Predicting antibody pharmacokinetic data using a gaussian process regression model

A Gaussian process regression model effectively predicts antibody pharmacokinetics using assay measurements, addressing data scarcity and complexity issues in conventional methods, enabling precise pharmacokinetic data generation.

WO2026235715A1PCT designated stage Publication Date: 2026-11-12GENZYME CORP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/US2025/037718
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-09
Filing Date
2025-07-15
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Conventional machine learning models require large amounts of training data to accurately predict pharmacokinetic data for antibodies, while mechanistic models fail to capture the real-world complexity of antibody pharmacokinetics due to simplifying assumptions, making both approaches inadequate for reliable prediction.

Method used

A Gaussian process regression model is used to predict pharmacokinetic data for antibodies based on assay measurements, leveraging a kernel function to model relationships between antibodies and incorporating uncertainty values, even with limited training data.

Benefits of technology

The system provides accurate and reliable predictions of antibody pharmacokinetics by modeling complex relationships with scarce data, overcoming limitations of conventional machine learning and mechanistic models.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure US2025037718_12112026_PF_FP_ABST
    Figure US2025037718_12112026_PF_FP_ABST
Patent Text Reader

Abstract

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating predicted pharmacokinetic data characterizing a test antibody. In one aspect, a method comprises: obtaining a plurality of assay measurements for a test antibody; processing a model input comprising the plurality of assay measurements for the test antibody using a Gaussian process regression model to generate, as a model output of the Gaussian process regression model, predicted pharmacokinetic data characterizing the test antibody, wherein the Gaussian process regression model generates the predicted pharmacokinetic data based at least in part on a kernel function evaluated between: (i) the plurality of assay measurements for the test antibody, and (ii) a respective plurality of assay measurements for each of a plurality of training antibodies; and outputting the predicted pharmacokinetic data characterizing the test antibody.
Need to check novelty before this filing date? Find Prior Art

Description

Attorney Docket No. 46567-1739WO1PREDICTING ANTIBODY PHARMACOKINETIC DATA USING A GAUSSIAN PROCESS REGRESSION MODEECROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of priority to Greek Patent Application No.20250100347, filed May 9, 2025, the contents of which are herein incorporated by reference.BACKGROUND

[0002] This specification relates to predicting antibody pharmacokinetic data using a Gaussian process regression model.

[0003] Antibodies are proteins produced by the immune system that specifically recognize and bind to antigens, such as pathogens or abnormal cells, through highly variable antigenbinding regions. Structurally, antibodies are composed of two heavy chains and two light chains forming a Y-shaped molecule, with the antigen-binding sites located at the tips of the arms. Monoclonal antibodies, which are identical antibodies derived from a single B-cell clone, have become valuable tools in medicine and biotechnology due to their uniform specificity and high affinity for particular target antigens.

[0004] Pharmacokinetic (PK) data can characterize how a molecule is absorbed, distributed, metabolized, or excreted in a biological system over time. Certain PK parameters provide quantitative measures of the behavior of a molecule in the body. A PK profile refers to the time course of the concentration of a molecule in the bloodstream or other biological compartments following administration and can be represented as a concentration-versus-time curve.SUMMARY

[0005] This specification generally describes a system implemented as computer programs on one or more computers in one or more locations that can generate predicted pharmacokinetic data characterizing an antibody using a Gaussian process regression model. As an example, the predicted pharmacokinetic data can include pharmacokinetic (PK) parameters, or a pharmacokinetic (PK) profile.

[0006] In this specification, “PK parameters” are global summan' statistics characterizing properties of a chemical compound, e.g., characterizing one or more of absorption, distribution, metabolism, or elimination of the chemical compound in a subject. For example, a subject can be rat, dog, pig, human, or another species. A “PK” profile includes data thatAttorney Docket No. 46567-1739WO1defines, for each time point in a sequence of time points, the concentration value of the chemical compound in a subject at the time point. In some cases, the PK profile for a particular chemical compound is different for different species.

[0007] According to a first aspect, there is a provided a method for obtaining a plurality of assay measurements for a test antibody, processing a model input comprising the plurality of assay measurements for the test antibody using a Gaussian process regression model to generate, as a model output of the Gaussian process regression model, predicted pharmacokinetic data characterizing the test antibody, wherein the Gaussian process regression model generates the predicted pharmacokinetic data based at least in part on a kernel function evaluated between: (i) the plurality' of assay measurements for the test antibody, and (ii) a respective plurality of assay measurements for each of a plurality of training antibodies, and outputting the predicted pharmacokinetic data characterizing the test antibody.

[0008] In some implementations, the model output of the Gaussian process regression model further comprises an uncertainty value defining an uncertainty’ in the predicted pharmacokinetic data charactenzing the test antibody.

[0009] In some implementations, the kernel function of the Gaussian process regression model comprises a plurality' of hyperparameters that have been numerically optimized by performing operations comprising, at each of a plurality of optimization iterations:computing gradients of a likelihood of the pharmacokinetic data of the plurality’ of training antibodies under the Gaussian process regression model implemented using current values of the plurality7of hyperparameters; andadjusting the current values of the plurality' of hyperparameters using the gradients.

[0010] In some implementations, the likelihood is a log marginal likelihood.

[0011] In some implementations, the kernel function comprises a respective length scale hyperparameter for each of the plurality7of assay measurements that controls a sensitivity of the predicted pharmacokinetic data generated by the Gaussian process regression model on the assay measurement; andwherein the lengths scale hyperparameters for the plurality of assay measurements have been optimized by gradient-based optimization.

[0012] In some implementations, the kernel function comprises a mixture of a squared exponential kernel and a linear kernel; orwherein the kernel function is parametrized by a neural network.Attorney Docket No. 46567-1739WO1

[0013] In some implementations, the Gaussian process regression model comprises a noise level parameter that is empirically calibrated based on a set of validation antibodies that are not used for configuring the Gaussian process regression model.

[0014] In some implementations, empirically calibrating the noise level parameter based on the set of validation antibodies comprises:instantiating, for each of a plurality of noise level parameter values, a respective Gaussian process regression model that comprises a noise level parameter having the noise level parameter value and that is configured using the plurality of training antibodies;evaluating, for each of the plurality of noise level parameter values, a prediction accuracy of the Gaussian process regression model that comprises the noise level parameter having the noise level parameter value on the set of validation antibodies; andselecting the noise level parameter value based on the prediction accuracies of the Gaussian process regression models comprising noise level parameters having the noise level parameter values.

[0015] In some implementations, selecting the noise level parameter value based on the prediction accuracies of the Gaussian process regression models comprising noise level parameters having the noise level parameter values comprises:selecting a noise level parameter value associated with a highest prediction accuracy.

[0016] In some implementations, a number of training antibodies used for configuring the Gaussian process regression model is less than 1000.

[0017] In some implementations, the number of training antibodies used for configuring the Gaussian process regression model is less than 100.

[0018] In some implementations, the predicted pharmacokinetic data characterizing the test antibody comprises one or more predicted pharmacokinetic parameters of the test antibody.

[0019] In some implementations, the one or more predicted pharmacokinetic parameters of the test antibody comprise one or more of:a Cmax parameter defining a maximum test antibody concentration in plasma; or a Tmax parameter defining a time at which maximum test antibody concentration in plasma occurs; oran AUC parameter defining total exposure to the test antibody over time; or a half time parameter that defines a duration of time for the test antibody concentration in plasma to decrease by half; ora clearance parameter that defines a rate at which the test antibody is removed from a subject.Attorney Docket No. 46567-1739WO1

[0020] In some implementations, the model input to the Gaussian process regression model additionally comprises a time value; andwherein the predicted pharmacokinetic data characterizing the test antibody comprises data defining a predicted pharmacokinetic profile of the test antibody evaluated at the time value.

[0021] In some implementations, the plurality of assay measurements for the test antibody comprise one or more of:an antibody cellular recycling FcRn assay measurement;a baculovirus particle assay measurement;a retention time on cross-interaction chromatography assay measurement; or an FcRn column assay measurement.

[0022] In some implementations, outputting the predicted pharmacokinetic data characterizing the test antibody comprises: presenting the predicted pharmacokinetic data on a display of a user device; or transmitting the predicted pharmacokinetic data over a data communications network; or storing the predicted pharmacokinetic data in a memory’.

[0023] In some implementations, one or more of the plurality of assay measurements for the test antibody are determined by performing physical assays on the test antibody.

[0024] In some implementations, one or more of the plurality' of assay measurements for the test antibody are synthetic assay measurements generated by performing operations comprising:processing data characterizing an amino acid sequence of the antibody using an assay prediction machine learning model and in accordance with trained values of a set of assay prediction machine learning model parameters to generate the synthetic assay measurement as a model output of the assay prediction machine learning model.

[0025] In some implementations, the data characterizing the amino acid sequence of the antibody comprises an embedding of the amino acid sequence of the antibody that is generated by a protein language model.

[0026] In some implementations, the assay prediction machine learning model has been trained by a machine learning training technique on a set of training examples that each comprise: (i) a training input that characterizes an amino acid sequence of a training antibody, and (ii) a target output that defines one or more experimental assay measurements for the training antibody.Attorney Docket No. 46567-1739WO1

[0027] In some implementations, the method further comprises estimating a therapeutic dose of the test antibody based at least in part on the predicted pharmacokinetic data for the test antibody.

[0028] In some implementations, the predicted pharmacokinetic data for the test antibody comprises a predicted pharmacokinetic profde for the test antibody; andwherein the method further comprises:determining a total exposure of the test antibody over time from the predicted pharmacokinetic profile for the test antibody; andcomparing the total exposure to an efficacy threshold to predict efficacy of the test antibody, or comparing the total exposure to a toxicity threshold to predict toxicity of the test antibody.

[0029] In some implementations, the test antibody comprises a nanobody.

[0030] In some implementations, the model input to the Gaussian process regression model further comprises an output of a mechanistic pharmacokinetic model.

[0031] In some implementations, the method further comprises selecting the test antibody for physical synthesis based at least in part on the predicted pharmacokinetic data characterizing the test antibody.

[0032] In some implementations, the method further comprises physically synthesizing the test antibody.|0033| In another aspect, there is provided a system comprising one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform the method of any one of the example implementation methods described.

[0034] In another aspect, there is provided a computer storage medium encoded with a computer program, the program comprising instructions that are operable, when executed by data processing apparatus, to cause the data processing apparatus to perform the method of any one of the example implementation methods described.

[0035] Particular embodiments of the subject matter described in this specification can be implemented so as to realize one or more of the following advantages.

[0036] Pharmacokinetic data are important for understanding the behavior of molecules in vivo and for guiding the selection and optimization of drug candidates. However, obtaining pharmacokinetic data experimentally can be time-consuming, expensive, and often infeasible in drug development workflows that involve screening large numbers of candidate molecules. One approach to efficiently obtain pharmacokinetic data for candidate molecules is to train aAttorney Docket No. 46567-1739WO1machine learning model to process data characterizing an input molecule to generate predicted pharmacokinetic data for the input molecule. After training such a machine learning model can be used to screen large numbers of candidate molecules based on their predicted pharmacokinetic properties.

[0037] Training a conventional machine learning model to perform a pharmacokinetic prediction task can require a large amount of training data, in particular, to allow the machine learning model to leam to generalize over diverse regions of chemical space and to reduce the likelihood of overfitting. For certain types of molecules, e.g., small molecules, pharmacokinetic training data may be relatively plentiful and can be used for training machine learning models to perform pharmacokinetic prediction.

[0038] However, pharmacokinetic training data for antibodies tends be highly scarce, e.g.. as a result of such data being particularly expensive and cumbersome to obtain experimentally. For instance, the number of pharmacokinetic training examples available for antibodies may be less than 1000, or possibly even less than 100, which is insufficient to train most conventional machine learning models. Therefore, conventional machine learning approaches may be inadequate for predicting pharmacokinetic data for antibodies.

[0039] Another approach to generating pharmacokinetic data for antibodies is to use mechanistic models, such as compartmental models and physiologically based pharmacokinetic (PBPK) models. Mechanistic models can simulate absorption, distribution, metabolism, and excretion properties by solving systems of equations that represent biological processes. However, mechanistic models make simplifying assumptions about complex biological systems that often fail to capture the real world complexity of such systems, and as a result, their accuracy can be limited, particularly for novel molecules (e.g., antibodies). Therefore, mechanistic models may be inadequate for predicting pharmacokinetic data for molecules in general, and in particular for antibodies.

[0040] The system described in this specification addresses these issues to provide an innovative solution for accurately predicting pharmacokinetic data for antibodies. In particular, the system can predict pharmacokinetic data for antibodies using a Gaussian process regression model, which is a is a nonparametric, probabilistic model that can generate predicted pharmacokinetic data for a “tesf ’ antibody based at least in part on a kernel function evaluated between: (i) a plurality of assay measurements for the test antibody, and (ii) a respective plurality of assay measurements for each of a plurality of “training” antibodies. Gaussian process regression models can perform well with relatively small training datasets because they explicitly model the relationships between antibodies using aAttorney Docket No. 46567-1739WO1kernel function that encodes how pharmacokinetic data varies across the space of antibodies. This allows Gaussian process regression models to make coherent, data-efficient predictions by leveraging the structure inferred from training data. In contrast, many conventional machine learning models leam predictive rules through flexible function approximators that require large amounts of data to capture underlying patterns and achieve reliable generalization.

[0041] The system thus provides a technical improvement to the field of computational modeling of antibody pharmacokinetics by allowing reliable, precise prediction of antibody pharmacokinetics using a Gaussian process regression model. Modeling antibody pharmacokinetics using conventional machine learning approaches may be infeasible because of the highly limited amount of pharmacokinetic training data available for training such machine learning models. Instead of using a conventional machine learning model, the system models antibody pharmacokinetics using a Gaussian process regression model that can robustly generalize from limited and noisy pharmacokinetic training data for antibodies. Modeling antibody pharmacokinetics using mechanistic models may be infeasible because the strong simplifying biological assumptions encoded in mechanistic models may fail to capture the real-world complexity of pharmacokinetics. Instead of using a mechanistic model, the system models antibody kinetics using a Gaussian process regression model that is a nonparametric that does not rely on the simplifying biological assumptions of mechanistic models. The system thus provides a technical solution to technical issues (e.g., limited training data and inability to model real-world complexity ) that exist in alternative approaches to modeling antibody pharmacokinetics.

[0042] The system described in this specification can implement a Gaussian process regression model that is configured to generate predicted pharmacokinetic data for a test antibody by processing a model input that includes a set of assay measurements for the test antibody. Rather than obtaining the assay measurements for the test antibody7through performing physical assays, the system can leverage an assay prediction machine learning model to predict the assay measurement values for the test antibody. The system can train the assay prediction machine learning model to process data characterizing a test antibody (e.g.. a protein language modeling embedding of the test antibody) to generate predicted assay7measurement values for the test antibody. Training the assay prediction machine learning model in this manner can be feasible because training data characterizing assay measurement values for antibodies may be more plentiful, at least in comparison to training data characterizing pharmacokinetic data for antibodies. The system can thus provide a technicalAttorney Docket No. 46567-1739WO1improvement in pharmacokinetic modeling of antibodies by exploiting asymmetries in training data availability, e.g., with assay measurement training data possibly being more plentiful than pharmacokinetic training data for antibodies.

[0043] The details of one or more embodiments of the subject matter of this specification are set forth in the accompanying drawings and the description below. Other features, aspects, and advantages of the subject matter will become apparent from the description, the drawings, and the claims.BRIEF DESCRIPTION OF THE DRAWINGS

[0044] FIG. 1 is a system diagram of an example pharmacokinetic prediction inference system that includes a Gaussian process regression model.

[0045] FIG. 2 is a system diagram of an example pharmacokinetic prediction training system that can train a Gaussian process regression model for generating predicted pharmacokinetic data.

[0046] FIG. 3 depicts example PK profile prediction results of an example pharmacokinetic prediction inference system.

[0047] FIG. 4 is a flow diagram of an example process for generating predicted pharmacokinetic data characterizing an antibody using a Gaussian process regression model.

[0048] FIG. 5 is a flow diagram of an example process for training a Gaussian process regression model to generate predicted pharmacokinetic data.

[0049] Like reference numbers and designations in the various drawings indicate like elements.DETAILED DESCRIPTION

[0050] FIG. 1 shows an example pharmacokinetic prediction inference system 100. The pharmacokinetic prediction inference system 100 is an example of a system implemented as computer programs on one or more computers in one or more locations in which the systems, components, and techniques described below are implemented.

[0051] The pharmacokinetic (PK) prediction inference system 100 can process a model input 110 including test antibody assay measurements 112 to generate a model output 160 including predicted pharmacokinetic data 150, e.g., a pharmacokinetic (PK) parameter 152 characterizing the properties of the test antibody in vitro, or a pharmacokinetic (PK) profile 154 characterizing the concentration of the test antibody over time, using a Gaussian process regression model 140. A Gaussian process regression model can be used to model a complex function and theAttorney Docket No. 46567-1739WO1uncertainty of observations of the function using relatively scarce data, as is described in more detail below.

[0052] For example, the test antibody can be a nanobody, e.g., a stable immunoglobulin single variable domain that is capable of binding antigens without a light chain. The term “immunoglobulin single variable domain"’ (ISV), interchangeably used with “single variable domain”, defines immunoglobulin molecules wherein the antigen binding site is present on, and formed by, a single immunoglobulin domain. This sets immunoglobulin single variable domains apart from “conventional” immunoglobulins (e.g. monoclonal antibodies) or their fragments (such as Fab, Fab", F(ab’)2, scFv, di-scFv), wherein two immunoglobulin domains, in particular two variable domains, interact to form an antigen binding site. Typically, in conventional immunoglobulins, a heavy chain variable domain (VH) and alight chain variable domain (VL) interact to form an antigen binding site. In this case, the complementarity determining regions (CDRs) of both VH and VL will contribute to the antigen binding site, i.e. a total of 6 CDRs will be involved in antigen binding site formation.

[0053] In contrast, immunoglobulin single variable domains are capable of specifically binding to an epitope of the antigen without pairing with an additional immunoglobulin variable domain. The binding site of an immunoglobulin single variable domain is formed by a single VH, a single VHH or single VL domain. Hence, the antigen binding site of an immunoglobulin single variable domain is formed by no more than three CDRs.|0054| As such, the single variable domain may be a light chain variable domain sequence (e.g., a VL-sequence) or a suitable fragment thereof; or a heavy chain variable domain sequence (e.g., a VH-sequence or VHH sequence) or a suitable fragment thereof; as long as it is capable of forming a single antigen binding unit (i.e., a functional antigen binding unit that essentially consists of the single variable domain, such that the single antigen binding domain does not need to interact with another variable domain to form a functional antigen binding unit).

[0055] An immunoglobulin single variable domain (ISV) can for example be a heavy chain ISV, such as a VH, VHH, including a camelized VH or humanized VHH. In one embodiment, it is a VHH, including a camelized VH or humanized VHH. Heavy chain ISVs can be derived from a conventional four-chain antibody or from a heavy chain antibody.

[0056] For example, the immunoglobulin single variable domain may be a (single) domain antibody (or an amino acid sequence that is suitable for use as a single domain antibody), a "dAb" or dAb (or an amino acid sequence that is suitable for use as a dAb) or a Nanobody® ISV (as defined herein and including but not limited to a VHH); other single variable domains, or any suitable fragment of any one thereof.Attorney Docket No. 46567-1739WO1

[0057] In particular, the immunoglobulin single variable domain may be a Nanobody® ISV (such as a VHH, including a humanized VHH or camelized VH) or a suitable fragment thereof.[Note: Nanobody® is a registered trademark of Ablynx N.V.]

[0058] “VHH domains”, also known as VHHs, VHH antibody fragments and VHH immunoglobulins, have originally been described as the antigen binding immunoglobulin variable domain of “heavy chain antibodies” (i.e., of “antibodies devoid of light chains”; Hamers-Casterman et al. 1993 (Nature 363: 446-448). The term “VHH domain” has been chosen in order to distinguish these variable domains from the heavy chain variable domains that are present in conventional 4-chain antibodies (which are referred to herein as “VH domains”) and from the light chain variable domains that are present in conventional 4-chain antibodies (which are referred to herein as “VL domains”). For a further description of VHH’s, reference is made to the review article by Muyldermans 2001 (Reviews in Molecular Biotechnology 74: 277-302).

[0059] For the term “dAb's” and “domain antibody”, reference is for example made to Ward et al. 1989 (Nature 341: 544), to Holt et al. 2003 (Trends Biotechnol. 21: 484); as well as to for example WO 2004 / 068820, WO 2006 / 030220, WO 2006 / 003388 and other published patent applications of Domantis Ltd. It should also be noted that, although less preferred in the context of the present invention because they are not of mammalian origin, single variable domains can be derived from certain species of shark (for example, the so-called “IgNAR domains”, see for example WO 2005 / 18629).

[0060] Immunoglobulin sequences of different origin, comprising mouse, rat, rabbit, donkey, human and camelid immunoglobulin sequences can be used herein. Also, fully human, humanized or chimeric sequences can be used in the method described herein. For example, camelid immunoglobulin sequences and humanized camelid immunoglobulin sequences, or camelized domain antibodies, e.g. camelized dAb as described by Ward et al. 1989 (Nature 341: 544), WO 1994 / 04678, and Davis and Riechmann (1994, Febs Lett., 339:285-290; and 1996, Prot. Eng., 9:531-537) can be used herein. Moreover, the ISVs are fused forming a multivalent and / or multispecific construct (for multivalent and multispecific polypeptides containing one or more VHH domains and their preparation, reference is also made to Conrath et al. 2001 (J. Biol. Chem., Vol. 276, 10. 7346-7350) as well as to for example WO 1996 / 34103 and WO 1999 / 23221).

[0061] A “humanized VHH” comprises an amino acid sequence that corresponds to the amino acid sequence of a naturally occurring VHH domain, but that has been “humanized”, i.e. by replacing one or more amino acid residues in the amino acid sequence of said naturallyAttorney Docket No. 46567-1739WO1occurring VHH sequence (and in particular in the framework sequences) by one or more of the amino acid residues that occur at the corresponding position(s) in a VH domain from a conventional 4-chain antibody from a human being (e.g. indicated above). This can be performed in a manner known per se, which will be clear to the skilled person, for example on the basis of the prior art (e.g. WO 2008 / 020079). Again, it should be noted that such humanized VHHs can be obtained in any suitable manner known per se and thus are not strictly limited to polypeptides that have been obtained using a polypeptide that comprises a naturally occurring VHH domain as a starting material.

[0062] A “camelized VH” comprises an amino acid sequence that corresponds to the amino acid sequence of a naturally occurring VH domain, but that has been “camelized”, i.e. by replacing one or more amino acid residues in the amino acid sequence of a naturally occurring VH domain from a conventional 4-chain antibody by one or more of the amino acid residues that occur at the corresponding position(s) in a VHH domain of a (camelid) heavy chain antibody. This can be performed in a manner known per se, which will be clear to the skilled person, for example on the basis of the description in the prior art (e.g. Davies and Riechman 1994, FEBS 339: 285; 1995. Biotechnol. 13: 475; 1996. Prot. Eng. 9: 531; and Riechman 1999. J. Immunol. Methods 231: 25). Such “camelizing” substitutions are inserted at amino acid positions that form and / or are present at the VH-VL interface, and / or at the so-called Camelidae hallmark residues, as defined herein (see for example WO 1994 / 04678 and Davies and Riechmann (1994 and 1996, supra). In one embodiment, the VH sequence that is used as a starting material or starting point for generating or designing the camelized VH is a VH sequence from a mammal, such as the VH sequence of a human being, such as a VH3 sequence. However, it should be noted that such camelized VH can be obtained in any suitable manner known per se and thus are not strictly limited to polypeptides that have been obtained using a polypeptide that comprises a naturally occurring VH domain as a starting material.

[0063] The assay measurements 112 can be assays that are used to determine pharmacokinetic PK properties of the test antibody, e.g., the nanobody. In some cases, the assay measurements 112 can be based on how the test antibody interacts with the neonatal Fc receptor (RcRn), which is relevant for test antibody half-life, recycling and degradation, clearance, etc. For example, the test antibody assay measurements 112 can include a measurement of how much of the test antibody is recycled, e.g., an antibody cellular recycling neonatal Fc receptor (FcRn) assay measurement, e.g., or a measurement of test antibody binding affinity to FcRn, e.g., a baculovirus particle assay measurement, or a measure of test antibody interaction quality with FcRn, e.g., an FcRn column assay measurement. As another example, the test antibody assayAttorney Docket No. 46567-1739WO1measurements 112 can be related to clearance and cross-reactivity of the test antibody, e.g., a retention time on cross-interaction chromatography assay measurement.

[0064] In some cases, the system 100 can obtain one or more of the test antibody assay measurements 112 as a result of corresponding physical experiments. In particular, a user of the PK prediction inference system 100 can prepare the assays described above, e.g., in a laboratory, to yield the assay measurements for the test antibody 112. However, performing wet-lab experiments can be time-consuming and expensive, especially when assaying multiple test antibodies.

[0065] In other cases, the system 100 can obtain one or more of the test antibody assay measurements 112 as an output of an assay prediction machine learning model 120. In particular, the system 100 can leverage the availability of assay measurement data to predict the output of performing one or more assays, thereby facilitating the efficient generation of the test antibody assay measurements 112 and reducing the time required to generate the model output 160 using the system 100.

[0066] In this case, the system 100 can process data characterizing the amino acid sequence of the test antibody, e.g., the test antibody amino acid sequence data 105, using the assay prediction machine learning model 120 to generate one or more synthetic assay measurements for the test antibody 112. For example, the test antibody amino acid sequence data 105 can be an embedding of the amino acid sequence of the antibody that is generated using a protein language model (“protein language model embedding7’).

[0067] In this case, the protein language model can be a generative neural network that is configured to sequentially process amino acid sequence data and trained to perform next element prediction, e.g., similar to how a language processing model can process characters, words, sentences, etc. to generate an output. In particular, the protein language model can be transformer-based, e.g., an encoder-decoder transformer, an encoder-only transformer, or a decoder-only transformer, configured to perform parallel processing of the contents of the multimodal input using a multi-headed attention mechanism. In this context, the protein language model can have been finetuned to perform a synthetic assay measurement prediction task.

[0068] The assay prediction machine learning model 120 can have any appropriate machine learning architecture, e.g., a neural network, that can be configured to process the test antibody amino acid sequence data 105, e.g., the protein language model embedding representing the test antibody, to generate the synthetic assay measurements for the test antibody 112. In particular, the assay prediction machine learning model 120 can have anyAttorney Docket No. 46567-1739WO1appropriate number of neural network layers (e.g., 1 layer, 5 layers, or 10 layers) of any appropriate type (e.g., fully-connected layers, attention layers, convolutional layers, etc.) connected in any appropriate configuration (e.g., as a linear sequence of layers, or as a directed graph of layers).

[0069] In some cases, the system 100 can additionally process auxiliary' structural features, along with the test antibody acid sequence data 105, as input to the assay prediction machine learning model 120. In this case, the auxiliary structural features can include structural information that is not explicitly encoded in the protein language model embedding representing the test antibody. For example, the auxiliary structural features 105 can include explicit three-dimensional coordinates of the backbone or side chains of the protein. As another example, the auxiliary structural features can include a representation of disulfide bridges that stabilize tertiary or quaternary structure, allosteric interactions, or inter-subunit interfaces. As yet another example, the auxiliary structural features can include side chain rotamer states and binding pocket geometry information.

[0070] For example, the system 100, or another training system, can train the assay prediction machine learning model 120 on the set of training examples by a machine learning training technique to optimize an objective function. In particular, each training example can include data characterizing the amino acid sequence or related features of a particular training antibody as a training input, and a corresponding target output including one or more assay¬ measurements for the particular training antibody. In the case that the assay- is typically- performed using a plate, the assay measurements can be represented as an array of values, e.g., with each value corresponding to a well of the plate.

[0071] The objective function can measure, for each training example, a discrepancy between: (i) the target experimental assay measurements for the training antibody specified by the training example, and (ii) predicted experimental assay measurements generated by the assay- prediction machine learning model 120 by processing the training input of the training example. The objective function can measure a discrepancy between target and predicted experimental assays in any appropriate way. e.g., using an element-wise loss function, e.g., cross-entropy- loss or a mean squared error loss, for each predicted value corresponding to a well of the plate.

[0072] The machine learning training technique can be any technique appropriate for training the assay prediction machine learning model 120. For instance, for an assay prediction machine learning model 120 implemented as a neural network, the model 120 can be trained by calculating and backpropagating gradients of an objective function to update parameterAttorney Docket No. 46567-1739WO1values of the model, e.g., using the update rule of any appropriate gradient descent optimization algorithm, e.g.. RMSprop or Adam.

[0073] In either case, the system 100 can include the actual assay measurements, synthetic assay measurements, or both as the test antibody assay measurements 112 in the model input 110 to the Gaussian process regression model 140. In some cases, the model input 110 can also include a time value 114, e.g., in the case that the system is being used to generate a PK profile 154, as is described in more detail below.

[0074] In a further case, the model input 110 can additionally include an output of a mechanistic pharmacokinetic model, e.g., a mathematical or computational model that characterizes how a drug is absorbed, distributed, metabolized, and excreted by a test subj ect based on underlying biological principles. In particular, mechanistic pharmacokinetic models can evolve a system representing the test antibody’s interaction with the test subject using mass balance equations and physiological parameters. As an example, the model input 110 can include an output of a physiologically-based pharmacokinetic compartment model that accounts for, e.g., organ sizes / tissue volume, blood flow rates, membrane transport, and enzyme activity. In terms of machine learning, this can be classified as a multifidelity approach.

[0075] In some cases, the system 100 can process the model input 110 including the test antibody assay measurements 112 using a preprocessing engine 130. For example, the preprocessing engine 130 can transform each of the different assay measurements 112 to standardize or normalize the assay measurements 112 for processing using the Gaussian process regression model 140. That is, the system 100 can ensure that the model input 110 is in a compatible format for processing using the Gaussian process regression model 140, e.g., based on underlying assumptions that model outputs 160 of the Gaussian process regression model 140 are normally distributed.

[0076] As an example, for each assay measurement of the assay measurements 112, the preprocessing engine 130 can apply a log scale to each value in a particular assay measurement and then standardize the particular assay measurement to have a mean of zero and a standard deviation of one, e.g., by subtracting the mean from each value in a particular assay measurement and dividing by the standard deviation for the particular assay measurement. As another example, for each assay measurement of the assay measurements 112, the preprocessing engine 130 can normalize each value in a particular assay measurement using a scalar determined using the particular assay measurement.Attorney Docket No. 46567-1739WO1

[0077] As yet another example, the preprocessing engine 130 can apply filtering to identify missing values in the assay measurements 120. In this case, the engine 100 can correct for identified missing data by removing the missing data, e.g., removing rows of the array representing the assay, or imputing the values based on existing values in a particular assay measurement.

[0078] The system 100 can process the model input 110 using the Gaussian process regression model 140 to generate a model output 160 including predicted pharmacokinetic (PK) data 150 and an uncertainty value 165. In particular, the Gaussian process regression model 140 implemented in the inference system 100 has been trained to model a prior distribution over function outputs, e.g., the observations of ground truth PK data corresponding with the training assay measurements 146. under the assumption that a finite collection of function outputs is jointly Gaussian distributed. By modeling a prior over functions instead of using fixed parameters to calculate a regression output, the Gaussian process regression model 140 can generate accurate outputs, even with sparse data, through interpolation.

[0079] More specifically, the Gaussian process regression model 140 includes a kernel function 142 that encodes a correlation matrix specifying an expected correlation of model outputs based on model inputs and is defined by trained hyperparameters 144 and training assay measurements 146. In this context, the trained hyperparameters 144 have been numerically optimized using the training assay measurements 146 to define the mapping of the model input 110 to the Gaussian model into a kernel space that can be used to determine the model output 160. For example, the kernel function 142 can define that test antibody assay measurements 112 input to the Gaussian process regression model 140 that are mapped closer together in kernel space have correlated outputs 160.

[0080] As an example, the kernel function 142 can be a squared exponential kernel, e.g., a radial basis function, a linear kernel, or a mixture of both. For example, the kernel function 142 can include one or more trained length scale hyperparameters f that are used in a radial basis function kernel, e.g.,where A is a scale factor. In this case, the length scale factor f can control a sensitivity of the predicted pharmacokinetic data 150 generated by the Gaussian process regression model 140 on each of the test antibody assay measurements 120, e.g., each type of assay measurement canAttorney Docket No. 46567-1739WO1be associated with a respective length scale hyperparameter that is used in a radial basis kernel function.

[0081] As another example, the kernel function 142 can be parameterized using a neural network (not depicted). For example, the kernel function neural network can have any appropriate number of neural network layers (e.g., 1 layer, 5 layers, or 10 layers) of any appropriate type (e.g., fully-connected layers, attention layers, convolutional layers, etc.) connected in any appropriate configuration (e.g., as a linear sequence of layers, or as a directed graph of layers) that can be configured to generate a representation of the assay measurements in an intermediate embedding space that can be used to determine the model output 160. For example, the kernel function neural network can be an embedding neural network.

[0082] An example for determining the trained hyperparameters 144 using gradient-based optimization is described in more detail with respect to FIGS. 2 and 5.

[0083] In more detail, the kernel function 142 represents the prior assumptions of the model 140, given the training assay measurements 146, and can be used to convert the prior distribution into a posterior distribution based on the test antibody assay measurements 112. More specifically, the Gaussian process regression model 140 can use the test antibody assay measurements 112 to determine a posterior mean and variance at a new point as the predicted pharmacokinetic data 150 by conditioning the prior distribution represented by the kernel function 142 using the assay measurements for the test antibody 112.

[0084] The posterior mean and variance can be determined by evaluating a Gaussian distribution A(p, X) using the kernel function 142 K with different inputs, e.g.,where y* is the predicted PK data 150 for input test antibody assay measurements 112 x. , x is the training assay measurements 146, y is the ground truth PK data corresponding with the training assay measurements 146, with functions of the kernel K are defined as K = K(x-,x), K= K(x,x), K ** = K(x., X-), and where o is an empirically-calibrated noise parameter value 148. An example for calibrating the empirically-calibrated noise parameter value is described in more detail with respect to FIGS. 2 and 5.

[0085] More explicitly, the predicted PK data 150 can be determined as the mean of the posterior: K(x*,x)(K(x,x)+ o2ily, and the uncertainty value 165 can be determined as K(x*, x«) - K(x.,x) (K(x,x) + cr2 / )-1K(x,x.). Thus, the system 100 can process a model input 110Attorney Docket No. 46567-1739WO1using the Gaussian process regression model 140 to determine a model output 160 including predicted PK data 150 using the kernel function 142.

[0086] In the case that the system 100 processes the test antibody assay measurements 112, the model output 160 can be one or more PK parameter(s) 152. PK parameters are global summary statistics, e.g., that characterize one or more of: absorption, distribution, metabolism, or elimination (ADME) properties of the test antibody administered, e.g., in a drug, to a subject.

[0087] As an example, the predicted PK parameter(s) 152 can include a clearance of the test antibody, e.g., how fast a drug is eliminated or excreted from the body of the subject, a volume of distribution of the test antibody, a half-life of the test antibody, e.g., how long it takes the test antibody to be reduced to half its value, a maximum concentration of the test antibody, e.g., the maximum serum concentration that a test antibody achieves in a specified area of the body of the subject, or a time to maximum concentration of the test antibody, e.g., how long it takes for a drug to reach maximum concentration after administration and absorption. As another example, the predicted PK parameters can include an intrinsic clearance of the test antibody, e.g., the rate at which the liver can metabolize the drug with unrestricted access, a lipophilicity of the chemical compound, e.g., the ability of the test antibody to mix with an oily phase (as opposed to water), a solubility of the test antibody, e.g., the degree to which a substance dissolves in a solvent to make a solution, a fraction unbound of the test antibody, e.g., the portion that exerts a pharmacologic effect, and an acidic pKa and a basic pKa of the test antibody, e.g., an acid dissociation descriptor and a basic dissociation descriptor, respectively.

[0088] In the case that the system 100 processes the test antibody assay measurements 112 and a time value 114, the model output 160 can include a PK profile 154 defining a concentration value of the test antibody in a subject at the time point represented by the time value. In particular, the system 100 can iteratively process, for a sequence of time values, test antibody assay measurements at each of the time values to generate a PK profile 154 at the sequence of time values, e.g., as is depicted in FIG. 3.

[0089] The system 100 can provide the predicted PK data 150, e.g., by transmitting the predicted PK data 150 to a user device or to another downstream system for performing one or more downstream tasks. As another example, the system 100 can store the predicted PK data in memoiy.

[0090] In some cases, the predicted PK data 150 can be used to estimate a therapeutic dose of the test antibody, or to determine a total exposure of the test antibody in a subject over time, e.g., by determining the area under the curve (AUC) in the case that the predicted PK data 150Attorney Docket No. 46567-1739WO1is a PK profile 154. In this case, the system 100, or a downstream system, can compare the total exposure to an efficacy threshold to predict efficacy of the test antibody, or compare the total exposure to a toxicity threshold to predict toxicity of the test antibody.

[0091] In another case, the system 100 can be used to perform rank ordering of a set of test antibodies, e g., ascending or descending ordering with respect to one or more PK properties 152 or the AUC of the PK profile 154. The rank ordering can then be used to select one or more of the test antibodies for physical synthesis. As a related example, the system 100 can use the predicted PK profile 154 to experimentally validate the properties of the physically synthesized test antibodies.

[0092] FIG. 2 is a system diagram of an example pharmacokinetic prediction training system that can train a Gaussian process regression model for generating predicted pharmacokinetic data. For example, the pharmacokinetic prediction training system 200 can train the Gaussian process regression model 140 of FIG. 1 by determining the values of the trained hyperparameters 282 and the empirically-calibrated noise level parameter 284.

[0093] In the particular example depicted, training the Gaussian process regression model involves updating the hyperparameters of the kernel function, e.g., using a kernel parameter update subsystem 220, and calibrating the noise level parameter cr (described above), using a noise calibration subsystem 250. As described below, the system 200 can also be used for updating the hyperparameters of the kernel function, without calibrating the noise level parameter.

[0094] In the context of performing noise calibration, the system 200 can instantiate one or more Gaussian process regression models, e.g., the models 222, 242, 244, 246, corresponding with a number of different noise level parameter values 215, e.g., noise level parameter values 215(A), 215(B), 215(C). and 215(N), maintained by the noise calibration subsystem 250. In some cases, the noise level parameter values can be obtained as input to the system 200. Each Gaussian process regression model can have a particular noise level parameter, e.g., model 222 can have noise level parameter 215(A), model 242 can have noise level parameter 215(B), model 244 can have noise level parameter 215(C), and model 246 can have noise level parameter 215(N).

[0095] The kernel parameter update subsystem 220 can initiate update of the kernel hyperparameters 226 of each of the Gaussian process regression models 222, 242, 244, and 246 at each of a number of update iterations. While the update process is depicted in full detail wi th respect to Gaussian process regression model 222, the arrows extending from each of the Gaussian process regression models 242, 244, and 246 depict a simplified representation of theAttorney Docket No. 46567-1739WO1update process that is performed with each of the Gaussian process regression models, e.g., during the same duration of time as the update of model 222.

[0096] In particular, the subsystem 220 can evaluate the current values of the hyperparameters 226 of the kernel function 224 at each of a number of updates iterations using an evaluation engine 230. The subsystem 220 can process a set of training model inputs 210 including assay measurements for training antibodies 212 and target PK data 232 corresponding with the training antibodies to evaluate the likelihood of observing the target PK data 234 under the prior distribution represented by the Gaussian process regression model 222. More specifically, the subsystem 220 can evaluate the likelihood of observing the target, e.g., ground truth, PK data 234 from the Gaussian process regression model 222 given the assay measurements for training antibodies 212 with the values of the hyperparameters 226 at the current iteration using the target PK data 232.

[0097] The subsystem 220 can then compute gradients 235 of the likelihood of the target PK data based on the current values of the hyperparameters 226, and can update the values of the hyperparameters 226 using the gradients 235 to maximize the likelihood of observing the training data under the Gaussian process regression model. In the case of performing noise calibration, the noise level parameter cr is a static parameter that is not updated with the hyperparameters 226.

[0098] For example, the subsystem 220 can maximize the log marginal likelihood of the target PK data 234 over a number of training iterations:logL(where 9 are the hyperparameters 226 of the kernel function 222, x are the training assay measurements 212 of the training model inputs 210, y is the target PK data 232 corresponding with the training assay measurements 212 of the training model inputs 210, <7 is the static noise level parameter, and K is the kernel function K(x,x) determined from the training assay measurements 212, n is the number of training model inputs 210.

[0099] In particular, the subsystem 220 can update the hyperparameters 226 at each of a number of training iterations, e.g., N training iterations, where N is 10, 20, or 100 iterations, or until a training criterion is satisfied, e.g., when the likelihood 234 satisfies a threshold likelihood value.

[0100] After the update process is completed for each of the Gaussian process regression models 222, 242, 244, and 246, the noise calibration subsystem 250 can determine the empirically-calibrated noise level parameter 284 using a set of validation model inputs 260Attorney Docket No. 46567-1739WO1including assay measurements for validation antibodies 262 and target PK data 264 corresponding with the validation antibodies. In particular, the subsystem 250 can process the assay measurements for the validation antibodies 262 using each of the models 222, 242, 244, and 246 and compare the outputs to the target PK data 264 corresponding with the validation antibodies to determine a prediction accuracy for each of the models 222, 242, 244, and 246 with respective noise level parameter values 215(A), 215(B), 215(C). and 215(N).

[0101] The noise calibration subsystem 250 can then select the trained Gaussian process regression model 280 based on the prediction accuracies 270. For example, the subsystem 250 can select the trained Gaussian process regression model with the highest prediction accuracy as the model 280. Thus, the system 200 can determine the values of the trained hyperparameters 282 and empirically calibrate the noise level parameter 284 of the Gaussian process regression model deployed for inference, e.g., in the PK prediction inference system 100 of FIG. 1.

[0102] While described here in the context of performing noise calibration, in some cases, the system 200 can train a single Gaussian process regression model using the update process as described above, e.g., with respect to Gaussian process regression model 222, without calibrating the noise. In this case, the system 200 uses the kernel parameter update subsystem 220 to determine the values of the kernel hyperparameters 226, and, in some cases, to also determine the value of the noise level parameter 215(A). As another example, the system 200 can instantiate the noise level parameter 215(A) of the model 222 based on a user input and specify that the noise level parameter 215(A) should remain static during the update process.

[0103] FIG. 3 depicts example PK profile prediction results of an example pharmacokinetic prediction inference system.

[0104] Panels 300 and 330 depict example PK profile prediction graphs as results from processing assay measurements for a number of test antibodies at each of a sequence of time values using a Gaussian process regression model, e.g., that has been trained as described in FIG. 2. In the graphs in panels 300 and 330, the error bars represent the uncertainty value output from the Gaussian process regression model. As demonstrated by the overlap between the observed and predicted PK values at each of the sequence of time steps, the predicted PK profiles replicate the observed PK profiles with high fidelity.

[0105] Additionally, the results in graphs 302, 304, 332, 334, and 336 demonstrate that the Gaussian process regression model is able to capture complex PK behavior, e.g., triphasic behavior. In particular, graphs 302, 304, 332, 334, and 336 illustrate a steep drop in concentration during a distribution phase, a moderately steep slope in concentration during anAttorney Docket No. 46567-1739WO1elimination phase, and a flat tail during a terminal phase, thereby demonstrating the accuracy of the Gaussian process regression model.

[0106] FIG. 4 is a flow diagram of an example process for generating predicted pharmacokinetic data characterizing an antibody using a Gaussian process regression model. For convenience, the process 400 will be described as being performed by a system of one or more computers located in one or more locations. For example, a pharmacokinetic prediction inference system, e.g., the pharmacokinetic prediction inference system 100 of FIG. 1. appropriately programmed in accordance with this specification, can perform the process 400.

[0107] The system can obtain a number of assay measurements for a test antibody (step 410). For example, the test antibody can include a nanobody. As an example, the number of assay measurements for the test antibody can include one or more of an antibody cellular recycling FcRn assay measurement, a baculovirus particle assay measurement, a retention time on crossinteraction chromatography assay measurement, or an FcRn column assay measurement.

[0108] In some cases, one or more of the number of assay measurements can be determined hy¬ performing physical assays on the test antibody. In other cases, one or more of the number of assay measurements can be synthetic assay measurements generated using an assay prediction machine learning model.

[0109] In particular, the system can process data characterizing an amino acid sequence of the antibody using the assay prediction machine learning model, in accordance with trained values of a set of assay prediction machine learning model parameters, to generate the synthetic assay measurement as a model output. In this case, the data characterizing the amino acid sequence of the antibody can include an embedding of the amino acid sequence of the antibody that is generated by a protein language model. For example, the assay prediction machine learning model can have been trained by a machine learning training technique on a set of training examples that each include: (i) a training input that characterizes an amino acid sequence of a training antibody, and (ii) a target output that defines one or more experimental assay measurements for the training antibody.

[0110] The system can process a model input including the number of assay measurements using a Gaussian process regression model to generate predicted pharmacokinetic data characterizing the test antibody as output of the Gaussian process regression model (step 420). In an example, the model input can also include an output of a mechanistic pharmacokinetic model, e.g., a mathematical or computational model that characterizes how a drug is absorbed, distributed, metabolized, and excreted by a test subject using mass balance equations and physiological parameters to model the flow of the test antibody based on underlying biologicalAttorney Docket No. 46567-1739WO1principles. The model output can also include an uncertainty value defining an uncertainty in the predicted pharmacokinetic data characterizing the test antibody.

[0111] In some cases, the predicted pharmacokinetic data characterizing the test antibody can include one or more predicted pharmacokinetic parameters of the test antibody. As an example, the predicted pharmacokinetic parameter can include a Cmax parameter defining a maximum test antibody concentration in plasma. As another example, the predicted pharmacokinetic parameter can include a Tmax parameter defining a time at which maximum test antibody concentration in plasma occurs. As yet another example, the predicted pharmacokinetic parameter can include an area under the curve (AUC) parameter defining total exposure to the test antibody over time. As an additional example, the predicted pharmacokinetic parameter can include a half life parameter that defines a duration of time for the test antibody concentration in the plasma to decrease by half. As a further example, the predicted pharmacokinetic parameter can include a clearance parameter that defines a rate at which the test antibody is removed from a subject.

[0112] In the case that the model input to the Gaussian process regression model additionally includes a time value, the predicted pharmacokinetic data characterizing the test antibody can include data defining a predicted pharmacokinetic profile of the test antibody evaluated at the time value. That is, the system can iteratively process a sequence of time values and corresponding assay measurements for the test antibody at each of the sequence of time values to generate a predicted pharmacokinetic profile of the test antibody at the sequence of time values.

[0113] In particular, the Gaussian process regression model can generate the predicted pharmacokinetic data based at least in part on a kernel function evaluated between: (i) the number of assay measurements for the test antibody, and (ii) a respective number of assay measurements for each of a number of training antibodies. In some cases, the kernel function can include a mixture of a squared exponential kernel and a linear kernel. In other cases, the kernel function can be parameterized using a neural network.

[0114] More specifically, the kernel function of the Gaussian process regression model can include a number of hyperparameters that have been numerically optimized, e.g.. for a number of training antibodies. For example, the kernel function can include a length scale hyperparameter for each of the number of assay measurements that controls a sensitivity of the predicted pharmacokinetic data generated by the Gaussian process regression model on the assay measurement. In this case, the length scale hyperparameters can have been optimized by gradient-based optimization. An example method for optimizing the hyperparameters of theAttorney Docket No. 46567-1739WO1kernel function of the Gaussian process regression model will be described in more detail with respect to FIG. 5.

[0115] Furthermore, the Gaussian process regression model can include a noise level parameter that is empirically calibrated based on a set of validation antibodies that are not used for configuring the Gaussian process regression model, as is also described in more detail with respect to FIG. 5.

[0116] The system can output the predicted pharmacokinetic data characterizing the test antibody (step 430). As an example, the system can present the predicted pharmacokinetic data on a display of a user device. As another example, the system can transmit the predicted pharmacokinetic data over a data communications network. As yet another example, the system can store the predicted pharmacokinetic data in a memory.

[0117] For example, the system can estimate a therapeutic dose of the test antibody based at least in part on the predicted pharmacokinetic data for the test antibody. As another example, the system can select the test antibody for physical synthesis based at least in part on the predicted pharmacokinetic data characterizing the test antibody. In some cases, the system can physically synthesize the test antibody.

[0118] As yet another example, in the case that the predicted pharmacokinetic data for the test antibody is a predicted pharmacokinetic profile for the test antibody, the system can determine a total exposure of the test antibody over time from the predicted pharmacokinetic profile for the test antibody. In this case, the system can then compare the total exposure to an efficacy threshold to predict efficacy of the test antibody, or compare the total exposure to a toxicity threshold to predict toxicity of the test antibody.

[0119] FIG. 5 is a flow diagram of an example process for training a Gaussian process regression model to generate predicted pharmacokinetic data. Steps depicted in dashed lines are optional. For convenience, the process 500 will be described as being performed by a system of one or more computers located in one or more locations. For example, a pharmacokinetic prediction training system, e.g., the pharmacokinetic prediction training system 200 of FIG. 2. appropriately programmed in accordance with this specification, can perform the process 500.

[0120] In the case that the system is performing noise calibration (optional, as represented by the dashed lines), the system can instantiate a respective Gaussian process regression model for each of a number of noise level parameters (step 510). That is, the system can instantiate a respective Gaussian process regression model with a noise level parameter having a particular noise level parameter value.Attorney Docket No. 46567-1739WO1

[0121] The system can compute gradients of a likelihood of pharmacokinetic data for a number of training antibodies under the Gaussian process regression model at current iteration (step 520), e.g., for each Gaussian process regression model. For example, the number of training antibodies used for configuring the Gaussian process regression model can be less than 1000, or less than 100 training antibodies.

[0122] In this context, the likelihood of pharmacokinetic data for the number of training antibodies refers to evaluating the likelihood of observing the pharmacokinetic data for the training antibodies using the current values of the hyperparameters of the Gaussian process regression model. As an example, the likelihood can be the log marginal likelihood of the pharmacokinetic data for the number of training antibodies, e.g., given the current values of the hyperparameters.

[0123] The system can adjust current values of hyperparameters using gradients (step 530). For example, the system can adjust the current values of the hyperparameters, e.g., using gradient-descent, over a number of iterations until a training criterion is satisfied. In the case that the system is performing noise calibration, the system can perform steps 520-530 for each of the initiated Gaussian process regression models.

[0124] In the case that the system is performing noise calibration, the system can evaluate the prediction accuracy of each Gaussian process regression model on a set of validation antibodies (step 540). The set of validation antibodies can be separate from the training antibodies, e.g., the set of validation antibodies was not used for configuring the Gaussian process regression model in steps 520-530.

[0125] The system can then select the noise level parameter based on the prediction accuracies of the Gaussian process regression models having the noise level parameter values (step 550). For example, the system can select the noise level parameter value by selecting the noise level parameter associated with the highest prediction accuracy of the respective Gaussian process regression models.

[0126] This specification uses the term “configured” in connection with systems and computer program components. For a system of one or more computers to be configured to perform particular operations or actions means that the system has installed on it software, firmware, hardware, or a combination of them that in operation cause the system to perform the operations or actions. For one or more computer programs to be configured to perform particular operations or actions means that the one or more programs include instructions that, when executed by data processing apparatus, cause the apparatus to perform the operations or actions.Attorney Docket No. 46567-1739WO1

[0127] Embodiments of the subject matter and the functional operations described in this specification can be implemented in digital electronic circuitry, in tangibly-embodied computer software or firmware, in computer hardware, including the structures disclosed in this specification and their structural equivalents, or in combinations of one or more of them. Embodiments of the subject matter described in this specification can be implemented as one or more computer programs, i.e., one or more modules of computer program instructions encoded on a tangible non-transitory storage medium for execution by. or to control the operation of, data processing apparatus. The computer storage medium can be a machine-readable storage device, a machine-readable storage substrate, a random or serial access memory device, or a combination of one or more of them. Alternatively or in addition, the program instructions can be encoded on an artificially-generated propagated signal, e.g., a machine-generated electrical, optical, or electromagnetic signal, that is generated to encode information for transmission to suitable receiver apparatus for execution by a data processing apparatus.

[0128] The term “data processing apparatus” refers to data processing hardware and encompasses all kinds of apparatus, devices, and machines for processing data, including by way of example a programmable processor, a computer, or multiple processors or computers. The apparatus can also be, or further include, special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application-specific integrated circuit). The apparatus can optionally include, in addition to hardware, code that creates an execution environment for computer programs, e.g., code that constitutes processor firmware, a protocol stack, a database management system, an operating system, or a combination of one or more of them.

[0129] A computer program, which may also be referred to or described as a program, software, a software application, an app, a module, a software module, a script, or code, can be written in any form of programming language, including compiled or interpreted languages, or declarative or procedural languages; and it can be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment. A program may, but need not, correspond to a file in a file system. A program can be stored in a portion of a file that holds other programs or data, e.g., one or more scripts stored in a markup language document, in a single file dedicated to the program in question, or in multiple coordinated files, e.g., files that store one or more modules, sub-programs, or portions of code. A computer program can be deployedAttorney Docket No. 46567-1739WO1to be executed on one computer or on multiple computers that are located at one site or distributed across multiple sites and interconnected by a data communication network.

[0130] In this specification the term '‘engine” is used broadly to refer to a software-based system, subsystem, or process that is programmed to perform one or more specific functions. Generally, an engine will be implemented as one or more software modules or components, installed on one or more computers in one or more locations. In some cases, one or more computers will be dedicated to a particular engine; in other cases, multiple engines can be installed and running on the same computer or computers.

[0131] The processes and logic flows described in this specification can be performed by one or more programmable computers executing one or more computer programs to perform functions by operating on input data and generating output. The processes and logic flows can also be performed by special purpose logic circuitry, e.g., an FPGA or an ASIC, or by a combination of special purpose logic circuitry and one or more programmed computers.

[0132] Computers suitable for the execution of a computer program can be based on general or special purpose microprocessors or both, or any other kind of central processing unit. Generally, a central processing unit will receive instructions and data from a read-only memory or a random access memory or both. The essential elements of a computer are a central processing unit for performing or executing instructions and one or more memory devices for storing instructions and data. The central processing unit and the memory can be supplemented by, or incorporated in, special purpose logic circuitry. Generally, a computer will also include, or be operatively coupled to receive data from or transfer data to, or both, one or more mass storage devices for storing data, e.g., magnetic, magneto-optical disks, or optical disks. However, a computer need not have such devices. Moreover, a computer can be embedded in another device, e.g., a mobile telephone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a Global Positioning System (GPS) receiver, or a portable storage device, e.g., a universal serial bus (USB) flash drive, to name just a few.

[0133] Computer-readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media and memory devices, including by way of example semiconductor memory devices, e.g., EPROM, EEPROM, and flash memory devices; magnetic disks, e.g., internal hard disks or removable disks; magneto-optical disks; and CD-ROM and DVD-ROM disks.

[0134] To provide for interaction with a user, embodiments of the subject matter described in this specification can be implemented on a computer having a display device, e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor, for displaying information to theAttorney Docket No. 46567-1739WO1user and a keyboard and a pointing device, e.g., a mouse or a trackball, by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback, e.g., visual feedback, auditor}' feedback, or tactile feedback; and input from the user can be received in any form, including acoustic, speech, or tactile input. In addition, a computer can interact with a user by sending documents to and receiving documents from a device that is used by the user; for example, by sending web pages to a web browser on a user’s device in response to requests received from the web browser. Also, a computer can interact with a user by sending text messages or other forms of message to a personal device, e.g., a smartphone that is running a messaging application, and receiving responsive messages from the user in return.

[0135] Data processing apparatus for implementing machine learning models can also include, for example, special-purpose hardware accelerator units for processing common and compute-intensive parts of machine learning training or production, i.e., inference, workloads.

[0136] Machine learning models can be implemented and deployed using a machine learning framework, e.g., a TensorFlow framework, or a Jax framework.

[0137] Embodiments of the subject matter described in this specification can be implemented in a computing system that includes a back-end component, e.g., as a data server, or that includes a middleware component, e.g., an application server, or that includes a front-end component, e.g., a client computer having a graphical user interface, a web browser, or an app through which a user can interact with an implementation of the subject matter described in this specification, or any combination of one or more such back-end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication, e.g., a communication network. Examples of communication networks include a local area network (LAN) and a wide area network (WAN), e.g., the Internet.

[0138] The computing system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. In some embodiments, a server transmits data, e.g., an HTML page, to a user device, e.g., for purposes of displaying data to and receiving user input from a user interacting with theAttorney Docket No. 46567-1739WO1device, which acts as a client. Data generated at the user device, e.g.. a result of the user interaction, can be received at the server from the device.

[0139] While this specification contains many specific implementation details, these should not be construed as limitations on the scope of any invention or on the scope of what may be claimed, but rather as descriptions of features that may be specific to particular embodiments of particular inventions. Certain features that are described in this specification in the context of separate embodiments can also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments separately or in any suitable subcombination. Moreover, although features may be described above as acting in certain combinations and even initially be claimed as such, one or more features from a claimed combination can in some cases be excised from the combination, and the claimed combination may be directed to a subcombination or variation of a subcombination.

[0140] Similarly, while operations are depicted in the drawings and recited in the claims in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed, to achieve desirable results. In certain circumstances, multitasking and parallel processing may be advantageous. Moreover, the separation of various system modules and components in the embodiments described above should not be understood as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0141] Particular embodiments of the subject matter have been described. Other embodiments are within the scope of the following claims. For example, the actions recited in the claims can be performed in a different order and still achieve desirable results. As one example, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some cases, multitasking and parallel processing may be advantageous.

[0142] What is claimed is:

Claims

Attorney Docket No. 46567-1739WO1CLAIMS1. A method performed by one or more computers, the method comprising:obtaining a plurality of assay measurements for a test antibody;processing a model input comprising the plurality of assay measurements for the test antibody using a Gaussian process regression model to generate, as a model output of the Gaussian process regression model, predicted pharmacokinetic data characterizing the test antibody,wherein the Gaussian process regression model generates the predicted pharmacokinetic data based at least in part on a kernel function evaluated between: (i) the plurality of assay measurements for the test antibody, and (ii) a respective plurality of assay measurements for each of a plurality of training antibodies; andoutputting the predicted pharmacokinetic data characterizing the test antibody.

2. The method of claim 1 , wherein the model output of the Gaussian process regression model further comprises an uncertainty value defining an uncertainty in the predicted pharmacokinetic data characterizing the test antibody.

3. The method of any preceding claim, wherein the kernel function of the Gaussian process regression model comprises a plurality of hyperparameters that have been numerically optimized by performing operations comprising, at each of a plurality of optimization iterations:computing gradients of a likelihood of the predicted pharmacokinetic data of the plurality of training antibodies under the Gaussian process regression model implemented using current values of the plurality of hyperparameters; andadjusting the current values of the plurality of hyperparameters using the gradients.

4. The method of claim 3, wherein the likelihood is a log marginal likelihood.

5. The method of any preceding claim, wherein the kernel function comprises a respective length scale hyperparameter for each of the plurality of assay measurements that controls a sensitivity of the predicted pharmacokinetic data generated by the Gaussian process regression model on the assay measurement; andwherein the lengths scale hyperparameters for the plurality of assay measurements have been optimized by gradient-based optimization.Attorney Docket No. 46567-1739WO16. The method of any preceding claim, wherein the kernel function comprises a mixture of a squared exponential kernel and a linear kernel; orwherein the kernel function is parametrized by a neural network.

7. The method of any preceding claim, wherein the Gaussian process regression model comprises a noise level parameter that is empirically calibrated based on a set of validation antibodies that are not used for configuring the Gaussian process regression model.

8. The method of claim 7, wherein empirically calibrating the noise level parameter based on the set of validation antibodies comprises:instantiating, for each of a plurality of noise level parameter values, a respective Gaussian process regression model that comprises a noise level parameter having the noise level parameter value and that is configured using the plurality of training antibodies;evaluating, for each of the plurality of noise level parameter values, a prediction accuracy of the Gaussian process regression model that comprises the noise level parameter having the noise level parameter value on the set of validation antibodies; andselecting the noise level parameter value based on the prediction accuracies of the Gaussian process regression models comprising noise level parameters having the noise level parameter values.

9. The method of claim 8, wherein selecting the noise level parameter value based on the prediction accuracies of the Gaussian process regression models comprising noise level parameters having the noise level parameter values comprises:selecting a noise level parameter value associated with a highest prediction accuracy.

10. The method of any preceding claim, wherein a number of training antibodies used for configuring the Gaussian process regression model is less than 1000.

11. The method of any preceding claim, wherein the number of training antibodies used for configuring the Gaussian process regression model is less than 100.

12. The method of any preceding claim, wherein the predicted pharmacokinetic data characterizing the test antibody comprises one or more predicted pharmacokinetic parameters of the test antibody.Attorney Docket No. 46567-1739WO113. The method of claim 12, wherein the one or more predicted pharmacokinetic parameters of the test antibody comprise one or more ofa Cmax parameter defining a maximum test antibody concentration in plasma; or a Tmax parameter defining a time at which maximum test antibody concentration in plasma occurs; oran AUC parameter defining total exposure to the test antibody over time; or a half time parameter that defines a duration of time for the test antibody concentration in plasma to decrease by half; ora clearance parameter that defines a rate at which the test antibody is removed from a subject.

14. The method of any preceding claim, wherein the model input to the Gaussian process regression model additionally comprises a time value; andwherein the predicted pharmacokinetic data characterizing the test antibody comprises data defining a predicted pharmacokinetic profile of the test antibody evaluated at the time value.

15. The method of any preceding claim, wherein the plurality of assay measurements for the test antibody comprise one or more of:an antibody cellular recycling FcRn assay measurement;a baculovirus particle assay measurement;a retention time on cross-interaction chromatography assay measurement; or an FcRn column assay measurement.

16. The method of any preceding claim, wherein outputting the predicted pharmacokinetic data characterizing the test antibody comprises: presenting the predicted pharmacokinetic data on a display of a user device; or transmitting the predicted pharmacokinetic data over a data communications network; or storing the predicted pharmacokinetic data in a memory.

17. The method of any preceding claim, wherein one or more of the plurality of assay measurements for the test antibody are determined by performing physical assays on the test antibody.

18. The method of any preceding claim, wherein one or more of the plurality of assay measurements for the test antibody are synthetic assay measurements generated byAttorney Docket No. 46567-1739WO1performing operations comprising:processing data characterizing an amino acid sequence of the antibody using an assay prediction machine learning model and in accordance with trained values of a set of assay prediction machine learning model parameters to generate the synthetic assay measurement as a model output of the assay prediction machine learning model.

19. The method of claim 18, wherein the data characterizing the amino acid sequence of the antibody comprises an embedding of the amino acid sequence of the antibody that is generated by a protein language model.

20. The method of any one of claims 18-19. wherein the assay prediction machine learning model has been trained by a machine learning training technique on a set of training examples that each comprise: (i) a training input that characterizes an amino acid sequence of a training antibody, and (ii) a target output that defines one or more experimental assay¬ measurements for the training antibody.

21. The method of any preceding claim, further comprising estimating a therapeutic dose of the test antibody based at least in part on the predicted pharmacokinetic data for the test antibody.

22. The method of any preceding claim, wherein the predicted pharmacokinetic data for the test antibody comprises a predicted pharmacokinetic profile for the test antibody; and wherein the method further comprises:determining a total exposure of the test antibody over time from the predicted pharmacokinetic profile for the test antibody; andcomparing the total exposure to an efficacy threshold to predict efficacy of the test antibody, or comparing the total exposure to a toxicity threshold to predict toxicity of the test antibody.

23. The method of any preceding claim, wherein the test antibody comprises a nanobody.

24. The method of any preceding claim, wherein the model input to the Gaussian process regression model further comprises an output of a mechanistic pharmacokinetic model.Attorney Docket No. 46567-1739WO125. The method of any preceding claim, further comprising selecting the test antibody for physical synthesis based at least in part on the predicted pharmacokinetic data characterizing the test antibody.

26. The method of claim 25, further comprising physically synthesizing the test antibody.

27. A system comprising:one or more computers; andone or more storage devices communicatively coupled to the one or more computers, wherein the one or more storage devices store instructions that, when executed by the one or more computers, cause the one or more computers to perform operations of the respective method of any one of claims 1-25.

28. One or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations of the respective method of any one of claims 1-25.