, a 2D-CNN classification model is used for the energy features. However, there may be one or more alternative ways of feeding the energy features, such as a 1D flattened input or a 2D input with absolute residue energy values. These architectures were tested and their performance was compared to inform architecture selection. The results are shown in Fig. 12A For the energetics-only case, the 2D-CNN with categorical input shows improved performance compared to the other two architectures.

[0267] Additionally or alternatively, to reduce the variance of model performance, the performance of multiple models can be integrated by averaging their predictions and generating an ensemble of CNNs. Three supervised CNNs with the same architecture were trained on different dataset splits. The experimental set was shuffled to obtain three different pairs of training and holdout data. Fig. 12B and Fig. 12C The improved performance of predicting thermal stability using an ensemble of CNNs was demonstrated compared to the performance of a non-ensemble CNN.

[0268] For testing purposes, a randomly initialized CNN was used and the embeddings generated by this randomized model were compared to an ensemble of CNNs. As shown, the model developed by the inventors can better separate sequences based on their thermostability properties even with limited sequence size and training diversity.

[0269] The performance of the CNN ensemble was also tested on a blind dataset (i.e., test scFv sequences and isolated scFv sequences) with respect to a weighted random prediction model. For this case, the randomization was biased by the weight of the sample size for each class as observed from the training dataset. The random number generator then used this sample size as a probability to predict the class of the sequence. Although the data points were limited, non-uniform, and severely skewed, FIG. 13A to FIG. 13B It is shown that the thermal stability predictions obtained using the machine learning techniques developed by the inventors can be used to observe the thermal stability trends of scFvs compared to weighted random predictions. In particular, FIG. 13A to FIG. 13B A confusion matrix highlighting the predicted probabilities for each class is shown. The predictions for the top class are more biased towards the higher temperature region in the machine learning predictions compared to the weighted random predictions. This means that when predicting a blind sequence, if the top class (i.e., class 70 and above) is considered, the probability of actually selecting a thermally stable sequence (i.e., in the top 2 classes 60 to 70 or above 70) is higher. This is very important as it can help remove redundant, potentially not too thermally stable sequences, and with a well-curated training set, simple supervised networks can be used to make robust design estimates.

[0270] B. Pre-trained language model

[0271] A pre-trained language model (PTLM) was evaluated for its ability to predict the thermal stability of scFv sequences. Fig.14is a diagram depicting an example technique 1400 for predicting the thermal stability of scFv using PTLM according to some embodiments of the technology described herein. A pre-trained language model 1402 was evaluated for its ability to predict thermal stability using zero-shot learning predictions 1404 and fine-tuning predictions 1406.

[0272] Three pre-trained language models were evaluated. The first model, UniRep, is an mLSTM with 1900 hidden units pre-trained on the Pfam database. Multiple sequence alignments (MSAs) were collected for each sequence in the TS50 set as proposed by A Bateman et al., "The pfam protein families database" Nucleic acids research (2004), which is incorporated herein by reference in its entirety. These sequences were combined into a single dataset, and the model was pre-trained on this set of evolutionarily related sequences using the implementation described in Ma, Eric J. and Arkadij Kummer, "Reimplementing UniRepin JAX" bioRxiv (2020), which is incorporated herein by reference in its entirety.

[0273] In addition, ESM-1b and ESM-1v transformer models were considered. Both are 33-layer, 650M parameter transformer models, pre-trained with masked language modeling on the Uniref database. ESM-1b was trained on a 50% sequence identity filtered dataset (Unired50), while ESM-1v was trained on a 90% sequence identity filtered dataset (Unifer90). A Rives et al., "Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences" Proc. Natl. Acad. Sci. [Proceedings of the National Academy of Sciences of the United States of America] (2021) (which is incorporated herein by reference in its entirety) describes the ESM-1b transformer model. J Meier et al., “Language models enable zero-shot prediction of the effects of mutations on protein function” Adv. Neural Information Processing Systems (2021), which is incorporated herein by reference in its entirety, describes the ESM-1v transformer model.

[0274] i. Zero-Shot Learning Evaluation

[0275] Fig.15A is an illustration of an example pre-trained language model configured for zero-shot learning thermal stability prediction using sequence data according to some embodiments of the techniques described herein.

[0276] One way to predict thermal stability using a pre-trained language model is to use the model likelihood or pseudo-likelihood directly. Sequences that are more likely under the model are predicted to be more thermally stable. UniRep models the probability of each residue in a sequence of residues given all previous residues. Therefore, the likelihood of a sequence can be effectively evaluated as:

[0277]

[0278] ESM-1v models the probability of a masked residue given an unmasked residue. The pseudo-likelihood of the sequence can be obtained as follows:

[0279]

[0280] In practice, both log-likelihood and pseudo-log-likelihood are evaluated to ensure numerical stability. In addition, ESM-1v is an ensemble of five models trained with different random seeds. The predictions from all five models are averaged to obtain the final pseudo-log-likelihood.

[0281] ii. Fine-tuning evaluation

[0282] Fig. 15B is an illustration of an example fine-tuned pre-trained language model configured for thermal stability prediction using sequence data according to some embodiments of the techniques described herein.

[0283] For the UniRep model, the final hidden state is represented as a fixed-length vector of 3900 hidden units together with the average of the previous hidden states. For the ESM-1b model, each per-residue representation is projected down to 4 dimensions and then concatenated. This results in a fixed-length embedding of size 4L, where L is the maximum sequence length in the TS50 dataset. If a sequence has a length less than L, it is padded with zeros.

[0284] These embeddings pass through a linear layer with a hidden dimension of 512, followed by a tanh activation, and then to the final layer to predict the class logits. The parameters of the UniRep model and the ESM-1b model were frozen during training. The parameters of the head model (including the initial downcast of the ESM-1b) were optimized with the Adam optimizer and 10 -3 The learning rate is used for training.

[0285] For the TS50 data, the model was trained on all but one target and evaluated on the held-out target. For the non-TS50 data, an ensemble of TS50 models (one for each held-out target) was used to make predictions.

[0286] iii. Results

[0287] Fig. 15C , Fig.15D , Fig.15E and Fig.15F are graphs showing that fine-tuning predictions achieve improved correlation with thermal stability compared to zero-shot learning predictions according to some embodiments described herein.

[0288] In particular, Fig. 15C and Fig.15E It is shown that zero-shot learning predictions are generally not well correlated with thermal stability, either on the TS50 dataset described in the “Datasets” section or on a blind test set. Fig.15DIt is shown that when trained on the TS50 dataset, fine-tuned predictions from both ESM-1b and UniRep achieve moderate to high average Spearman correlations (0.63 and 0.45) on the retained targets. Fig.15E As shown in Figure 3, these predictions do not generalize well to the blind test set. This suggests that there is some underlying structure to the sequences in the TS50 dataset that the model can exploit to make predictions, but this structure does not generalize to new datasets.

[0289] C. Comparing the performance of machine learning models

[0290] Experiments were conducted to evaluate the ability of the supervised and pre-trained learning models described in this paper to distinguish between thermostable and thermodegenerate mutations.

[0291] Thermal aggregation experiments of point mutations of anti-VEGF antibodies (PDB ID: 2FJG / 2FJF) are detailed in P Koenig et al., "Mutational landscape of antibody variable domains reveal a switch modulating the interdomain confirmational dynamics and antigen binding", Proc. Natl. Acad. Sci. United States Am. (2017) and S Warszawski et al., "Optimizing antibody affinity and stability by the automated design of the variable light-heavy chain interfaces", PLoS Comput. Biol. (2019), each of which is incorporated herein by reference in its entirety. In both studies, deep mutation scanning (DMS) experiments were performed on antibodies, and fragment antigen binding (Fab) melting temperatures (T) of point mutations that improved binding enrichment compared to wild-type were analyzed. m These point mutations (20 mutations compiled from the two studies) were used as test cases to evaluate whether networks trained on TS50 temperature measurements can gain insights into relevant temperature-dependent properties such as thermal aggregation; and whether they distinguish between thermal-enhancing and thermal-hindering mutations.

[0292] To determine whether the predictive model described herein has potential for protein design, a computational DMS was created on an anti-VEGF antibody (PDBID: 2FJG). Each residue position in the sequence was mutated to 19 other amino acids to obtain mutant sequences. Each sequence was one-hot encoded to obtain sequence data, and an energetics dataset (e.g., Figure 2B The sequence and energy inputs were fed into the model, and point mutants classified into temperature categories above 70 predicted by machine learning were cross-validated with experimental results.

[0293] Fig.16 is an illustration according to some embodiments of the technology described herein showing that thermal stability predictions determined using machine learning techniques are consistent with experimentally determined thermal stability. Fig.16 The spheres shown in FIG. 1 indicate improved thermal stability of the scFv (e.g., T m ). Starred balls indicate that the machine learning technique accurately predicted the mutation and residue position that produced the most thermostable scFv. Light grey balls indicate that the machine learning technique accurately predicted the residue position that produced the most thermostable scFv, but did not predict the mutation. Ringed balls indicate mutations that were not observed using the machine learning technique.

[0294] As shown, the CNN was able to correctly identify five of the 20 mutations. Further, for 18 of the 20 mutations, the CNN could correctly identify the residue position despite predicting a different amino acid mutation to be the most thermostable. Of the 4540 point mutations analyzed (N res = 227 residues, 20 amino acids per residue), only experimental data for 20 point mutations are available. Since only 0.44% of the total possible mutations in anti-VEGF antibodies were evaluated for melting temperature in the experiment, there are few validation data sets for thermal stability. Furthermore, although TS50 and T m is a temperature-specific property, but they are different experimental measurements and not perfectly correlated. Therefore, it is noteworthy that the CNN can predict the thermostable residue position in 90% of the cases, with 25% of the predictions being successful (the correct residue and amino acid residue). By extrapolating the network trained on the TS50 measurements to the surrogate thermal aggregation experiment (in this case T m ), demonstrating that such a model can capture the intrinsic thermal properties.

[0295] Furthermore, when comparing residue positions that violate the germline consensus sequence of anti-VEGF Ab, e.g. FIG. 17A to FIG. 17BAs shown, different amino acid mutations were observed, highlighting the ability of the machine learning techniques described in this paper to provide mutations that are orthogonal to traditional germlining methods. With more diverse and larger training datasets, it will be possible to develop more robust models. The results suggest that these networks can serve as a useful tool for screening or filtering antibody sequences for use in temperature-specific antibody design pipelines.

[0296] FIG. 18A to FIG. 18B are graphs illustrating that thermal stability predictions output by a supervised convolutional neural network achieve improved correlation with thermal stability compared to thermal stability predictions made by an unsupervised pre-trained language model according to some embodiments of the techniques described herein.

[0297] Fig.19 A chart comparing the performance of various machine learning models described herein for predicting the thermal stability of scFv according to some embodiments of the technology described herein is shown. As shown, training and prediction using a supervised CNN model are generally more accurate than training and prediction using a pre-trained language model.

[0298] Fig. 20 Included are graphs showing that training and prediction based on residue pair interaction energy metrics are more accurate than training and prediction based on coding sequences and training and prediction based on both coding sequences and residue pair interaction energy metrics according to some embodiments of the technology described herein.

[0299] Fig.21A and Fig.21B are graphs showing that training and prediction based on residue pair interaction energy metrics are more accurate than training and prediction based on residue pair interaction energy metrics and coding sequences according to some embodiments of the technology described herein.

[0300] D. Dataset

[0301] In order to learn temperature-specific contextual patterns in sequence data, a machine learning model was developed and trained for thermal stability prediction using scFv sequences. The temperature data was collected from various antibody engineering studies to develop thermally stable scFv antibodies. The sequence data contained scFv sequences assembled by mutating heavy and light chains from multiple germlines. 2,700 scFv sequences from 17 germlines (further referred to as the experimental set) were collated to form the sequence data. In addition, sequences from another scFv study (currently in trial) and a separate scFv data set formed a blind test set.

[0302] For each sequence, thermal stability was assessed using a TS50 measurement representing the temperature at half maximum for target binding, and this measurement was used as a temperature annotation. TS50 data can also be divided into four categories. For example, TS50 data can be divided into below 50°C, 50°C to 60°C, 60°C to 70°C, and above 70°C.

[0303] The experimental dataset is non-uniform and may be skewed towards higher temperature categories (i.e., category 60°C to 70°C and category above 70°C). The distribution of the training dataset, validation dataset, and test dataset is as follows: Fig. 22 The sequence data is represented in such a way that the higher the bar, the greater the commonality. H (Heavy) chain and V L Gly between (light) chains 4 The / Ser linker region is evident. For machine learning model training, the heavy and light chain sequences were separated from the linker. Fig.23 The temperature distribution of the TS50 measurements is highlighted to show the skewed nature of the experimental dataset.

[0304] i. Generation of scFv

[0305] The scFv with a (G4S)3 linker was cloned as a single construct into a pTT vector with a puromycin selection marker. The construct was transfected into a mammalian CHO-K1 cell line and stably expressed at a scale of 4 mL. 21 days after transfection, VCD and viability were measured, and the expression level of secreted protein in the conditioned medium was analyzed by non-reducing SDS PAGE gel. The cells were further incubated overnight with magnetic beads coupled to proA (for scFv with a λ variable domain) or proL (for scFv with a κ variable domain). The beads were separated from the cell culture medium and then washed three times with PBS and 2 times with water. The scFv was eluted from the magnetic beads with a low pH buffer (100mM glycine, pH2.7) and neutralized with 3M Tris (pH11). Differential scanning fluorescence (DSF) was performed to determine the melting temperature of the purified material. In brief, the molecules were heated at 1.0°C / min on a nanoDSF instrument. Changes in tryptophan fluorescence were monitored to assess protein unfolding and aggregation. T m Reported as the midpoint between the start of expansion and the maximum expanded state.

[0306] ii. TS50 screening assay

[0307] The thermal stability of scFv was screened by determining the loss of target binding after high temperature stress. To this end, soluble scFv (VH-(G4S)3-VL) containing C-terminal FLAG tags (DYKDDDDK) and 6xHis tags were produced in E. coli TG1 (Agilent, Santa Clara, USA) in 10mL LB culture. Protein production was induced with 1mM IPTG. The bacteria were then centrifuged and the cell pellet was resuspended in 1mL Gibco TM In DPBS. Cells were lysed with 4 freeze / thaw cycles and residual cells and cell debris were removed by two centrifugation steps. 100 μL of these crude extracts were transferred to 0.2 mL tubes and subjected to different temperatures (4°C, 50°C, 60°C, 70°C) for 5 min in a water bath. After incubation, the tubes were directly transferred to ice and human target transfected CHO cells were incubated with 50 μL of lysate. The bound scFvs were detected and analyzed by flow cytometry. The median fluorescence intensity values ​​were determined and plotted. The temperature corresponding to the half-maximal binding of each scFv was calculated (TS50). The scFv sequences were further classified into sets based on the identity of the antigen to which the scFv sequences bind (non-random). Since these sets are not curated for machine learning tasks, there is a lack of uniform distribution across sets.

[0308] iii. nanoDSFT m method

[0309] Thermal melting (T) was determined by running the Trp Shift Study on Prometheus, NT.48. m ) temperature. A thermal ramp was applied at 1.0°C / min, with a starting temperature of 25°C and an end temperature of 95°C. Unfolding was measured by the fluorescence ratio 350nm / 330nm. Data analysis and T were performed using PR.ThermControl v2.0.4 m Prior to TM analysis, samples were normalized to 1.0 mg / mL in formulation buffer.

[0310] iv. Dataset Curation for Supervised Models

[0311] To generate sequence input, the datasets of TS50 measurements of scFvs from all experimental sets were aggregated together to form a single dataset. The scFv sequences consisted of a Gly-Ser (G4S) x The dataset was created using scFv sequences, split into their respective heavy and light chain sequences, and instead of classifying the sequences based on their thermal stability, their TS50 measurements were included. Fig.23The distribution of scFv sequences across the experimental and test datasets is shown in Figure 2. Sets P and Q were removed along with the sequences of the test antibodies and the isolated scFvs to form the retained set. The amino acid sequences were one-hot encoded to form a matrix of dimension (V H +V L +3)×21 input, where V H and V L The additional markers of the one-hot encoding of amino acids correspond to the separators at the start and end of the scFv sequence, as well as the separator between the heavy and light chains indicating chain breaks.

[0312] In order to obtain energetics input, first make the sequence pass through the structure module, that is, the DeepAb scheme for protein structure prediction.Rosetta Relax and refinement scheme (XML script in supplement) for side chain repacking are run for each predicted structure.Energy estimation in Rosetta starts with energy relaxation step to reduce steric conflict (RosettaRelax), wherein there is constraint on the starting coordinates so that the accuracy of the backbone structure (predicted by DeepAb) will not be reduced.The all-atom model is further refined with 4 side chain stacking cycles to obtain a robust structure, and the lowest energy structure is selected for further calculation.For each refined model, residue-residue interaction energy metrics are estimated with residue energy decomposition application.Monomer energy and dimer energy are converted to two-dimensional ij matrix, which is used as energy information for supervising CNN model training.

[0313] The energy values ​​in the ij matrix are further classified into 20 categories between the lower and upper energies of |-25,10|REU, respectively. Additional categories are included for start markers, end markers, and chain break markers, respectively. The dimension of the pairwise energy data is L×L×21.

[0314] Fig.24 The following illustrates a method that can be combined with the techniques described herein (e.g., FIG. 2A to FIG. 2B , FIG. 4A to FIG. 4B and Figure 6An illustrative embodiment of a computer system 2400 for use with any embodiment of the method of the present invention. The computer system 2400 includes one or more processors 2410 and one or more articles of manufacture including non-transitory computer-readable storage media (e.g., memory 2420 and one or more non-volatile storage media 2430). The processor 2410 can control writing data to and reading data from the memory 2420 and the non-volatile storage media 2430 in any suitable manner, as the various aspects of the technology described herein are not limited to any particular technology for writing or reading data. In order to perform any functionality described herein, the processor 2410 can execute one or more processor-executable instructions stored in one or more non-transitory computer-readable storage media (e.g., memory 2420), which can be used as a non-transitory computer-readable storage medium to store processor-executable instructions for execution by the processor 2410.

[0315] The computer system 2400 may also include a network input / output (I / O) interface 2440 via which the computing device may communicate with other computing devices (e.g., over a network), and one or more user I / O interfaces 2450 via which the computing device may provide output to a user and receive input from a user. The user I / O interfaces may include devices such as a keyboard, a mouse, a microphone, a display device (e.g., a monitor or a touch screen), a speaker, a camera, and / or various other types of I / O devices.

[0316] The above-described embodiments may be implemented in any of a variety of ways. For example, embodiments may be implemented using hardware, software, or a combination thereof. When implemented in software, the software code may be executed on any suitable processor (e.g., a microprocessor) or processor set, whether provided in a single computing device or distributed among multiple computing devices. It should be understood that any component or component set that performs the above-described functions may be generally considered to be one or more controllers that control the above-described functions. The one or more controllers may be implemented in a variety of ways, such as using dedicated hardware or using general-purpose hardware (e.g., one or more processors) that is programmed using microcode or software to perform the above-described functions.

[0317] In this regard, it should be understood that one implementation of the embodiments described herein includes at least one computer-readable storage medium (e.g., RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disk (DVD) or other optical disk storage, cassette, tape, magnetic disk storage or other magnetic storage device, or other tangible non-transitory computer-readable storage medium), the at least one computer-readable storage medium is encoded with a computer program (i.e., a plurality of executable instructions) that performs the above-mentioned functions of one or more embodiments when executed on one or more processors. The computer-readable medium may be transportable so that the program stored thereon can be loaded onto any computing device to implement various aspects of the technology described herein. In addition, it should be understood that the reference to a computer program that performs any of the above-mentioned functions when executed is not limited to an application program running on a host. Rather, the terms computer program and software are used herein in a general sense to refer to any type of computer code (e.g., application software, firmware, microcode, or any other form of computer instructions) that can be employed to program one or more processors to implement various aspects of the technology described herein.

[0318] The above description of the embodiments provides illustration and description, but is not intended to be exhaustive or limit the embodiments to the precise form disclosed. Modifications and variations are possible in accordance with the above teachings, or may be obtained from the practice of the embodiments. In other embodiments, the methods depicted in these figures may include fewer operations, different operations, operations in different orders, and / or additional operations. Further, non-dependent blocks may be executed in parallel.

[0319] It is apparent that in the embodiments shown in the figures, the example aspects described above may be implemented in many different forms of software, firmware, and hardware. Further, certain portions of these embodiments may be implemented as "modules" that perform one or more functions. The module may include hardware (such as a processor, an application specific integrated circuit (ASIC), or a field programmable gate array (FPGA)) or a combination of hardware and software.

[0320] Having described several aspects and embodiments of the technology set forth in this disclosure, it should be understood that various changes, modifications and improvements will be readily conceived by those skilled in the art. Such changes, modifications and improvements are intended to be within the spirit and scope of the technology described herein. For example, a person of ordinary skill in the art will easily envision various other means and / or structures for performing the functions described herein and / or obtaining results and / or one or more advantages, and each of such changes and / or modifications is considered to be within the scope of the embodiments described herein. Those skilled in the art will recognize or be able to determine many equivalents of the specific embodiments described herein using only routine experiments. Therefore, it should be understood that the aforementioned embodiments are presented only by way of example, and within the scope of the appended claims and their equivalents, the inventive embodiments may be practiced in a manner different from that specifically described. In addition, any combination of two or more features, systems, articles, materials, kits and / or methods described herein, if such features, systems, articles, materials, kits and / or methods are not mutually inconsistent, is included within the scope of this disclosure.

[0321] The above embodiments can be implemented in any of a variety of ways. One or more aspects and embodiments of the present disclosure related to the execution of a process or method can utilize program instructions that can be executed by a device (e.g., a computer, a processor, or other device) to execute these processes or methods or control their execution. In this regard, different inventive concepts can be embodied as a computer-readable storage medium (or multiple computer-readable storage media) (e.g., a computer memory, one or more floppy disks, compact disks, optical disks, tapes, flash memories, circuit configurations in field programmable gate arrays or other semiconductor devices, or other tangible computer storage media), which is encoded with one or more programs, which execute one or more methods of implementing the various embodiments described above when executed on one or more computers or other processors. The one or more computer-readable media can be transportable, so that one or more programs stored thereon can be loaded onto one or more different computers or other processors to implement various aspects of the above aspects. In some embodiments, the computer-readable medium can be a non-transient medium.

[0322] The term "program" or "software" is used herein in a general sense to refer to any type of computer code or computer executable instruction set that can be employed to program a computer or other processor to implement the various aspects described above. In addition, it should be understood that, according to one aspect, one or more computer programs that perform the methods of the present disclosure when executed need not reside on a single computer or processor, but can be distributed in a modular manner among multiple different computers or processors to implement various aspects of the present disclosure.

[0323] Computer executable instructions can be in many forms, such as program modules, that are executed by one or more computers or other devices. Typically, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. Typically, in various embodiments, the functionality of the program modules can be combined or distributed as desired.

[0324] Moreover, the data structure can be stored in a computer-readable medium in any suitable form. For simplicity of presentation, the data structure can be shown as having fields that are related by location in the data structure. Similarly, such relationships can be achieved by assigning locations in a computer-readable medium that convey the relationship between the fields for storage of the fields. However, any suitable mechanism can be used to establish the relationship between the information in the fields of the data structure, including by using pointers, tags, or other mechanisms that establish relationships between data elements.

[0325] When implemented in software, the software code can be executed on any suitable processor or collection of processors, whether provided in a single computer or distributed among multiple computers.

[0326] Moreover, the computer may have one or more input and output devices. Among other things, these devices may be used to present a user interface. Examples of output devices that may be used to provide a user interface include a printer or display screen for visually presenting the output and a speaker or other sound generating device for auditorily presenting the output. Examples of input devices that may be used for a user interface include a keyboard and a pointing device, such as a mouse, a touch pad, and a digitized tablet computer. As another example, the computer may receive input information by voice recognition or in other audible formats.

[0327] Such computers can be interconnected by one or more networks of any suitable form, including local area networks or wide area networks (such as enterprise networks) and intelligent networks (IN) or the Internet. Such networks can be based on any suitable technology and can operate according to any suitable protocol, and can include wireless networks, wired networks or fiber optic networks.

[0328] Also, as described, some aspects may be embodied as one or more methods. The actions performed as part of a method may be ordered in any suitable manner. Thus, embodiments may be constructed in which the actions are performed in an order different from that shown, which may include performing some actions simultaneously (even though the actions are shown as being sequential in the illustrative embodiments).

[0329] All definitions, as defined and used herein, should be understood to control over dictionary definitions, definitions in documents incorporated by reference, and / or ordinary meanings of the defined terms.

[0330] The indefinite articles "a" and "an" as used herein in the specification and claims, unless expressly indicated to the contrary, should be understood to mean "at least one".

[0331] The phrase "and / or" as used herein in the specification and claims should be understood to mean "either or both" of the elements so combined, i.e., elements that are present in combination in some cases and separately in other cases. Multiple elements listed with "and / or" should be interpreted in the same manner, i.e., "one or more" of the elements so combined. Other elements other than the elements specifically identified by the "and / or" clause may optionally be present, whether related or unrelated to those elements specifically identified. Thus, as a non-limiting example, when used in conjunction with open language such as "comprising", a reference to "A and / or B" may refer to only A (optionally including elements other than B) in one embodiment; to only B (optionally including elements other than A) in another embodiment; to both A and B (optionally including other elements) in yet another embodiment; etc.

[0332] As used herein in the specification and claims, when referring to a list of one or more elements, the phrase "at least one" should be understood to mean at least one element selected from any one or more of the elements in the list of elements, but not necessarily including at least one of each element specifically listed within the list of elements, and not excluding any combination of elements in the list of elements. This definition also allows that elements other than the elements specifically identified within the list of elements to which the phrase "at least one" refers may optionally be present, whether related to or unrelated to those elements specifically identified. Thus, as a non-limiting example, "at least one of A and B" (or equivalently "at least one of A or B", or equivalently "at least one of A and / or B") may refer to at least one (optionally including more than one) A in one embodiment, where B is absent (and optionally including elements other than B); to at least one (optionally including more than one) B in another embodiment, where A is absent (and optionally including elements other than A); to at least one (optionally including more than one) A and at least one (optionally including more than one) B (and optionally including other elements) in yet another embodiment; etc.

[0333] In the claims and the above specification, all transitional phrases, such as "comprising", "including", "carrying", "having", "containing", "involving", "holding", "composed of", etc., should be understood as open-ended, that is, meaning including but not limited to. Only the transitional phrases "consisting of" and "consisting essentially of" should be closed or semi-closed transitional phrases, respectively.

[0334] The terms "approximately," "substantially," and "about" may be used in some embodiments to mean within ±20% of a target value, in some embodiments to mean within ±10% of a target value, in some embodiments to mean within ±5% of a target value, and in some embodiments to mean within ±2% of a target value. The terms "approximately," "substantially," and "approximately" may include the target value.

Claims

1. A method for computationally screening a set of single-chain variable fragments (scFvs) based on the thermal stability of the scFvs predicted by a trained machine learning model, the set of scFvs comprising scFvs having different residue sequences, the method comprising: determining a thermal stability indicator for each scFv in the panel of scFvs using the trained machine learning model and at least one computer hardware processor to obtain a plurality of thermal stability indicators, the panel of scFvs including a first scFv having a first residue sequence, the determining comprising: obtaining an interaction energy metric for each of a plurality of pairs of residues in the first residue sequence using information indicative of the three-dimensional (3D) structure of the first scFv; generating a first feature set to provide as input to the trained machine learning model, the generating comprising including the interaction energy metrics in the first feature set; and providing the first feature set as input to the trained machine learning model to obtain a corresponding output indicative of a first thermal stability of the first scFv; identifying a subset of the panel of scFvs for subsequent production based on the plurality of thermal stability indicators; and At least one scFv among the scFvs in the identified subset is produced.

2. The method of claim 1, wherein: The panel of scFvs further comprises a second scFv different from the first scFv, the second scFv having a second residue sequence, and wherein determining the thermal stability indication of each scFv in the panel of scFvs further comprises: obtaining a second interaction energy metric for each pair of the second plurality of pairs of second residues, These second residues are in this second residue sequence; generating a second feature set to provide as input to the trained machine learning model, the generating comprising including the second interaction energy metrics in the second feature set; and The second feature set is provided as input to the trained machine learning model to obtain a corresponding output indicative of a second thermal stability of the second scFv.

3. The method of claim 1 or any other preceding claim, wherein: The output indicative of a first thermal stability of the first scFv indicates a first temperature at which the first scFv is thermally stable.

4. A method as claimed in claim 3 or any other preceding claim, wherein: The first temperature is an estimate of the temperature corresponding to half-maximal binding of the first scFv.

5. The method of claim 1 or any other preceding claim, wherein: The output indicative of a first thermal stability of the first scFv is indicative of a first temperature range including at least one temperature at which the first scFv is thermally stable.

6. A method as claimed in claim 5 or any other preceding claim, wherein: The first temperature range is an estimate of the temperature range that includes the temperature corresponding to half-maximal binding of the first scFv.

7. The method of claim 1 or any other preceding claim, wherein: Providing the first feature set as input to the trained machine learning model to obtain the output indicative of a first thermal stability of the first scFv comprises: Using the trained machine learning model, the first scFv is classified into one of a plurality of categories using the first feature set, wherein each of the plurality of categories corresponds to a respective temperature range.

8. The method of claim 1 or any other preceding claim, wherein: Obtaining these interaction energy measures includes: Information indicative of the 3D structure of the first scFv is determined by generating information indicative of the 3D structure from the first residue sequence using protein structure prediction software.

9. The method of claim 8 or any other preceding claim, wherein: Obtaining these interaction energy measures includes: The interaction energy metrics are determined using molecular modeling software to generate the interaction energy metrics using information indicative of the 3D structure of the first scFv.

10. The method of claim 1 or any other preceding claim, wherein: Generating the first feature set includes: For each specific one of these interaction energy metrics, generating a corresponding two-dimensional (2D) matrix of values ​​of the particular energy metric, wherein rows and columns of the 2D matrix correspond to corresponding residues in the first residue sequence, and wherein entries in the i-th row and j-th column of the 2D matrix correspond to values ​​of the particular energy metric for the i-th residue in the first residue sequence and the j-th residue in the first residue sequence; and The generated 2D matrix is ​​included in the first feature set.

11. The method of claim 10 or any other preceding claim, wherein: The generated 2D matrix includes rows for at least 75% of the residues in the first residue sequence.

12. The method of claim 11 or any other preceding claim, wherein: The generated 2D matrix includes rows for at least 90% of the residues in the first residue sequence.

13. The method of claim 12 or any other preceding claim, wherein: The generated 2D matrix includes rows for at least 95% of the residues in the first residue sequence.

14. The method of claim 13 or any other preceding claim, wherein: The generated 2D matrix includes rows for at least 99% of the residues in the first residue sequence.

15. The method of claim 14 or any other preceding claim, wherein: The generated 2D matrix includes a row for each residue in the first residue sequence.

16. The method of claim 1 or any other preceding claim, wherein: Generating the first feature set further includes: encoding the first residue sequence to obtain a coding sequence; and The encoded sequence is included in the first feature set.

17. The method of claim 16 or any other preceding claim, wherein: Encoding the first residue sequence includes: The first residue sequence is one-hot encoded to obtain the encoding sequence, which includes the one-hot encoded version of the first residue sequence.

18. The method of claim 1 or any other preceding claim, wherein: The trained machine learning model includes a trained neural network model.

19. The method of claim 18 or any other preceding claim, wherein: The trained neural network model includes a trained convolutional neural network (CNN) model having multiple 2D convolutional layers.

20. The method of claim 19 or any other preceding claim, wherein: The trained CNN model further includes a fully connected layer.

21. The method of claim 19 or any other preceding claim, wherein: The trained CNN model is configured to output a plurality of probabilities that the scFv is thermally stable within each of a plurality of temperature ranges.

22. The method of claim 21 or any other preceding claim, wherein: Providing the first set of features as input to the trained machine learning model to obtain the corresponding output indicative of a first thermal stability of the first scFv comprises: providing the first feature set to the trained CNN model to obtain a first plurality of probabilities that the first scFv is thermostable within each of the plurality of temperature ranges; and The first thermal stability is determined as: (i) a temperature range in the plurality of temperature ranges that is associated with a highest probability in the first plurality of probabilities; or (ii) a temperature determined as a weighted linear combination of averages of the plurality of temperature ranges weighted by the probabilities in the first set of probabilities.

23. The method of claim 1 or any other preceding claim, wherein: Identifying a subset of the panel of scFvs for subsequent production based on the determined thermal stability indicators comprises: determining whether a first thermal stability of the first scFv satisfies at least one criterion; and After determining that the first thermal stability satisfies the at least one criterion, the first scFv is identified for subsequent production.

24. The method of claim 1 or any other preceding claim, further comprising: The thermal stability of at least one of the scFvs is tested in an in vitro assay.

25. A method for predicting the thermal stability of a single-chain variable fragment (scFv) using a trained machine learning model, the method comprising: determining a first thermal stability indicator for a first scFv using the trained machine learning model and at least one computer hardware processor, the first scFv having a first residue sequence: obtaining an interaction energy metric for each of a plurality of pairs of residues in the first residue sequence using information indicative of the three-dimensional (3D) structure of the first scFv; generating a first feature set to provide as input to the trained machine learning model, the generating comprising including the interaction energy metrics in the first feature set; as well as The first feature set is provided as input to the trained machine learning model to obtain a corresponding output indicative of a first thermal stability of the first scFv.

26. The method of claim 25, further comprising determining a thermal stability indicator for each scFv in a set of scFvs including the first scFv using the trained machine learning model and the at least one computer hardware processor to obtain a plurality of thermal stability indicators.

27. The method of claim 26 or any other preceding claim, further comprising identifying a subset of the panel of scFvs for subsequent production based on the plurality of thermal stability indicators.

28. A system comprising: at least one computer hardware processor; as well as At least one non-transitory computer-readable storage medium storing processor-executable instructions that, when executed by the at least one computer hardware processor, cause the at least one computer hardware processor to perform a method for predicting thermal stability of a single-chain variable fragment (scFv) using a trained machine learning model, the method comprising: determining a first thermal stability indicator for a first scFv using the trained machine learning model and at least one computer hardware processor, the first scFv having a first residue sequence: obtaining an interaction energy metric for each of a plurality of pairs of residues in the first residue sequence using information indicative of the three-dimensional (3D) structure of the first scFv; generating a first feature set to provide as input to the trained machine learning model, the generating comprising including the interaction energy metrics in the first feature set; and The first feature set is provided as input to the trained machine learning model to obtain a corresponding output indicative of a first thermal stability of the first scFv.

29. At least one non-transitory computer-readable storage medium storing processor-executable instructions that, when executed by at least one computer hardware processor, cause the at least one computer hardware processor to perform a method for predicting thermal stability of a single-chain variable fragment (scFv) using a trained machine learning model, the method comprising: determining a first thermal stability indicator for a first scFv using the trained machine learning model and at least one computer hardware processor, the first scFv having a first residue sequence: obtaining an interaction energy metric for each of a plurality of pairs of residues in the first residue sequence using information indicative of the three-dimensional (3D) structure of the first scFv; generating a first feature set to provide as input to the trained machine learning model, the generating comprising including the interaction energy metrics in the first feature set; as well as The first feature set is provided as input to the trained machine learning model to obtain a corresponding output indicative of a first thermal stability of the first scFv.

30. A method for computationally screening a panel of monoclonal antibodies (mAbs) based on thermal stability of the panel of mAbs predicted by a trained machine learning model, the panel of mAbs comprising mAbs having different residue sequences, the method comprising: determining a thermal stability indicator for each mAb in the panel of mAbs using the trained machine learning model and at least one computer hardware processor to obtain a plurality of thermal stability indicators, the panel of mAbs including a first mAb having a first residue sequence, the determining comprising: obtaining an interaction energy metric for each of a plurality of pairs of residues in the first residue sequence using information indicative of the three-dimensional (3D) structure of the first mAb; generating a first feature set to provide as input to the trained machine learning model, the generating comprising including the interaction energy metrics in the first feature set; and providing the first feature set as input to the trained machine learning model to obtain a corresponding output indicative of a first thermal stability of the first mAb; identifying a subset of the panel of mAbs for subsequent production based on the plurality of thermal stability indicators; and At least one mAb among the mAbs in the identified subset is produced.

31. The method of claim 30, further comprising: The thermal stability of at least one of the mAbs was tested in an in vitro assay.

Images