Methods and systems for biopharmaceutical development

Abstract

This provides methods for biopharmaceutical development. [Solution] A method is disclosed that includes determining experimental data associated with one or more monoclonal antibodies (mAbs), determining computationally derived data associated with one or more mAbs, wherein the computationally derived data includes one or more computational parameters weighted based on the exposed area (ASA) of one or more residues of one or more mAbs, determining a plurality of candidate predictive models based on the experimental data and the computationally derived data, determining the optimal predictive model from the plurality of candidate predictive models, and outputting the optimal predictive model.

Description

[Technical Field]

[0001] (Cross-reference of related applications) This application claims priority to U.S. Provisional Patent Application No. 63 / 108,716, filed 2 November 2020, which is incorporated herein by reference in its entirety. [Background technology]

[0002] Since muromonab, a mouse CD3-specific IgG2a monoclonal antibody (mAb) for acute organ transplant rejection, was the first monoclonal antibody to receive FDA approval in 1986, more than 64 mAbs have been approved by the FDA. The popularity of this therapeutic platform is evident from the increasing number of ongoing clinical trials each year and the expansion of its use into a wide variety of different therapeutic portfolios. Therapeutic mAbs are most commonly administered via three routes of administration: intravenous (IV), intramuscular (IM), and subcutaneous (SC) injections, with the choice based on various contributing factors including safety, efficacy, patient satisfaction, and pharmacoeconomics. IV administration can usually be delivered at controllable high doses in a clinic setting and is therefore typically more expensive for both patients and clinicians. The shift from IV administration to SC administration for most immunoglobulins could potentially reduce overall healthcare costs by easing the economic burden on the healthcare system through faster self-administration by patients at home or administration by healthcare professionals in clinics, reducing long hospital stays, and improving the overall quality of care. Despite offering many advantages over IV administration, SC administration presents several significant challenges for drug development and drug delivery. The main drawbacks of SC administration are innate tolerance to the extracellular matrix and volume limitations, requiring high concentrations of antibody solution (≥150 mg / mL) to be administered in limited injectable volumes (approximately 2-3 mL) for optimal PK / PD outcomes and user convenience.

[0003] High-concentration antibody solutions are difficult to develop in the pharmaceutical field because high protein concentrations can present significant technical challenges, such as high solution viscosity and protein aggregation rates. High-viscosity antibodies also present difficulties related to manufacturing processes and drug delivery. Protein aggregation can lead to decreased antibody activity and, due to its greater immunogenic potential, can affect the pharmacokinetics and safety of the protein. Therefore, appropriate formulations should be developed for high-concentration therapeutic antibodies to stabilize both colloidal and stereostructural structures, thereby reducing viscosity and aggregation tendencies, and ensuring both an acceptable shelf life and compatibility with manufacturing processes.

[0004] Conventional approaches to developing high-concentration protein formulations generally involve empirical methods utilizing a wide range of biophysical and analytical techniques. For example, solution viscosity is measured by direct viscometers, and the osmotic pressure quadratic virial coefficient (B) is used. 22 ) measurement, and diffusion interaction parameter (K D ) can be predicted experimentally by established tools such as measurements. Such tools, for example, B 22 The measurements require a considerable amount of material and are quite labor-intensive. Protein aggregation and association are thought to be the result of the interaction between steric stability (i.e., macroscopic and microscopic perturbations in the protein structure) and colloidal stability (i.e., native intermolecular interactions).

[0005] Various established biophysical and analytical tools and approaches are routinely used to assess overall protein stability and predict the individual contributions of structural and colloidal stability. These stabilities can be measured and predicted by various established techniques that enable the quantification of thermal stress, stirring stress, and freeze-thaw stress. Stability prediction approaches include thermal stress stability studies, thermal denaturation temperature (T). m ), chemical denaturation temperature, coagulation temperature (T agg ), cloud point temperature (T cloudThese include measurements of zeta potential, direct surface hydrophobicity measurements using hydrophobic interaction chromatography, zeta potential, and higher-order structure estimation. All of these techniques require physical materials, and some are complex and time-consuming.

[0006] Furthermore, experimentally developed predictive models fail in most cases. Most of the aforementioned techniques for measuring and predicting viscosity and aggregation rates are costly, time-consuming, and require physical materials. Therefore, the development of novel experimental and / or computer-aided tools to predict viscosity values ​​and aggregation tendencies, or to rank antibodies based on their developmental potential, is essential for rapidly screening many mAb candidates during early formulation development and drug discovery.

[0007] The antibody sequence, homology model, and physical properties obtained from molecular dynamics (MD) simulations of individual antibody molecules can be used as parameters for developing predictive models or ranking schemes. These models can be used to predict viscosity, steric stability, colloidal stability, and manufacturability. Furthermore, these rapid, material-free tools provide molecular insights into antibody molecules and their interactions.

[0008] Sharma et al. developed a viscosity predictive model based on the variable domain Fv of 14 IgG1 homology models. Physical parameters, including but not limited to charge and hydrophobicity, correlated with viscosity values ​​measured in a major component regression model. Agrawal et al. also developed a viscosity scoring function to rank IgG1 antibodies based on the partial charge of surface-exposed residues in the Fv region of the homology models and proposed a threshold for distinguishing highly viscous antibodies from others.

[0009] Tomar et al. predicted the concentration-dependent viscosity curves of antibody solutions based on electrostatic and hydrophobic descriptors obtained from full-length homology models of 16 IgG1, IgG2, and IgG4 antibodies. The slope of the linearized concentration-dependent viscosity curve was predicted using the hydrophobic surface area of ​​the full-length antibodies and the charge relative to the Fv and hinge regions.

[0010] Furthermore, researchers in this field have developed various methods for predicting aggregation-prone regions of peptides and therapeutic proteins. The TANGO statistical mechanics algorithm was developed based on the physicochemical principles of β-sheet formation to predict sequence-based aggregation. Waltz, a web-based tool, was designed to identify amyloid-forming regions in protein sequences using a site-specific scoring matrix. Chennamsetty et al. developed a method to predict aggregation-prone regions based on the dynamic exposure of hydrophobic patches obtained from atomic simulations of antibodies. A comprehensive list of computational methods developed to predict aggregation and aggregation-prone regions of therapeutic proteins can be found in chapters and summaries of published books.

[0011] Overall, there is a need for more robust and predictable models for viscosity and aggregation tendencies to facilitate drug development. [Overview of the project]

[0012] A method is described that includes determining experimental data associated with one or more monoclonal antibodies (mAbs), determining computationally derived data associated with one or more mAbs, wherein the computationally derived data includes one or more computational parameters weighted based on the exposed area (ASA) of one or more residues of one or more mAbs, determining a plurality of candidate predictive models based on the experimental data and the computationally derived data, determining the optimal predictive model from the plurality of candidate predictive models, and outputting the optimal predictive model.

[0013] The report also describes a method that includes receiving computationally derived data related to monoclonal antibodies (mAbs), providing the computationally derived data to a predictive model, and determining a viscosity score related to the mAb based on the predictive model.

[0014] A method is also described that includes receiving computationally derived data related to monoclonal antibodies (mAbs), providing computationally derived data to a predictive model, and determining an agglutination score related to the mAbs based on the predictive model.

[0015] Further advantages of the disclosed methods and compositions may be, in part, described in the following description, inferred from the description, or learned through the practice of the disclosed methods and compositions. The advantages of the disclosed methods and compositions will be realized and achieved by the elements and combinations specifically indicated in the appended claims. It should be understood that both the above general description and the following detailed description are illustrative and descriptive of the claimed invention and not limiting.

[0016] The accompanying drawings incorporated herein and forming part of this specification illustrate some embodiments of the disclosed methods and compositions and serve to illustrate, together with the description, the principles of the disclosed methods and compositions. [Brief explanation of the drawing]

[0017] [Figure 1] Figure 1 is an example flowchart for generating a predictive model to assist in screening and / or selection for treatment. [Figure 2] Figure 2 shows an example of a block diagram for generating a predictive model. [Figure 3] Figure 3 is a flowchart illustrating an example of a training method. [Figure 4]Figure 4 is a diagram illustrating an exemplary process flow for determining whether a nucleotide sequence is a promoter using a machine learning-based classifier. [Figure 5] Figure 5 shows an example of a diffusion interaction parameter (KD) calculated based on fitting a line to the diffusion coefficients at various protein concentrations. This graph shows, as an example, how the KD was calculated for mAb4. [Figure 6A] Figures 6A and 6B are tables showing the parameters calculated for the 16 complete antibody models used in this study. These physical properties were obtained from the complete antibody homology models used to develop predictive models for protein solution viscosity. ZVL, ZVH, ZCL, ZCH1, ZHinge, ZCH2, and ZCH3 are the effective charges for the VL, VH, CL, CH1, Hinge, CH2, and CH3 regions, respectively; ZmAb is the total antibody charge; Z*VL, Z*VH, Z*CL, Z*CH1, Z*Hinge, Z*CH2, and Z*CH3 are the net solvent-exposed area (SAS) adjusted charges for the VL, VH, CL, CH1, Hinge, CH2, and CH3 regions, respectively; HI is the hydrophobicity index; DmAb is the averaged total dipole moment of the antibody; PISequence and PIStructure are the sequence-based and structure-based isoelectric points, respectively; and AP is the aggregation tendency predicted by Chennamsetty. [Figure 6B] Same as above. [Figure 7]Figure 7 is a table showing the calculated parameters of the Fab models. In this study, calculated parameters for 14 Fab models were used. These physical properties were obtained from Fab homology models and molecular dynamics simulations used to develop predictive models for aggregation tendencies. ZVL, ZVH, ZCL, and ZCH1 are the effective charges for the VL, VH, CL, and CH1 regions, respectively; ZFab is the total Fab region charge; Z*VL, Z*VH, Z*CL, and Z*CH1 are the net solvent-exposed area (SAS) adjusted charges for the VL, VH, CL, and CH1 regions, respectively; HI is the hydrophobicity index; DmAb is the averaged total dipole moment of the Fab region; PISequence and PIStructure are the sequence-based and structure-based isoelectric points, respectively; AP is the aggregation tendency predicted by Chennamsetty; and RMSD is the averaged mean squared deviation (Å) obtained from molecular dynamics simulations of the Fab region. [Figure 8] Figure 8 shows the broad distribution of protein solution viscosity values ​​measured for the 16 mAbs used in this study. IgG1 and IgG4 candidates are shown in gray and black, respectively. Our dataset indicates that IgG1 antibodies tend to exhibit lower viscosity values ​​compared to IgG4 candidates. [Figure 9] Figures 9A, 9B, and 9C show examples of measurements for 15 mAbs used in this study. The strong correlation between B22 and KD values ​​for the current dataset is observed by (a) the osmotic second-order virial coefficient (B22), (b) the diffusion interaction parameter (KD), and (c) the high correlation coefficient (R). MAb1 was excluded from these plots due to the lack of material for KD measurement. Candidate IgG1 and IgG4 are colored gray and black, respectively. [Figure 10]Figures 10A and 10B show examples of correlations between measured values ​​and protein solution viscosity: a) osmotic second-order virial coefficient (B22), and b) diffusion interaction parameter (KD). B22 values ​​were measured for 16 mAbs used in this study, and KD values ​​were measured for 15 of them because sufficient material was not available for mAb1. Linear correlation coefficients (R) and regression lines are shown in each graph. [Figure 11A] Figures 11A, 11B, and 11C show the linear relationship between experimental viscosity values ​​and calculated parameters. The linear fit equation and correlation coefficient (R) are shown for each plot. ZVL, ZVH, ZCL, ZCH1, ZHinge, ZCH2, and ZCH3 are the effective charges for the VL, VH, CL, CH1, Hinge, CH2, and CH3 regions, respectively; ZmAb is the total antibody charge; Z*VL, Z*VH, Z*CL, Z*CH1, Z*Hinge, Z*CH2, and Z*CH3 are the net solvent-exposed area (SAS) adjusted charges for the VL, VH, CL, CH1, Hinge, CH2, and CH3 regions, respectively; HI is the hydrophobicity index; DmAb is the averaged total dipole moment of the antibody; PISequence and PIStructure are the sequence-based and structure-based isoelectric points, respectively; and AP is the aggregation tendency predicted by Chennamsetty. [Figure 11B] Same as above. [Figure 11C] Same as above. [Figure 12] Figure 12 shows an example of a linear regression line between the calculated predicted viscosity score (PVS) and the measured viscosity value. The correlation coefficient (R) and the squared correlation coefficient (R²) are shown in the graph. The area between the dashed lines represents the 95% confidence interval. [Figure 13]Figure 13 shows a representative example of the size exclusion chromatography (SEC) signal for one of the antibody mAb3s used in this study, illustrating incubation over days 0 and 28 at 40°C and 75% relative humidity. Increased high molecular weight (HMW) species formation is observed as a result of aggregation. Data for day 0 and day 28 are shown by the red dotted line and the black solid line, respectively. [Figure 14] Figure 14 shows the %ΔHMW, which is the relative percentage of high molecular weight (HMW) species formation at 7, 14, and 28 days compared to day 0, measured by size exclusion chromatography (SEC) for the 14 mAbs used in this study. The samples were incubated at 40°C and 75% relative humidity for 7, 14, and 28 days. MAb9 and mAb16 were excluded due to limited material availability. [Figure 15] Figure 15 shows the %ΔHMW / day, which is the daily high molecular weight (HMW) species formation rate for the 14 mAbs used in this study, calculated based on the %ΔHMW of the 28-day data points divided by 28. IgG1 and IgG4 are colored gray and black, respectively. [Figure 16A] Figures 16A, 16B, and 16C show the mean squared deviation (RMSD) of the configurational structures of the 14 mAbs used in this study relative to their initial structures. These structures were obtained from three molecular dynamics simulations of 2.0 ns each for the Fab region of each antibody. The first, second, and third simulations for each mAb are color-coded in black, red, and blue, respectively. [Figure 16B] Same as above. [Figure 16C] Same as above. [Figure 17]Figure 17 shows examples of averaged mean squared deviations (RMSDs) of skeletal atoms in the Fab model from the initial structures of 14 mAbs used in the current experiment to develop a predictive model for aggregation. RMSD values ​​can be used as descriptors of steric stability to distinguish antibodies from one another. The RMSD value for each mAb is the average of three molecular dynamics (MD) simulations. mAb9 and mAb16 were excluded due to limited material availability. [Figure 18A] Figures 18A, 18B, and 18C show the averaged mean squared deviation (RMSD) for three 2.0 ns molecular dynamics simulations of the Fab region in each mAb. The RMSD for each simulation was calculated for each three-dimensional structure relative to the initial structure. [Figure 18B] Same as above. [Figure 18C] Same as above. [Figure 19A] Figures 19A, 19B, and 19C show the linear relationship between the measured daily high molecular weight (HMW) species formation rate, %ΔHMW / day, and the calculation parameters, including the mean squared deviation (RMSD) obtained from molecular dynamics simulations. [Figure 19B] Same as above. [Figure 19C] Same as above. [Figure 20] Figure 20 shows an example of a linear regression line between the calculated predicted agglutination score (PAS) and the measured daily high molecular weight (HMW) formation rate. The correlation coefficient (R) and the squared correlation coefficient (R²) are shown in the graph. The area between the dashed lines represents the 95% confidence interval. [Figure 21] Figure 21 shows the correlation between the predicted viscosity score and the validation experimental data. [Figure 22] Figure 22 shows the correlation between the predicted agglomeration score and the validation experiment data. [Figure 23] Figure 23 shows an exemplary operating environment. [Figure 24] Figure 24 shows an exemplary method. [Figure 25] Figure 25 shows an exemplary method. [Figure 26]Figure 26 shows an exemplary method. [Modes for carrying out the invention]

[0018] By referring to the modes for carrying out the invention with respect to the specific embodiments and examples contained herein, as well as the drawings and their accompanying descriptions, an understanding of the disclosed methods and compositions can be facilitated.

[0019] Naturally, the methods and compositions described herein are not limited to the specific methodologies, protocols, and reagents described, for they may be modified. It should also be understood that the terms used herein are intended solely to describe specific embodiments and do not limit the scope of the invention as defined solely by the appended claims.

[0020] It should be noted that, as used herein and in the appended claims, the singular forms "a," "an," and "the" include plural references unless the context explicitly indicates otherwise. For example, a reference to "one antibody" includes multiple such antibodies, and a reference to "antibody" refers to one or more antibodies and their equivalents known to those skilled in the art.

[0021] As used herein, the term “antibody” refers to the entire antibody. An antibody is a glycoprotein comprising at least two heavy (H) chains and two light (L) chains interconnected by disulfide bonds. Each heavy chain consists of a heavy chain variable region (abbreviated herein as VH) and a heavy chain constant region. The heavy chain constant region consists of three subdomains: CH1, CH2, and CH3. Each light chain consists of a light chain variable region (abbreviated herein as VL) and a light chain constant region. The light chain constant region consists of one subdomain: CL. The VH and VL regions can be further subdivided into hypervariable regions called complementarity-determining regions (CDRs), which are dotted with more conserved regions called framework regions (FRs). Each VH and VL consists of three CDRs and four FRs arranged from the amino terminus to the carboxyl terminus in the following order: FR1, CDR1, FR2, CDR2, FR3, CDR3, FR4. The variable regions of the heavy and light chains contain binding domains that interact with the antigen. The constant region of the antibody can mediate the binding of immunoglobulins to host tissues or factors, including various cells of the immune system (e.g., effector cells) and the first component (Clq) of the classical complement system. In some embodiments, the antibody may be a chimeric antibody, a monoclonal antibody, and / or a humanized antibody.

[0022] An antibody fragment can refer to any smaller portion of the entire antibody. Antibody fragments can be described in terms of proteolytic fragments, including but not limited to Fv (variable region fragment), Fab (antibody-binding region fragment), Fab', and F(ab')2 (antibody-binding fragment with a portion of the hinge region added) fragments. Such fragments may be prepared by standard methods (see, for example, Coligan et al., Current Protocols in Immunology, John Wiley & Sons, 1991–1997, incorporated herein by reference). The antibody may contain at least three proteolytic fragments (i.e., fragments produced by cleavage with papain), namely, two Fab fragments (designated herein as "Fab heavy-chain domains"), each containing a light-chain domain and a heavy-chain domain, and one Fc fragment containing two Fc domains. Each light-chain domain contains a VL subdomain and a CL subdomain, each Fab heavy-chain domain contains a VH subdomain and a CH1 subdomain, and each Fc domain contains a CH2 subdomain and a CH3 subdomain. In some embodiments, the antibody fragment may be a chimeric antibody fragment, a monoclonal antibody fragment, and / or a humanized antibody fragment.

[0023] As used herein, the terms “monoclonal antibody” or “monoclonal antibody fragment” refer to an antibody or antibody fragment obtained from a single clonal population of immunoglobulins that bind to the same epitope of an antigen. Monoclonal antibodies have the same Ig gene rearrangement and therefore exhibit identical binding specificity. Methods for preparing monoclonal antibodies are known in the art.

[0024] As used herein, "humanized monoclonal antibody" or "humanized monoclonal antibody fragment" may refer to a monoclonal antibody or fragment thereof having at least a human constant region and antigen-binding regions of non-human species, such as one, two, or three CDRs. Humanized antibodies or fragments specifically recognize a target antigen but do not induce an immune response to the antibody itself in humans.

[0025] As used herein, the terms “chimeric antibody” or “chimeric antibody fragment” refer to a monoclonal antibody or fragment thereof comprising a variable region from one source (e.g., species) and at least a portion of a constant region derived from a different source. In some embodiments, the chimeric antibody comprises a mouse variable region and a human constant region.

[0026] Throughout this specification and the claims, the word “comprise” and its variations, such as “comprising” and “comprises,” mean “including but not limited to,” and are not intended to exclude, for example, other additional things, components, integers, or processes. In particular, where a method is described as including one or more processes or operations, each process is specifically intended to include what is listed (unless the process includes a limiting term such as “consisting of”), and it is not intended that each process excludes, for example, other additional things, components, or processes not listed in the process.

[0027] During drug discovery and early development, most drug candidates are initially screened and selected based on affinity and functionality. However, other properties and attributes must be considered in biopharmaceutical development. For example, protein yield, viscosity, aggregation, chemical stability (e.g., susceptibility to degradation by oxidation and deamidation), compatibility, and immunogenicity should form part of a comprehensive development risk assessment. The concept of development is used to define the suitability of a drug candidate (e.g., an antibody) to be developed as a therapeutic / medicinal agent. Once an antibody is identified and developed as a drug, it can be administered to patients.

[0028] A method for predicting antibody viscosity is disclosed. This viscosity prediction tool can be used in the early stages of the drug development process, enabling the ranking and selection of lead antibodies with reduced risk of being viscous.

[0029] Antibody drugs require high-concentration formulations, which can result in cumbersome and viscous solutions, both in the manufacturing process and during injection to the end user. Antibody viscosity is often only discovered in the later stages of drug development, after significant R&D investment. The methods and systems described herein can predict antibody viscosity in the early stages of the drug development process, enabling prioritization of low-viscosity antibody candidates. While viscosity prediction tools are related to antibody analysis, it should be understood that these techniques can also be applied to other proteins. Proteins may, but are not limited to, clinical candidates.

[0030] A method for predicting antibody aggregation is disclosed. This aggregation prediction tool can be used in the early stages of the drug development process, enabling the ranking and selection of lead antibodies with reduced aggregation risk. Protein aggregation, a common problem during biopharmaceutical development, can occur at different stages of the manufacturing and development process, including fermentation, purification, formulation, filling and final formulation, and storage. Aggregation can affect not only the manufacturing process but also the target product profile, product efficacy, delivery, and, importantly, patient safety. Protein aggregates have been reported to contribute to cases of immune response in patients.

[0031] These aggregates can manifest as reversible oligomers, invisible or visible particles, or precipitates. The protein aggregation process is determined by many factors, including the amino acid composition and sequence, environmental factors such as pH, concentration, buffers / excipients and shear forces during the protein production process, and final formulation and storage conditions.

[0032] In some embodiments, aggregation prediction can be used in combination with other computer-based prediction tools to screen and select target antibodies. For example, the disclosed aggregation prediction model can be combined with the disclosed viscosity prediction model, or with known immunogenicity or degradation prediction tools. These combinations of tools allow for the selection of one or more antibodies with reduced risks of aggregation, viscosity, degradation, and / or immunogenicity, which can then be proceeded to in vitro expression and characterization analysis.

[0033] One embodiment shown in Figure 1 describes a method 100 for generating a predictive model to assist in treatment screening, ranking, and / or selection. In method 110, experimental parameters may be determined. The experimental parameters may relate, for example, to protein yield, viscosity, aggregation, chemical stability (e.g., susceptibility to degradation by oxidation, deamidation), compatibility, and / or immunogenicity.

[0034] Experimental parameters may be determined from experimental data. Experimental data may be data generated, for example, by measurements, test methods, experimental design, and / or quasi-experimental design. In clinical research, the data produced are the results of clinical trials. Experimental data may be qualitative or quantitative, each suitable for different investigations. Experimental data may include values ​​of experimental parameters obtained by performing one or more experiments related to antibodies.

[0035] In one embodiment, experimental parameters related to viscosity may be determined. In some embodiments, techniques for measuring viscosity measure how a sample responds to flow rate, velocity, and time. For example, a capillary viscometer can be used to measure the time it takes for a sample to pass through a tube. Similar to the use of a capillary viscometer, the Zahn cup method can be used, in which a small hole is placed at the bottom of a cup, and the time it takes for the sample to pass through the hole is measured. Viscosity can also be measured using the falling ball viscometer technique, in which a ball of known density is dropped into the sample, and the time it takes for the ball to fall to a specified point is recorded. In some embodiments, a vibrating viscometer is used to measure the damping of a vibrating electromechanical resonator immersed in the sample. Rotational viscometer techniques may also be used, which measure the torque required to rotate an object in the sample as a function of the viscosity of the sample.

[0036] In one embodiment, experimental parameters related to antibody aggregation may be determined. These parameters can be determined using any known protein aggregation technique. For example, biochemical assays for measuring aggregation include, but are not limited to, ultracentrifugation, size exclusion chromatography, gel electrophoresis, dynamic light scattering, or turbidiometry. Many of these techniques take into account the size difference between protein monomers and aggregates. Fluorescence-based assays may also be used, in which the fluorescence yield is increased by fluorophores in the presence of protein aggregates.

[0037] In one embodiment, experimental parameters related to protein yield can be performed using techniques known in the art. Protein concentration is similar to protein yield, but establishes the amount of protein in a specific volume of solution. Protein concentration is most often determined using a spectrophotometer. Once the protein concentration is determined, the protein yield can be determined. Therefore, if the sample has a protein concentration of 5 mg / ml, the total protein yield when the protein yield is 100 ml is 500 mg.

[0038] In one embodiment, experimental parameters related to antibody-antigen docking can be performed using known techniques. A “standard criterion” for obtaining this data is experimentally determining the 3D structure of the antibody-antigen complex using X-ray crystallography. Other structural methods such as cryo-electron microscopy (cryoEM) or nuclear magnetic resonance (NMR) can also be used, although the latter is difficult due to the size of the complex. These experimental data, which indicate the potential for binding between the antibody and its antigen, may result in different conformational changes that may occur during binding.

[0039] In one embodiment, experimental parameters related to immunogenicity may be determined. The immunogenicity of therapeutic antibodies can cause adverse side effects. Immunogenicity can be determined experimentally using animal experiments. Antibodies may be administered to animals (such as mice or rabbits), and then, at different time points, serum derived from the animals can be tested for the immune response to the antibody (particularly the T-cell and B-cell response). In most cases, the lower the immunogenicity, the better the choice of therapeutic antibody. In some embodiments, the immunogenicity of an antibody can be altered by humanizing the antibody.

[0040] In one embodiment, experimental parameters related to chemical stability may be determined. Chemical stability can be an important attribute of therapeutic proteins, particularly antibodies. In most cases, the more likely an antibody is to degrade, the less desirable it is as a therapeutic agent. The most common method for experimentally determining chemical stability is to use gel electrophoresis. Pulse chase assays can also be used. pH, temperature, and proteases are all factors affecting chemical stability. Therefore, even slight changes in formulation can affect chemical stability.

[0041] At 120, the computational parameters may be determined. The computational parameters may be determined by computational analysis and / or simulation. The computational parameters may be determined from computationally derived data. Computationally derived data may be data generated by, for example, sequence analysis, antibody numbering, complete FV region modeling, Ab-specific side chain prediction, antibody-specific loop prediction, side chain prediction, first-principles loop prediction, CDR canonical structure prediction, VH / VL orientation, paratope prediction, protein contact prediction, Ab-specific epitope prediction, Ab-specific docking, non-specific docking, structure prediction, homology modeling, protein-protein docking simulation, molecular dynamics simulation, etc. Experimental data may include values ​​of experimental parameters obtained by performing computational analysis related to the antibody.

[0042] In one embodiment, computational parameters may be determined via antibody numbering. Antibody sequences may be mapped onto a standardized reference framework. The raw nucleotide sequences of the variable regions can be translated into amino acids by aligning them with germline sequences, thereby identifying the V, D, and J regions. This can be achieved by programs such as IgBLAST or IMGT V-Ques and several other tools intended for processing raw antibody data. Similarity between antibody amino acid sequences further enables the generation of a standardized reference framework, or numbering scheme, assigning an identifier to each variable region amino acid. The numbering scheme describes the context of each position within the antibody structure and enables rapid depiction of the CDR and framework regions. Antibody numbering may be the first step in computational antibody analysis, such as homology modeling.

[0043] In one embodiment, computational parameters may be determined via antibody modeling. Structural antibody modeling generates 3D structures from antibody sequences based on existing knowledge of antibody structures in particular, and protein structures in general. Advanced antibody sequence and structure conservation in framework regions and five canonical loops results in overall high accuracy of antibody homology modeling. Antibody modeling generally involves selecting a suitable framework template that can accommodate the CDR loop. This can be achieved by searching for sequence matches that closely resemble the H and L chains within available databases. The relative orientation of the VH and VL domains is determined, which affects the shape of the paratope. The CDR loop is then modeled. An antibody-specific knowledge-based approach can be used to predict the CDR loop according to the template. If a suitable template is unavailable, as is often seen with CDRH3, a more computationally expensive first-principles approach can be used, which generates a large set of novel loops to select the best loop model. The side chains are then constructed and refined. Approaches focused on proteins in general and / or antibodies may be employed. The final antibody model can be further refined by optimizing the energy packing of the molecule. For example, various modeling tools can be used, such as Biovia from Accelrys (https: / / www.3dsbiovia.com / ), SmrtMolAntibody from Macromoltek (https: / / www.macromoltek.com / ), MOE from CCG (https: / / www.chemcomp.com / ), and BioLuminate from Schrodinger Inc. (https: / / www.schrodinger.com / products / bioluminate). The modeling tools are accurate to antibody F with a mean squared deviation (RMSD) of 1.1 Å on average. VThe entire model can be generated, and the most difficult region is CDRH3, which is modeled with an RMSD > 5 Å for some targets. Such results are not usually comparable to the accuracy of structures derived from experiments, but models with an RMSD of 1.0 Å can be used as a quick proxy for depicting the structural features of the molecule. The model structure can be used to characterize the binding to selected surface-exposed paratope residues or to homologous epitopes due to mutations. Using accurate structural information, various development potential indicators such as hydrophobicity, which depends on an accurate model of the molecular surface of the paratope and epitope, can be evaluated.

[0044] In one embodiment, the computational parameters related to residue charge may be determined via antibody homology modeling. The homology model of the full antibody and / or Fab (antigen-binding fragment) may be constructed via modeling software using a Protein Data Bank (PDB) crystal structure as a template. In one embodiment, the full antibody and / or Fab homology model may be constructed to determine computational parameters of protein viscosity and / or protein aggregation tendency. As described herein, the computational parameters may be determined using the full antibody and / or Fab homology model in molecular dynamics simulations. The energy of the antibody structure may be determined based on the homology model and then minimized by geometric optimization. The antibody structure may be protonated, and then computational parameters such as the charge and average dipole moment for the residues may be determined.

[0045] Antibodies (Z in the full antibody model) containing variable and constant regions of both the light and heavy chains VL Z VH Z CL Z CH1 Z Hinge Z CH2 Z CH3 Z Total Z in the Fab model VL Z VH Z CL ZCH1 , Z Total The charge for one or more regions of the ) may be determined as a computational parameter. In one embodiment, the effective charge of each residue may be adjusted by considering the relative solvent exposure area (SAS) of that residue within the homology model of the corresponding antibody. In one embodiment, the total exposed surface area of ​​each amino acid can be determined using a built-in algorithm in the Discovery Studio software. In this approach, depending on the model used, the charge for each residue may be multiplied by a weighting coefficient calculated using the SAS of the residue relative to the total SAS of either the complete antibody or Fab. For example, in a variable light chain, the adjusted charge for each residue may be calculated using Equation 3, and the total SAS adjusted charge for this region may be calculated using Equation 4. These SAS adjusted charges are calculated in the complete model Z * VL , Z * VH , Z * CL , Z * CH1 , Z * Hinge , Z * CH2 , Z * CH3 , Z * Total In the Fab model, Z * VL , Z * VH , Z * CL , Z * CH1 , Z * Total It may also be labeled as such.

[0046]

number

[0047] In the formula, i = any residue in the variable light chain (VL), and n = the number of residues in the complete model or Fab model of the specified antibody.

[0048]

number

[0049] In the formula, m = the number of residues in the variable light chain (VL). In one embodiment, the hydrophobicity index (HI) as a computational parameter may be determined using a complete antibody and / or Fab homology model. The HI of the variable fragment (Fv) is:

[0050]

number

[0051] It may be determined as follows, where i represents a hydrophobic amino acid, e.g., A, C, F, G, I, L, M, P, V, W, Y; j represents a hydrophilic amino acid, e.g., D, E, H, K, N, Q, R, S, T; n is the number of each amino acid; and E is the Eisenberg scale value of each amino acid. In one embodiment, the mean dipole moment (HI) as a computational parameter may be determined using a complete antibody and / or a Fab homology model. The mean dipole moment of the complete model and the Fab model can be determined from the protonated structure.

[0052] In one embodiment, the isoelectric point (pI) may be determined using a complete antibody homology model, a Fab homology model, and / or antibody sequence data. In one embodiment, the atom-by-atom aggregation tendency (AP) score may be determined using a complete antibody and / or a Fab homology model. The AP score may be determined, for example, based on the CHARMM force field and SAS patch of hydrophobic residues exposed at a radius of 10 Å. The total aggregation score for each antibody may be determined as the sum of the aggregation scores of all residues in either the complete antibody homology model or the Fab homology model. Examples of calculation parameters that may be determined based on the antibody homology model and / or Fab homology model are shown in Table 1.

[0053] [Table 1]

[0054] In one embodiment, the computational parameters may be determined based on molecular dynamics (MD) simulations. MD simulations may be used to include conformational changes in the Fab region related to aggregation tendencies. The atoms of the Fab structure can be assigned force field parameters by structurally matching each residue to its template. These structures may be explicitly solvated within the truncated octahedral box of the TIP3P water molecule. The counterion is Na. + and Cl - It may be added to an explicitly solvated system to neutralize the system. In each simulation, the system energy may be minimized using the steepest descent algorithm, and then further minimized using the Adopted Basis Newton-Raphson (ABNR) minimization method to remove large strains in the system. The system may be gradually heated under constant volume (NVT) and simulated at constant temperature and pressure. Long-range electrostatics can be determined using the particle mesh Ewald (PME) method and the cutoff distance of the van der Waals interaction. The SHAKE algorithm may be used in each simulation to constrain the bond length for all hydrogen atoms. Simulations may be performed for each system to allow scrutiny of the reproducibility of the results, differing only in the initial distribution of velocities. The trajectory and time-varying atomic coordinates of each simulation may be captured. In one embodiment, the skeletal mean square deviation (RMSD) of the configurational structure relative to the initial structure after rigid alignment in each simulation may be determined as a computational parameter, which is therefore a descriptor of structural stability.

[0055] In one embodiment, computational parameters may be determined via interface prediction and antibody-antigen docking. A computational method may be employed to predict the antibody-antigen contact surface. The computational method may predict, for example, a paratope, an epitope, or the entire antibody-antigen complex. Approximately half of the 40-50 residues in the CDR are in direct contact with the antigen and form a paratope. Statistical methods such as Antibody i-Patch assign a score to each residue regarding its tendency to become part of a paratope, with high-scoring residues providing candidates with potential for mutagenesis. Since not all paratope residues are constrained by the CDR, their positions within the framework region that may contribute to antigen recognition can be computationally identified.

[0056] Computational methods for epitope prediction can be divided into linear epitope predictors, which focus on identifying the continuous extension of the primary amino acid sequence, and structural epitope predictors, which aim to identify the 3D configuration of the epitope. While paratope and epitope prediction can provide useful information regarding antibody-antigen recognition, these methods do not provide information about the specific interactions involved in antibody-antigen binding. This problem is addressed by antibody-antigen docking, a specialized application of the broader field of molecular docking. Molecular docking predicts biological complexes starting from unbound proteins. Typically, it involves two steps: a sampling step in which thousands of possible complex structures are generated, and a scoring step in which the structures are ranked according to a specific scoring function to identify models that are close to the native structure.

[0057] In one embodiment, computational parameters may be determined through the evaluation of the "humanization" of a therapeutic agent via sequence analysis. Most antibodies currently under development are discovered through immunization of animals. Molecules produced in animals such as mice carry the risk of inducing an immunological response in humans in the form of anti-drug antibodies (ADAs). To avoid such problems, animal-derived antibodies undergo a process called humanization. During this process, (typically) CDRs from mouse-derived antibodies are transplanted onto a human framework, or, alternatively, the mouse-derived framework is manipulated to resemble the human framework. Traditionally, humanization involves comparing the animal-derived sequence to approximately 1000 human germline sequences and then selecting a suitable template. However, germline sequences provide only a limited perspective on overall mutant antibody diversity, which can be addressed by computational humanization, comparing the animal-derived therapeutic agent to the distribution of amino acids in human antibody sequences. In one embodiment, a computational method may be employed that compares a query therapeutic sequence to a set of recombinant variable region sequences that serve as a reference in humanization. In one embodiment, a calculation method may be employed to evaluate the "human-likeness" of a query therapeutic sequence by determining how closely the amino acid content of the query therapeutic sequence resembles the human amino acid distribution.

[0058] In one embodiment, computational parameters may be determined through computational prediction of immune epitopes and ADAs generated against a biopharmaceutical. Generating an immune response to a biopharmaceutical requires multiple steps beyond the reproduction of the diversity of human antibody sequences. Humanized antibodies and even fully human antibodies may induce an immune response in patients receiving such treatment and form ADAs in those patients. ADA formation is a multifactorial issue and may depend, for example, on the patient's genetic background, medical history, protein aggregates in the therapeutic agent, and other degradation products. The components of ADA formation are the binding of short peptide fragments derived from the biopharmaceutical to major histocompatibility complex class II (MHC II) molecules. Therefore, computational methods may be used to identify the potential of MHC I and MHC II-binding T cell epitopes, as well as the three-dimensional B cell epitopes and T cell epitopes.

[0059] In one embodiment, computational parameters related to the biophysical properties of the therapeutic agent may be determined. Examples include biophysical properties such as colloidal stability, concentration-dependent viscosity behavior, and physicochemical degradation of the antibody solution. Solubility avoids aggregation, which could lead to loss of activity, antibody degradation, or immunogenicity. From a general standpoint, protein aggregation has two aspects: mechanistic and kinetic. The mechanistic aspect focuses on protein instability and identifying potential APRs, primarily hydrophobic patches on the protein surface, which may potentially be nuclear aggregation. Computational methods may be used to predict APRs in biopharmaceuticals, according to sequence analysis, for example, the presence of multiple well-defined agglutinating motifs (often located within CDRs). These APRs located within CDRs may contribute to antigen binding. Furthermore, sequence analysis can be used to identify aggregation enhancers and mitigating mutations in the protein. Computational methods may be used to predict solubility. Sequence analysis can be used to determine the presence of one or more predictors of solubility and APR in the protein. Computational methods may be used to predict hydrophobicity. The identification of hydrophobic regions may be performed using homology models.

[0060] In step 130, one or more candidate predictive models may be determined. In one embodiment, experimental and computational parameters may be analyzed to determine one or more predictive models that depend on computational parameters that are determined to have a significant influence on the experimental parameters. One or more computational methods can be used to determine one or more predictive models, including, for example, adaptive context tree weighting, neural networks, CART (classification trees and regression trees), projection tracking regression, stepwise regression, linear regression, elastic networks, multivalued models, MARS (multivariate adaptive regression splines), power laws, linear graphical LASSO, ridge regression, and general additive models (GAM).

[0061] In one embodiment, one or more predictive models may be determined using stepwise multiple regression (including forward selection or backward reduction), forced input, forced elimination, and hierarchical multiple regression. For example, multiple regression analysis may be used to establish relationships between all independent variables (e.g., experimental parameters) and dependent variables (computational parameters). The relationships establish the relative influence of the independent variables. Next, forward selection (related to stepwise regression) may be used to determine the relationships of the independent variables. Forward selection can be started with no independent variables in the equation (related to multiple regression). Independent variables that show the highest correlation or influence with the dependent variable may be added to the equation. The performance of the resulting predictive model may be determined using evaluation techniques. Evaluation techniques (e.g., "goodness-of-fit" analysis techniques) such as the Akaike Information Criterion (AIC), R², RMS, p-value, F-ratio, and standard error may be used to establish the performance characteristics of the relationships. For example, techniques such as R² to establish the rate of variance of the dependent variable (e.g., computational parameters) collectively explained by the independent variables (e.g., experimental parameters). Using R², for example, it is possible to evaluate which relationship best explains the variance of the dependent variable in response to the independent variable. Techniques such as AIC serve as estimators of within-sample prediction error and therefore as indicators of the relative quality of the prediction model.

[0062] The forward selection process can be iterative, allowing for the addition of another independent variable (and associated coefficients) to the equation and then evaluation of the equation. After all independent variables have been added, evaluation metrics (e.g., AIC, R2) may be compared to determine the equation that best describes the relationship. The variable in the equation that best describes the relationship may be considered the most relevant variable, and other variables may be ignored. For example, a decision may be made regarding which variable configuration yielded the lowest AIC, and / or which variable configuration yielded the highest R2 or a significant improvement in R2. In another example, the relationship may be evaluated each time an independent variable is added to determine whether a significant improvement (e.g., a substantial decrease in AIC) has been observed. If the evaluation metrics have not changed significantly, the process may be stopped, and the independent variable currently forming the relationship may be considered the most relevant.

[0063] The backward reduction process (related to stepwise regression) begins with all independent variables in the equation and, like the forward process, sequentially removes them to determine the desired relationships. For example, after establishing the relative influence of the independent variables, the least influential independent variable may be removed from the equation. If the resulting AIC does not decrease significantly, the process can be repeated. In one embodiment, stepwise regression may be used when constructing the equation or to reduce the variables used in establishing the predictive model.

[0064] In step 140, the predictive model may be selected from the candidate predictive models generated in step 130. In one embodiment, a validation technique such as one-miss cross-validation (LOOCV) may be used to select the predictive model. LOOCV is a method in which data points are systematically excluded from the dataset, and then their endpoint values ​​are predicted by relationships derived from the remaining data points (see Cramer et al., Quant. Struct-Act. Relat. 7:18-25, 1998, incorporated herein by reference). Cross-validation is useful for determining the reliability of relationships, especially when a validation dataset is unavailable. The mean and standard deviation of the errors of the predicted LOOCV values ​​from the experimental values ​​may be used as criteria for comparing and selecting predictive models.

[0065] Once selected, the predictive model is presented with new computational parameters that enable it to make predictions related to experimental parameters. For example, the predictive model may be trained according to experimental parameters related to mAb solution viscosity and computational parameters related to the charge values ​​of mAb residues. The predictive model may also be presented with the type of computational parameters on which it was trained, and the predictive model makes predictions related to the experimental parameters on which it was trained.

[0066] For example, a predictive model may be generated according to experimental parameters derived through viscosity measurements consisting of multiple mAb solutions. Viscosity measurements may be obtained via the use of a viscometer. A homology model may be computationally generated and used to determine charge values ​​associated with mAb residues. The charge values ​​may be weighted based on whether the residue is determined to be a surface-exposed residue. The charge values ​​and / or weighted charge values ​​may be used as computational parameters. The predictive model may be generated according to experimental and computational parameters. The predictive model may be configured to generate a score indicating viscosity. The predictive model may model an untrained mAb (e.g., an input mAb) to generate charge values / weighted charge values.

[0067] Charge values ​​and / or weighted charge values ​​may be provided to a predictive model that generates a score indicating viscosity associated with the input mAb. For example, the predictive model may be generated according to experimental parameters generated through aggregation measurements consisting of multiple mAb solutions. Aggregation measurements may be obtained via the use of dynamic light scattering. A homology model may be computationally generated and used to determine charge values ​​associated with residues of the mAb. The charge values ​​may be weighted based on whether the residue is determined to be a surface-exposed residue. Charge values ​​and / or weighted charge values ​​may be used as computational parameters. The predictive model may be generated according to experimental and computational parameters. The predictive model may be configured to generate a score indicating aggregation. The predictive model may model an mAb that has not been trained (e.g., an input mAb) to generate charge values / weighted charge values. Charge values ​​and / or weighted charge values ​​may be provided to a predictive model that generates a score indicating aggregation associated with the input mAb.

[0068] In some embodiments, computational derivation data can be provided to an optimal predictive model, and based on the optimal predictive model, a viscosity score associated with the query mAb can be determined. Based on the viscosity score, an appropriate formulation composition or protein engineering strategy can be adjusted to mitigate specific challenges of the drug candidate under development, such as adjusting the amount of viscosity-reducing agent in the solution associated with the query mAb. In some embodiments, the same can be done for the aggregation score in addition to, or instead of, the viscosity score. When high viscosity and aggregation scores are calculated for a target mAb, various formulation development or protein engineering strategies can be designed. In some embodiments, high aggregation and viscosity scores may indicate the presence of intermolecular interactions that can be determined by a combination of colloidal and steric interactions. Various generally recognized safe (GRAS) excipients are known to stabilize colloidal and steric instability, and combinations of such excipients can be used to stabilize the mAb structure and reduce viscosity. In some embodiments, high viscosity and low aggregation scores may indicate that the intermolecular interactions are transient and determined primarily by colloidal interactions. Here too, various GRAS excipients are known to reduce electrostatic and hydrophobic interactions between mAbs in solution. In some embodiments, protein engineering can be used to substitute specific amino acids involved in such interactions. In some embodiments, high aggregation scores and low viscosity scores may indicate aggregation mainly due to conformational instability. Excipients such as sucrose, various diols and salts have been shown to conform to the conformational stability of the protein and can be used in such cases.

[0069] Referring here to Figure 2, an additional method for generating a predictive model is illustrated. The described method may use machine learning ("ML") techniques to train at least one ML module 230 configured to predict protein viscosity scores and / or protein aggregation scores for any antibody based on an analysis of one or more training datasets 210 by a training module 220.

[0070] The training dataset 210 may include experimental parameters related to the direct measurement of the viscosity of antibody solutions and / or antibody aggregation. The experimental parameters are related to computational parameters related to the corresponding antibodies. The computational parameters may be related to the charge values ​​of residues on the corresponding antibodies determined through computational modeling. For example, the measurement of the viscosity of a first mAb solution may be related to the charge value of the first mAb. Such data can be derived entirely or partially from the experimental data and / or computationally derived data described herein.

[0071] A subset of experimental parameters related to computational parameters may be randomly assigned to the training dataset 210 or the test dataset. In some implementations, the assignment of data to the training or test dataset may not be entirely random. In this case, one or more criteria may be used during the assignment. In general, any preferred method may be used to assign data to the training or test dataset, while ensuring that the yes and no labeling distributions are somewhat similar in the training and test datasets.

[0072] The training module 220 may train the ML module 230 by extracting a feature set from computational parameters (e.g., labeled by experimental parameters) in the training dataset 210 using one or more feature selection techniques. The training module 220 may also train the ML module 230 by extracting a feature set from the training dataset 210 that contains statistically significant features.

[0073] The training module 220 may extract feature sets from the training dataset 210 in various ways. The training module 220 may perform feature extraction multiple times each time using different feature extraction techniques. For example, feature sets generated using different techniques may each be used to generate different machine learning-based classification models 240. For example, the feature set with the highest quality metric may be selected for use in training. The training module 220 may use the feature sets to construct one or more machine learning-based classification models 240A to 240N configured to show the calculated viscosity and / or calculated agglutination score of a new mAb (e.g., having an unknown viscosity and / or unknown agglutination).

[0074] The training dataset 210 may be analyzed to determine any dependencies, relationships, and / or correlations between features and experimental parameters in the training dataset 210. Identified correlations may take the form of a list of features. As used herein, the term “feature” can refer to any feature of an item in the data that can be used to determine whether an item in the data falls into one or more specific categories. For example, features described herein include Z VL (V L Charge relative to the region), Z VH (V H Charge relative to the region), Z CL (C L Charge relative to the region), Z CH1 (C H Charge in a region, Z Hinge (Charge relative to the hinge region), Z CH2 (C H Charges in 2 regions), Z CH3 (C H Charges in 3 regions), Z mAb (total charge), Z * VL (V L Solvent exposure area (SAS) relative to the region (adjusted charge), Z * VH (V H (SAS-adjusted charge relative to the region), Z *CL (C L (SAS-adjusted charge relative to the region), Z * CH1 (C H (SAS-adjusted charge for one region), Z * Hinge (SAS-adjusted charge relative to the hinge region), Z * CH2 (C H (SAS-adjusted charge for 2 regions), Z * CH3 (C H SAS-adjusted charge for 3 regions), HI (hydrophobicity index), D mAb Or D Fab (Total dipole moment), pI Sequence (Sequence-based pI), pI Structure It may include one or more of the following: (structure-based pI), AP (predictive agglomeration tendency), and / or RMSD (mean squared deviation of steric change).

[0075] A feature selection technique may include one or more feature selection rules. One or more feature selection rules may include feature generation rules. Feature generation rules may include determining which features occur a threshold number of times in the training dataset 210, and identifying those features that satisfy the threshold as features.

[0076] A single feature selection rule may be applied to select features, or multiple feature selection rules may be applied to select features. Feature selection rules may be applied in a cascading manner, where they are applied in a specific order and applied to the results of previous rules. For example, a feature generation rule may be applied to the training dataset 210 to generate a first list of features. The final list of features may be analyzed by further feature selection techniques to determine one or more feature sets (e.g., a set of features that can be used to predict viscosity and / or aggregation). Feature sets can be identified using any feature selection technique, such as a filtering method, a wrapping method, and / or an embedding method, using any suitable computational technique. One or more feature sets may be selected according to a filtering method. Filtering methods include, for example, Pearson correlation, linear discriminant analysis, analysis of variance (ANOVA), chi-squared, and combinations thereof. Feature selection according to a filtering method is independent of any machine learning algorithm. Alternatively, features may be selected based on their correlation with an outcome variable, based on scores in various statistical tests.

[0077] As another example, one or more feature sets may be selected by a wrapper method. A wrapper method may be configured to use a subset of features and train a machine learning model using that subset of features. Features may be added to and / or removed from the subset based on inferences drawn from previous models. Wrapper methods include, for example, forward feature selection, backward feature reduction, recursive feature reduction, and combinations thereof. As an example, forward feature selection may be used to identify one or more feature sets. Forward feature selection is an iterative method that starts with no features in the machine learning model. In each iteration, features that best improve the model are added until adding new variables no longer improves the performance of the machine learning model. As an example, backward reduction may be used to identify one or more feature sets. Backward reduction is an iterative method that starts with all features in the machine learning model. In each iteration, the lowest-performing features are removed until no improvement is observed when features are removed. As an example, recursive feature reduction may be used to identify one or more feature sets. Recursive feature reduction is a greedy optimization algorithm that aims to find the feature subset that performs best. Recursive feature reduction iteratively constructs a model, setting aside the best-performing or worst-performing features in each iteration. Recursive feature reduction continues until all features are exhausted, building the next model with remaining features. Finally, recursive feature reduction ranks the features based on the order in which they were removed.

[0078] As a further example, one or more feature sets may be selected by an embedding method. The embedding method combines the qualities of the filtering method and the wrapping method. Embedding methods include, for example, the least absolute shrinkage and selection operator (LASSO) and ridge regression, which implement penalty functions to reduce overfitting. For example, LASSO regression implements L1 regularization, which adds a penalty equivalent to the absolute value of the coefficient magnitude, and ridge regression implements L2 regularization, which adds a penalty equivalent to the square of the coefficient magnitude.

[0079] After the training module 220 generates a feature set, the training module 220 may generate a machine learning-based classification model 240 based on the feature set. A machine learning-based classification model can refer to a complex mathematical model for data classification generated using machine learning techniques. In one example, the machine learning-based classification model 240 may include a map of support vectors representing boundary features. In this example, the boundary features may be selected from and / or represent the highest-ranking features in a given feature set.

[0080] The training module 220 may use feature sets determined or extracted from the training dataset 210 to construct machine learning-based classification models 240A to 240N. In some examples, the machine learning-based classification models 240A to 240N may be combined into a single machine learning-based classification model 240. Similarly, the ML module 230 may represent a single classifier containing one or more machine learning-based classification models 240, and / or multiple classifiers containing one or more machine learning-based classification models 240.

[0081] Features may be combined in classification models trained using machine learning approaches such as discriminant analysis; decision trees; nearest neighbor (NN) algorithms (e.g., k-NN models, replicator NN models, etc.); statistical algorithms (e.g., Bayesian networks, etc.); clustering algorithms (e.g., k-means, mean shifts, etc.); neural networks (e.g., reservoir networks, artificial neural networks, etc.); support vector machines (SVM); logistic regression algorithms; linear regression algorithms; Markov models or chains; principal component analysis (PCA) (e.g., for linear models); multilayer perceptron (MLP) ANN (e.g., for nonlinear models); reservoir network replication (e.g., for nonlinear models, usually for time series); random forest classification; combinations thereof and / or similar. The resulting ML module 230 may include determination rules or mappings for each feature to determine the viscosity and / or aggregation of the antibody.

[0082] In one embodiment, the training module 220 may train a machine learning-based classification model 240 as a convolutional neural network (CNN). The CNN includes three fully connected layers, each consisting of at least one convolutional feature layer and a final classification layer (softmax). The final classification layer may be applied to combine the outputs of the fully connected layers using a softmax function known in the art.

[0083] Features (multiple) and ML Module 230 can be used to predict viscosity and / or aggregation from experimental parameters in a test dataset. In one example, the prediction result for each sequence includes a confidence level corresponding to the likelihood or probability that the computational parameters of the mAb in the test dataset are associated with low / high viscosity and / or low / high aggregation. The confidence level can be a value between 0 and 1. In one example, when there are two states (e.g., low and high), the confidence level may correspond to a value p, which indicates the likelihood that a particular mAb belongs to the first state (e.g., low). In this case, the value 1-p may indicate the likelihood that a particular sequence belongs to the second state (e.g., high). In general, multiple confidence levels may be provided for each mAb in the test dataset, and for each feature if there are three or more states. The feature with the best performance may be determined by comparing the results obtained for each test mAb with known experimental parameters for each test mAb. In general, the feature with the best performance will have results that closely match known yes / no promoter states. The most effective feature can be used to predict the viscosity and / or aggregation state of mAbs.

[0084] Figure 3 is a flowchart illustrating an example training method 300 for generating an ML module 230 using the training module 220. The training module 220 can implement supervised, unsupervised, and / or semi-supervised (e.g., augmentation-based) machine learning-based classification models 240. The method 300 illustrated in Figure 3 is an example of a supervised learning method, and variations of this example of the training method are considered below; however, other training methods can be implemented similarly to train unsupervised and / or semi-supervised machine learning models.

[0085] The training method 300 may determine data (e.g., access, receive, retrieve, etc.) in step 310. The data may include experimental parameters related to the direct measurement of the viscosity of the antibody solution and / or antibody aggregation. The experimental parameters are related to computational parameters related to the corresponding antibody. The computational parameters may be related to the charge values ​​of residues on the corresponding antibody determined via computational modeling.

[0086] The training method 300 may generate a training dataset and a test dataset in step 320. The training dataset and the test dataset may be generated by randomly assigning computational parameters and associated experimental parameters to either the training dataset or the test dataset. In some implementations, the assignment of computational parameters and associated experimental parameters as training or test data may not be entirely random. For example, the training dataset may be generated using the majority of the computational parameters and associated experimental parameters. For instance, the training dataset may be generated using 75% of the computational parameters and associated experimental parameters, and the test dataset using 25%. In another example, the training dataset may be generated using 80% of the computational parameters and associated experimental parameters, and the test dataset using 20%.

[0087] The training method 300 may, in step 330, determine (e.g., extract, select, etc.) one or more features that can be used by the classifier to distinguish between different classifications of viscosity and / or coagulation state (e.g., low and high). As an example, the training method 300 may determine a set of features from computational parameters and associated experimental parameters. In a further example, the set of features may be determined from data different from the computational parameters and associated experimental parameters in either the training dataset or the test dataset. Such computational parameters and associated experimental parameters or other data may be used to determine an initial set of features, which may be further reduced using the training dataset.

[0088] The training method 300 allows one or more machine learning models to be trained in process 340 using one or more features. In one example, the machine learning models may be trained using supervised learning. In another example, other machine learning techniques may be used, including unsupervised and semi-supervised learning. The machine learning models trained in 340 may be selected based on different criteria depending on the problem to be solved and / or the data available in the training dataset. For example, machine learning classifiers may be subject to different degrees of bias. Thus, two or more machine learning models can be trained in 340, optimized, improved, and cross-validated in process 350.

[0089] The training method 300 may select one or more machine learning models to build a predictive model in 360. The predictive model may be evaluated using a test dataset. The predictive model may analyze the test dataset and generate predicted viscosity and / or aggregation states in process 370. The predicted viscosity and / or aggregation states can be evaluated in process 380 to determine whether these values ​​have achieved the desired level of accuracy. The performance of the predictive model can be evaluated in a number of ways based on the classification of a large number of true positives, false positives, true negatives, and / or false negatives of multiple data points shown by the predictive model.

[0090] For example, false positives in a predictive model may refer to the number of times the predictive model incorrectly classified mAbs that are actually high viscosity or highly aggregated as low viscosity or low aggregated. Conversely, false negatives in a predictive model may refer to the number of times the machine learning model classified mAbs that are actually low viscosity or low aggregated as high viscosity or highly aggregated. True negatives and true positives may refer to the number of times one or more mAbs were correctly classified by the predictive model. Related to these measurements are the concepts of recall and precision. Generally, recall refers to the ratio of true positives to the sum of true positives and false negatives, thereby quantifying the sensitivity of the predictive model. Similarly, precision refers to the ratio of true positives to false positives. Once such a desired level of precision is reached, the training period may end and the predictive model (e.g., ML module 230) may be output in step 390; however, if the desired level of precision has not been reached, subsequent iterations of the training method 300 may be initiated in step 310, with variations such as considering a larger collection of sequence data.

[0091] Figure 4 is a diagram of an exemplary process flow for using a machine learning-based classifier to determine whether an mAb is associated with low viscosity and / or low aggregation. As shown in Figure 4, the unclassified computational parameters of mAb 410 may be provided as input to ML module 230. ML module 230 may process the unclassified computational parameters of mAb 410 using a machine learning-based classifier(s) to arrive at a prediction result 420.

[0092] The prediction result 420 may identify one or more properties of the unclassified computational parameters of mAb 410. For example, the classification result 420 may identify the viscosity and / or aggregation state of the unclassified computational parameters of mAb 410 (e.g., whether the mAb has low / high viscosity and / or low / high aggregation). [Examples]

[0093] A. Computer-aided predictive models for protein solution viscosity and aggregation tendency to accelerate drug development. In this study, we developed two predictive models: (1) a predictive model of solution viscosity based on experimental measurement of the viscosity of a mixture of 16 IgG1 and IgG4 antibodies and computational complete antibody homology modeling of the corresponding antibodies; and (2) a predictive model of aggregation tendency based on experimental measurement of high molecular weight (HMW) species formation under accelerated thermal stress and computational antigen-binding fragment (Fab) homology modeling and MD simulation of the corresponding Fab regions. The approach in this study is to adjust the charge of each residue in the homology model by weighting coefficients based on the relative solvent exposure area (SAS) of the exposed residues. With the help of machine learning algorithms, we evaluated the calculated electrostatic and hydrophobic parameters, as well as the conformational changes obtained from the homology model and MD simulation, respectively, to construct robust predictive models for the viscosity and aggregation tendency of protein solutions.

[0094] 1. Method i. Viscosity measurement of protein solutions Sixteen IgG1 and IgG4 antibody solutions (mAb1-mAb16), formulated with a protein concentration of 150 mg / mL, 10 mM histidine buffer, and pH 6.0, were prepared for viscosity measurement. The dynamic viscosity of the solutions was measured using an m-VROC viscometer (Rheosense, San Ramon, California) at a shear rate of 1420 S. -1 The viscosity was measured at a rate of 100 μL / min at 20°C. Three viscosity measurements were recorded over a 100-second period.

[0095] ii. Osmotic second virial coefficient (B 22 )measurement The antibody sample was diluted with the corresponding buffer to reach a final protein concentration of 10 mg / mL. Then, the sample was filtered through a 0.22 μm Millex-GV syringe filter unit (EMD Millipore, Billerica, Massachusetts). Static light scattering at room temperature was measured using a fully automated compositional gradient multi-angle static light scattering (CG-MALS) apparatus with a triple syringe pump Calypso-II sample preparation and delivery unit (Wyatt Technology, Santa Barbara, California). Both light scattering and protein concentration were measured using a Mini Dawn Treos light scattering apparatus (Wyatt Technology, Santa Barbara, California) equipped with a 658 nm laser and an Optilab Rex refractive index detector (Wyatt Technology, Santa Barbara, California). The Rayleigh ratio light scattering intensity was obtained over a protein concentration range of 2 - 8 mg / mL. The data of light scattering and protein concentration were fitted to Equation 1, which is the virial expansion of a non-ideal solution, to estimate the B 22 value.

[0096]

Equation

[0097] R θ is the Rayleigh ratio, M w is the molecular weight, and c is the protein concentration (mg / mL). B 22 represents the osmotic second virial coefficient that remains unrestricted during data fitting. B 22 provides useful insights regarding the intermolecular interactions between protein molecules in a dilute solution. B 22 A negative value of B indicates that the overall interaction between protein molecules is attractive, while a positive value indicates that the overall interaction is repulsive. K in Equation 1 is an optical constant described by Equation 2.

[0098]

Number

[0099] n is the refractive index of the solvent (1.33), N A is Avogadro's number (mol -1 ), dn / dc is the increment of the refractive index of the protein / solvent pair (0.185 mL / g), and λ is the wavelength of the incident light in vacuum.

[0100] iii. Measurement of the diffusion interaction parameter (K D ) The antibody solution was diluted with 10 mM histidine buffer (pH 6.0), and samples of each mAb were prepared at protein concentrations of 10, 5, 2.5, and 0.1 mg / mL. MAb1 was excluded from the K D measurement due to limited availability of the material. The samples were centrifuged at 12,000 × g for 5 minutes and analyzed to remove microbubbles in the solution. Dynamic light scattering (DLS) was measured using a DynaPro plate reader (Wyatt Technology, Santa Barbara, CA). Measurements were carried out 15 times for 15 seconds each, collected, and averaged to determine the diffusion coefficient of each sample. The interaction parameter K D was calculated based on the following equation:

[0101]

Number

[0104] v. Calculated parameters Charge over the entire region of the antibody, including both the variable and constant regions of the light and heavy chains (Z in the complete antibody model) VL , Z VH , Z CL , Z CH1 , Z Hinge , Z CH2 , Z CH3 , Z Total In the Fab model, Z VL , Z VH , Z CL , Z CH1 , Z TotalThe effective charge of each residue was calculated using the procedure described in the section on computational homology modeling. The effective charge of each residue was adjusted by considering the relative solvent exposure area (SAS) of that residue within the homology model of the corresponding antibody. In this approach, depending on the model used, the charge for each residue was multiplied by a weighting coefficient calculated using the SAS of the residue relative to the total SAS of either the complete antibody or Fab. For example, in the variable light chain, the adjusted charge for each residue was calculated using Equation 3, and the total SAS adjusted charge for this region was calculated using Equation 4. These SAS adjusted charges are used in the complete model Z * VL , Z * VH , Z * CL , Z * CH1 , Z * Hinge , Z * CH2 , Z * CH3 , Z * Total In the Fab model, Z * VL , Z * VH , Z * CL , Z * CH1 , Z * Total It is labeled as such.

[0105]

number

[0106] In the formula, i = any residue in the variable light chain (VL), and n = the number of residues in the complete model or Fab model of the specified antibody.

[0107]

number

[0108] In the formula, m = the number of residues in the variable light chain (VL). The hydrophobicity index (HI) of the variable fragment (Fv) is determined using the method described by Sharma et al.

[0109]

number

[0110] The formula is calculated as follows: i represents hydrophobic amino acids, e.g., A, C, F, G, I, L, M, P, V, W, Y; j represents hydrophilic amino acids, e.g., D, E, H, K, N, Q, R, S, T; n is the number of each amino acid; and E is the Eisenberg scale value of each amino acid. As described above, the mean dipole moments of the complete and Fab models were calculated for the protonated structure at pH 6.0. The isoelectric point (pI) of the antibodies used in this study was calculated from both the structural homology model and the sequence. The atom-by-atom aggregation tendency (AP) score is calculated based on the CHARMM force field and SAS patch of exposed hydrophobic residues at a radius of 10 Å. The total aggregation score for each mAb is calculated as the sum of the aggregation scores of all residues in either the complete or Fab model. A comprehensive list of computational parameters for the complete antibody and Fab models can be found in Table 1, and the values ​​of these computational parameters for each mAb are listed in Figures 6A and 6B for the complete antibody model and in Figure 7 for the Fab model.

[0111] Table 1. Correlation coefficients (R) obtained from linear regression between experimentally measured viscosity values ​​and each calculated parameter in the complete antibody model, and between experimentally measured high molecular weight species formation rate (%ΔHMW / day) per day in the Fab model and each calculated parameter. * . JPEG2026094431000012.jpg178170

[0112] i. Mathematical Predictive Modeling of Protein Viscosity The experimental viscosity values ​​and calculated parameters were fed into a stepwise regression algorithm as dependent and independent variables, respectively. This algorithm generates new linear regression models by adding significant parameters and removing non-significant parameters from the list of parameters, and compares the generated models based on the Akaike Information Criterion (AIC), which is an estimator of the relative quality of statistical models for any dataset, based on the number of estimated parameters in the model and the maximum value of the model's likelihood function. Combinations of parameters in Wilkinson notation were also considered to generate more models. The results are shown in their R 2 These are potential predictive models that can be compared together based on their p-values ​​and the mean and standard deviation of the error in the predicted viscosity score (PVS) from experimental values. Furthermore, the models were evaluated using the one-miss cross-validation (LOOCV) method to ensure the models' effectiveness in predicting unpredictable datasets. LOOCV evaluated the robustness of the models in predicting the excluded data point by removing one mAb at a time from the training set. This analysis was repeated 16 times for each model by removing computed and experimental data for a different mAb each time. The mean and standard deviation of the error in the predicted LOOCV viscosity score from experimental values ​​were used as another criterion for comparing the models. The script was developed in the R environment to facilitate the construction of predictive models in a high-throughput automated pipeline.

[0113] ii. Measurement of accelerated thermal stress stability and aggregation dynamics Accelerated stability testing is routinely performed by pharmaceutical companies during formulation development. In this study, the effect of thermal stress on overall stability was evaluated by incubating 14 mAb samples used in viscosity testing at 40°C and 75% relative humidity for 0, 7, 14, and 28 days. Due to limited material availability, mAb9 and mAb16 were excluded from the current dataset for the development of a predictive model for aggregation tendency. The amount of high molecular weight (HMW) species formation was measured using size exclusion chromatography (SEC). %ΔHMW, the relative percentage of HMW formation at 7, 14, and 28 days, was calculated by comparing it to day 0. Furthermore, the rate of %ΔHMW formation per day was calculated based on %ΔHMW obtained by dividing the 28-day data point by 28.

[0114] iii. Molecular Dynamics (MD) Simulation To include the conformational changes of the Fab region in the predictive model of aggregation tendency, MD simulations of the Fab model were used. The atoms of the minimized Fab structures of 14 antibodies were assigned to CHARMM36 force field parameters by structurally matching each residue to its template. These structures were then explicitly solvated in a truncated octahedral box of TIP3P water molecules with the counterion Na. + and Cl - The system was explicitly solvated and neutralized with an ionic concentration of 0.145 mol / L. In each simulation, the system energy was initially minimized using a 1000-step steepest descent algorithm, followed by a further 2000-step minimization using ABNR to remove large strains within the system. The system was gradually heated from 50.0 to 300.0 K at 4 ps in 50 K intervals under a constant volume (NVT) ensemble with a time step of 2.0 fs. Each system was then further equilibrated for 10 ps at a target temperature of 300.0 K and a time step of 2.0 fs under an isotropic pressure of 1.0 bar. Finally, each system was simulated over 2000 ps (i.e., 2.0 ns) at a constant temperature of 300.0 K and a pressure of 1.0 bar with a time step of 2.0 fs.

[0115] Long-range electrostatics were employed using the particle mesh Ewald (PME) method and a 10 Å cutoff distance for van der Waals interactions. The SHAKE algorithm was applied to each simulation, constraining the bond lengths for all hydrogen atoms to allow for a 2.0 fs time step. Three 2.0 ns simulations (6.0 ns in total) could be performed for each system to allow for scrutiny of the reproducibility of the results, differing only in the initial velocity distribution. The trajectories and time-varying atomic coordinates for each simulation were captured every 1.0 ps (i.e., 2000 stereostructures in total for each simulation). The skeletal mean square deviation (RMSD) of the configurational structure relative to the initial structure after rigid alignment in each simulation was calculated as a descriptor of stereostructural stability.

[0116] iv. Mathematical Predictive Modeling of Aggregation Tendencies As previously mentioned, protein aggregation is a collective effect of colloidal stability (i.e., intermolecular interactions) and conformational stability (i.e., changes in protein structure). To construct predictive models of aggregation tendencies, the experimental rate of HMW formation (i.e., %ΔHMW / day), the physical computational parameters of the Fab model (Table 1) as a descriptor of colloidal stability, and the averaged RMSD in MD simulations as a descriptor of conformational stability were used in the same protocol as described in the viscosity section. In summary, the most statistically significant models were created by using a stepwise regression algorithm to correlate the computational parameters of colloid and conformational stability with the measured aggregation dynamics. These models were analyzed using AIC numbers, p-values, and R 2 , adjusted R 2 The antibodies were compared to each other according to the mean and standard deviation of the absolute error, the performance of the LOOCV, and the structural symmetry of the antibody structure.

[0117] 2.Results i. Viscosity of protein solutions Viscosity values ​​were measured for 16 mAbs at a protein concentration of 150 mg / mL. Overall, viscosity values ​​showed a broad distribution ranging from 5.5 to 32.0 cP (Table 2 and Figure 8). According to the dataset, IgG1 antibodies tended to show lower viscosity values ​​compared to IgG4 candidates (Figure 8).

[0118] Table 2. Measured viscosity and calculated predicted viscosity score (PVS) values ​​for the 16 mAbs used in this study. Viscosity values ​​were measured with formulations at a protein concentration of 150 mg / mL, 10 mM histidine buffer, and pH 6.0. PVS values ​​were calculated based on Equation 5. The absolute error between the PVS value and the measured viscosity value was calculated. Absolute errors in PVS and single-out cross-validation (LOOCV) are also shown. JPEG2026094431000013.jpg135170

[0119] ii. B22 and KD show a strong correlation with each other, but no correlation with viscosity. Osmotic pressure secondary virial coefficient (B 22 ) and diffusion interaction parameter (K D ) was measured for 16 and 15 mAbs, respectively. For the current dataset, B 22 The value is -1.461 × 10 -05 ~2.939 × 10 -04 mol ml g -2 It changes between K D The values ​​vary between -11.604 and 61.114 mL / g (Figures 9A-9C). Based on the dataset, the IgG1 antibody showed a higher positive B compared to the IgG4 candidate. 22 Value and K D It tends to show a value. B 22 Value and K D The values ​​are strongly correlated with each other for the current dataset, with a linear correlation coefficient (R) of 0.99 (Figures 9A-9C). This observation is consistent with previous research published by other researchers in the same field.

[0120] B 22 and K DBoth measurements were measures of pairwise interactions, which were predominantly dominant at dilution concentrations. However, as the concentration increased, higher-level interactions involving multiple molecules also contributed significantly to solution viscosity. Therefore, B measured in the dilution solution 22 Value and K D The value is not a direct measure of protein-protein interactions at high concentrations. However, the literature discusses the effectiveness of using B22 and Kd values ​​at dilution concentrations as predictors of viscosity at high protein concentrations. To evaluate this, B 22 or K D A linear correlation was determined between either of the measured viscosity values ​​(Figures 10A-10B). Based on the dataset, B 22 or K D There is a directional decreasing trend (i.e., a negative correlation) between either of these and the viscosity value, but there is no strong correlation (Figures 10A-10B). Therefore, B 22 and K D This is insufficient for predicting the viscosity of protein solutions when mAb concentrations increase.

[0121] iii. Each single selected computational parameter contributes to the prediction of the overall viscosity value. Table 1 shows the linear correlation coefficient (R) values ​​obtained from the regression line between measured viscosity values ​​and calculated parameters from a complete antibody homology model of 16 mAbs. Furthermore, Figures 11A–11C show plots of the linear relationship between experimental viscosity values ​​and calculated parameters. The correlation R values ​​vary between -0.68 and 0.54. The calculated parameters (i.e., Z) used in the final prediction model are shown. * VL , Z * CL , Z * Hinge , Z * CH2 , Z * CH3The parameters (HI) were selected based on the stepwise protocol and the structural symmetry of the antibody, as described in the section on mathematical modeling. While each of these parameters contributes to the prediction of the overall viscosity value, it is not possible to predict the viscosity value with a single computational parameter alone, as observed at moderate R values. This was actually expected, as the nature of viscosity is a multivariate phenomenon involving intermolecular interactions defined by hydrophobic and electrostatic properties in various regions.

[0122] iv. Predictive models for protein viscosity: Predictive viscosity score (PVS) The final predictive model for protein viscosity was selected based on a stepwise protocol and chosen calculated parameters, as described in the previous section. The predicted viscosity score (PVS) is the V of the complete antibody model. L area, C L Region, hinge region, C H 2 regions and C H This is a predictive model of protein viscosity that takes into account the solvent exposure area (SAS) for three regions, the adjusted charge, and the hydrophobicity of the variable region (Equation 5).

[0123] Table 3 shows the constants C0 to C6 in the PVS model.

[0124]

number

[0125] Table 3: Constant coefficients and predicted viscosity score (PVS) in Equation 5. JPEG2026094431000015.jpg63170

[0126] The PVS model is R 2 0.884, adjusted R 2The R-value is 0.807. The correlation R-value of 0.94 indicates a strong correlation between PVS and measured viscosity values ​​in the current dataset, and the adjusted R-value of 0.90 indicates that the model fits any dataset by considering the number of parameters in the model without concerns about data overfitting. The p-value of the PVS model is 0.0009, providing strong evidence against the null hypothesis that the PVS model cannot predict viscosity values ​​with 95% confidence (p-value < 0.05). The mean absolute error of 2.7 and standard deviation of 1.8 observed between PVS and experimental values ​​demonstrate the validity of this model. The minimum and maximum residuals between PVS and measured viscosity are -7.9 and 3.9, respectively (Table 2). The linear regression line between the calculated PVS values ​​and measured viscosity values ​​shows that the majority of data points fall within the 95% confidence interval (Figure 12). More importantly, the PVS model performs well in LOOCV analysis, with a mean absolute error of 4.6 and a standard deviation of 2.3 between the LOOCV-based PVS and experimental viscosity values ​​(Table 2). 2 The values ​​are in the range of 0.863 to 0.925, and the adjusted R 2 The values ​​ranged from 0.760 to 0.868. These results confirm that PVS represents a statistically significant predictive model between the selected calculation parameters and experimental viscosity values.

[0127] v. Results of thermal stress stability Fourteen mAbs were incubated at 40°C and 75% relative humidity. Figure 13 shows representative SEC signals for mAb3 over incubation periods of 0 and 28 days, as well as the increase in HMW peaks resulting from aggregate formation. For the 14 mAbs used in this study, the relative percentage of HMW formation at 7, 14, and 28 days compared to day 0, %ΔHMW, is shown in Figure 14. Longer incubation times resulted in the formation of more aggregates (Figure 14). The rate of %ΔHMW formation per day, calculated based on %ΔHMW obtained by dividing the 28-day data points by 28, ranged from 0.0564 to 0.1600 (Table 4 and Figure 15). IgG1 antibody molecules, which tend to have lower viscosity values, tended to have higher average %ΔHMW / day compared to IgG4 candidates in the current dataset (Figure 15).

[0128] Table 4. Measured daily high molecular weight species formation rate (%ΔHMW / day) and calculated predicted aggregation score (PAS) values ​​for the 14 mAbs used in this study. mAb9 and mAb16 were excluded from the current dataset for the development of a predictive model for aggregation tendency due to limited material availability. %ΔHMW / day values ​​were calculated after 28 days of incubation at 40°C and 75% relative humidity. PAS values ​​were calculated based on Equation 6. The absolute error between the PAS value and the measured %ΔHMW / day value was calculated. Absolute errors in PAS and single-out cross-validation (LOOCV) are also shown. JPEG2026094431000016.jpg118170

[0129] vi. Further verification of PVS To further test the predictive capabilities of the PVS model (Equation 5), four IgG1 and IgG4 mAbs (Table 5) that were not part of the 16 mAbs used in the development of the predictive model were evaluated. The viscosity of these four mAbs was measured using the same formulation as described above, with the same protocol. The measured viscosity values ​​ranged from 4.1 to 22.0 cP (Table 5). The structures of these four mAbs were modeled according to the protocol described in this work. Hydrophobic and electrostatic parameters were calculated for use in the PVS model (Table 5). The absolute error between PVS and experimental viscosity values ​​ranged from 2.8 to 5.6 (Table 5), demonstrating that the PVS model can accurately predict the viscosity values ​​of mAbs not included in the training dataset.

[0130] Table 5: For the development of the viscosity prediction model, calculation parameters for four mAbs (not part of the 16 mAbs), measured viscosity values, and predicted viscosity scores (PVS) were used. * VL , Z * CL , Z * Hinge , Z * CH2 and Z * CH3 These are, respectively, V L area, C L Region, hinge region, C H PVS is the adjusted charge of the solvent-exposed area for two regions, where HI is the hydrophobicity index. PVS values ​​were calculated based on Equation 5, and viscosity values ​​were measured for formulations with a protein concentration of 150 mg / mL, 10 mM histidine buffer, and pH 6.0. The absolute error between the PVS value and the measured viscosity value was calculated. JPEG2026094431000017.jpg59170

[0131] vii.MD Simulation Results Three separate MD simulations were performed for each Fab model of the 14 mAbs to assess the consistency of observations. The RMSD of the configurational structure relative to the initial structure of each mAb was plotted over a simulation time of 2.0 ns (Figures 16A-16C) to demonstrate the reproducibility of the overall simulation for each mAb. For each mAb, the RMSD value at each time point was averaged across the three simulations, and the averaged RMSD was plotted against a simulation time of 2.0 ns in Figure 17. Furthermore, Figures 18A-18C show the averaged RMSD for each mAb in separate plots. The averaged RMSD values ​​across the three simulations for each mAb were averaged over the final 1.5 ns to obtain a single number as the averaged RMSD for each mAb. For the Fab regions of the 14 mAbs used in this study, the average RMSD ranged from 1.785 to 3.159 Å (Figure 7).

[0132] viii. Each single selected computational parameter contributes to the prediction of the overall cohesive trend. The calculated parameters from the Fab homology model and the RMSD values ​​from the MD simulations are descriptors for colloidal stability and stereostructural stability, respectively. Table 1 shows the linear correlation coefficient (R) values ​​obtained from the regression lines between the measured %ΔHMW / day values ​​and the calculated parameters or RMSD values, obtained from MD simulations of 14 mAbs. Furthermore, Figures 19A–19C show plots of the linear relationship between experimental %ΔHMW / day values ​​and calculated parameters, including RMSD. The correlation R values ​​vary between -0.59 and 0.72. The final predicted model (i.e., Z) is shown. * VL , Z * CH1 RMSD, HI, D Fab , and PI SequenceThe calculated parameters used were selected based on the stepwise protocol and structural symmetry of the Fab region described in the Mathematical Modeling section. Each of these parameters contributes to the prediction of the overall %ΔHMW / day value, but as observed with moderate R values, a single calculated parameter cannot predict the %ΔHMW / day value alone. This was expected as a result of the aggregation nature being a multivariate phenomenon involving both colloidal and stereostructural stability conditions.

[0133] ix. Predictive model for aggregation tendency: Predictive aggregation score (PAS) The final predictive model for aggregation tendency was selected based on a stepwise protocol and selected calculated parameters, as described in the Test Methods section. The Predictive Aggregation Score (PAS) is a predictive model of protein aggregation dynamics that includes both calculated descriptors of the colloidal and three-dimensional structures of the Fab region. This model is V L Region and C H1 The SAS-adjusted charge for the region, the averaged RMSD of the steric change from the initial structure, the hydrophobicity of the variable region, the dipole moment of the Fab region, and the isoelectric point of the antibody obtained from its sequence were considered (Equation 6). The constants C0 to C6 in the PAS model are shown in Table 6.

[0134]

number

[0135] Table 6: Constant coefficients and predicted agglomeration score (PAS) in Equation 6. JPEG2026094431000019.jpg67170

[0136] The PAS model is R 2 0.883, adjusted R 2The R-value is 0.782. The correlation R-value of 0.94 indicates a strong correlation between PAS and the measured %ΔHMW / day values ​​in the current dataset, and the adjusted R-value of 0.88 indicates that the model fits the arbitrary dataset well by considering the number of parameters in the model without concerns about data overfitting. The p-value of the PAS model is 0.0057, providing strong evidence against the null hypothesis that the PAS model cannot predict %ΔHMW / day values ​​with 95% confidence (p-value < 0.05). The mean absolute error of 0.0084 and standard deviation of 0.0083 observed between PAS and experimental values ​​demonstrate the effectiveness of this model. The minimum and maximum residual errors between PAS and the measured %ΔHMW / day are -0.0290 and 0.0216, respectively (Table 3).

[0137] The linear regression line between the calculated PAS values ​​and the measured %ΔHMW / day values ​​shows that the majority of the data points fall within the 95% confidence interval (Figure 20). More importantly, the PAS model performs well in LOOCV analysis, with a mean absolute error of 0.0173 and a standard deviation of 0.0151 between the PVS based on the LOOCV method and the experimental %ΔHMW / day (Table 4). During LOOCV, R 2 The value is in the range of 0.858 to 0.949, and the adjusted R 2 The values ​​ranged from 0.716 to 0.897. These results confirm that PAS represents a statistically significant predictive model between the calculation parameters for the selected colloidal and stereostructures and the experimental %ΔHMW values.

[0138] 3. Essay The production of monoclonal antibodies is increasing in the pipelines of biopharmaceutical companies. The shift from IV to SC administration presents challenges in the formulation development of these candidates, which require higher potency doses due to viscosity and mAb aggregation issues. More robust methods should be developed to enable early prediction of viscosity and aggregation tendencies in formulation development and drug discovery. A combination of computer-aided tools and experimental techniques is promising for addressing these challenges. Here, a scheme was developed that utilizes the capabilities of homology modeling and MD simulation to generate predictive models of the viscosity and aggregation tendencies of antibody solutions. These models, PVS and PAS, allow for the comparison of different antibodies together in drug development and even in the early stages of drug discovery, without the need for physical materials.

[0139] Compared to similar predictive models in the literature regarding the viscosity and aggregation tendency of protein solutions, the PVS and PAS models developed in this study have fewer predictive errors and are R 2 Value, adjusted R 2 The reliability is demonstrated by analyzing the values, p-values, absolute error values, and LOOCV analysis. The robustness of the PVS and PAS models is obtained by utilizing novel computational parameters that adjust the charge distribution for each residue in different regions of the antibody for SAS, and by considering the structural symmetry. In the PVS model, the calculated electrostatic and hydrophobic parameters from the antibody homology model are considered to reflect intramolecular and intermolecular interactions in the protein solution.

[0140] This is the first time that both colloidal stability parameters and stereostructural stability parameters have been used in a predictive model of aggregation tendency via whole-atom MD simulation. Therefore, the PAS model can predict aggregation tendency more realistically by considering stability at the atomic level. It goes without saying that the constant coefficients of the PVS and PAS predictive models developed for viscosity and aggregation tendency are specific to the buffer systems and respective protein concentrations used in this study. However, the overall scheme and calculated parameters can be extended to other buffer systems and protein concentrations.

[0141] The robustness and accuracy of predictive models in machine learning algorithms and statistical methods depend on the number of data points in the training and validation datasets. However, limitations in the nature of experimental viscosity, cohesiveness measurements, and the availability of physical materials limit researchers in this field's ability to acquire a large number of data points. Therefore, all data points were used as the training dataset, and LOOCV analysis was performed to evaluate the predictability of the models. The PVS and PAS models were developed based on viscosity and cohesiveness measurements of 16 mAbs and 14 mAbs, respectively. The robustness and accuracy of these predictive models can be improved by adding more data points measured under the same conditions to the datasets developed for these models. The new data points can also be used as validation datasets to further evaluate the predictability of the PVS and PAS models. Furthermore, in this study, a stepwise regression algorithm was used to generate predictive models from the calculated parameters. Further statistical and machine learning algorithmic techniques, such as LASSO (least absolute shrinkage and selection operator) regression and random forest regression, can be explored to develop more robust models.

[0142] As mentioned earlier, MD simulations were performed on the Fab region of the antibody to consider the conformational stability in the aggregation tendency model. The protocol described in this work can be extended to perform MD simulations on complete antibodies to generate calculated conformational parameters for the complete antibody model. These parameters may improve the effectiveness and reliability of the aggregation tendency model. Furthermore, whole-atom and coarse-grained MD simulations of complete antibodies can provide more clues about intramolecular and intermolecular interactions of antibody molecules. As an example, Cloutier et al. analyzed the effect of excipients on aggregation and viscosity through whole-atom MD simulations of three IgG1 mAbs. Kastelic et al. performed coarse-grained MD simulations to evaluate the binding interactions of fragment antigens (Fab-Fab) or crystallizable fragments (Fab-Fc) and proposed a strategy to control the viscosity of antibody solutions through control of their binding sites.

[0143] In the current test, each mAb was dissolved in a truncated octahedron water box, minimizing the number of water atoms to prioritize computationally intensive simulations. Even with this scheme, each solvated complete antibody model or Fab model consists of approximately 275,000 or 51,000 atoms, respectively. Such large systems require significant computing power, potentially limiting simulation time to available infrastructure. Advances in graphics processing units (GPUs) enable longer MD simulations for Fab regions and complete antibodies. Longer MD simulations can provide more clues about the three-dimensional structure and inherent instability of antibodies in solution, potentially leading to a better understanding of intramolecular and intermolecular interactions. Furthermore, other schemes, such as rectangular water boxes, can be used to account for the interaction of more water molecules with specific antibodies. The box shape used in the simulation can influence the dynamic behavior of proteins and computational properties.

[0144] Furthermore, all MD simulations in this study were performed in aqueous solution. This is the most common approach in MD simulations of biopharmaceuticals, as most force fields are optimized for water interactions. However, performing MD simulations in a buffer environment similar to that of the final drug can yield computational parameters that are more relevant to experimental viscosity and stability measurements. When ranking antibodies, as long as all mAb candidates are treated in the same buffer system, they can be used in predictive models and compared to one another.

[0145] B. Model Verification Validation experiments were conducted on both PVS (Predicted Viscosity Score) and PAS (Predicted Agglomeration Score) models and algorithms using a set of 10 mAbs. The viscosity and agglomeration prediction algorithms were validated using statistical correlations between the experimental dataset and the predicted scores.

[0146] 1. Test Method For 10 additional mAbs, the predictive model was validated using experimental data of dynamic viscosity (cP) and total cohesiveness at 40°C. The experimental data was blinded to users to remove known and unknown biases. The predicted viscosity and cohesiveness scores were then compared to the experimental data and correlated using a linear regression model. Given the small size of the dataset, a correlation score greater than 0.75 was considered acceptable.

[0147] Figure 21 shows model validation using a test set (data-blinded correlation) for dynamic viscosity. The PVS (Equation 5) and PAS (Equation 6) models were validated using data from 10 mAbs (a mixture of IgG1 and IgG4). To ensure unbiased validation, the data was blinded to users. The color coding for risk ranking (separated by dashed lines and labeled by color) is based on past development goals across the biopharmaceutical industry and does not reflect any regulatory requirements. JPEG2026094431000020.jpg68170

[0148] Figure 22 shows a model validation of high molecular weight (HMW) species formation using a test set (data-blinded correlation). %ΔHMW was calculated as follows:

[0149]

number

[0150] The PVS and PAS models were validated using data from 10 mAbs (a mixture of IgG1 and IgG4). To ensure unbiased validation, the data was blinded to users.

[0151] The color coding used for risk ranking (separated by dashed lines and labeled with colors) is based on past development goals across the biopharmaceutical industry and does not reflect any regulatory requirements. JPEG2026094431000022.jpg64170

[0152] 2.Results Both the aggregation and viscosity prediction scores showed a high correlation with the validation experimental data (R² value greater than 0.8). This strong statistical correlation further improves the reliability of both the prediction model and its underlying AI algorithm.

[0153] Figure 23 is a block diagram depicting an environment 2300 including a non-limiting example of computing devices 2301 and servers 2302 connected via a network 2304. In one embodiment, some or all steps of any described method can be performed on the computing devices described herein. The computing device 2301 may comprise one or more computers configured to store one or more of the following: experimental data 2320, computationally derived data 2322, prediction modules 2326 (e.g., ML modules 230, including any auxiliary training modules). The server 2302 may comprise one or more computers configured to store experimental data 2320 and / or computationally derived data 2322. Multiple servers 2302 can communicate with the computing device 2301 via the network 2304. In one embodiment, the server 2302 may comprise a repository for data generated by one or more experiments.

[0154] The computing device 2301 and server 2302 may be digital computers, with respect to their hardware architecture, generally including a processor 2308, a memory system 2310, an input / output (I / O) interface 2312, and a network interface 2314. These components (2308, 2310, 2312, and 2314) are communicatively connected via a local interface 2316. The local interface 2316 may be, for example, one or more buses or other wired or wireless connections known in the art, but is not limited thereto. The local interface 2316 may have additional elements (omitted for simplicity) to enable communication, such as controllers, buffers (caches), drivers, repeaters, and receivers. Furthermore, the local interface may include address, control, and / or data connections to enable appropriate communication between the aforementioned components.

[0155] The processor 2308 may be a hardware device for executing software, particularly stored in the memory system 2310. The processor 2308 can be any custom-made or commercially available processor, a central processing unit (CPU), an auxiliary processor among several processors associated with the computing device 2301 and server 2302, a semiconductor-based microprocessor (in the form of a microchip or chipset), or any device in general for executing software instructions. When the computing device 2301 and / or server 2302 are operating, the processor 2308 may be configured to execute software stored in the memory system 2310 to communicate data to and from the memory system 2310, and to generally control the operation of the computing device 2301 and server 2302 according to the software.

[0156] The I / O interface 2312 can be used to receive user input from and / or provide system output to one or more devices or components. User input may be provided, for example, via a keyboard and / or mouse. System output may be provided via a display device and / or printer (not shown). The I / O interface 2312 may include, for example, a serial port, a parallel port, a Small Computer System Interface (SCSI), an infrared (IR) interface, a radio frequency (RF) interface, and / or a Universal Serial Bus (USB) interface.

[0157] The network interface 2314 can be used to send and receive data from the computing device 2301 and / or from the server 2302 on the network 2304. The network interface 2314 may include, for example, a 10BaseT Ethernet adapter, a 100BaseT Ethernet adapter, a LAN PHY Ethernet adapter, a Token Ring adapter, a wireless network adapter (e.g., WiFi, cellular, satellite), or any other suitable network interface device. The network interface 2314 may include address, control, and / or data connections to enable proper communication on the network 2304.

[0158] The memory system 2310 may include one or a combination of volatile memory elements (e.g., random access memory (RAM such as DRAM, SRAM, SDRAM, etc.)) and non-volatile memory elements (e.g., ROM, hard drive, tape, CD-ROM, DVD-ROM, etc.). Furthermore, the memory system 2310 may incorporate electronic, magnetic, optical, and / or other types of storage media. It should be noted that the memory system 2310 may have a distributed architecture in which various components are located apart from each other but can be accessed by the processor 2308.

[0159] The software in the memory system 2310 may include one or more software programs, each of which includes an ordered list of executable instructions for performing a logical function. In the example in Figure 23, the software in the memory system 2310 of the computing device 2301 may include experimental data 2320, computationally derived data 2322, a prediction module 2326, and a preferred operating system (O / S) 2318. In the example in Figure 23, the software in the memory system 2310 of the server 2302 may include experimental data 2320, computationally derived data 2322, and a preferred operating system (O / S) 2318. The operating system 2318 essentially controls the execution of other computer programs and provides scheduling, input-output control, file and data management, memory management, and communication control, as well as related services.

[0160] For illustrative purposes, application programs and other executable program components, such as the operating system 2318, are illustrated herein as separate blocks, but it is recognized that such programs and components may exist at different times within different storage components of the computing device 2301 and / or server 2302. Implementations of the prediction module 2326 may be stored on or transmitted on some form of computer-readable medium. Any of the methods of this disclosure can be executed by computer-readable instructions embodied on a computer-readable medium. The computer-readable medium can be any available medium accessible by a computer. For example, and not intended to limit, the computer-readable medium may include “computer storage medium” and “communication medium.” The “computer storage medium” may include volatile and non-volatile removable and non-removable media, implemented by any method or technique for storing information, such as computer-readable instructions, data structures, program modules, or other data. Exemplary computer storage media may include RAM, ROM, EEPROM, flash memory or other storage technologies, CD-ROM, digital versatile disk (DVD) or other optical storage devices, magnetic cassettes, magnetic tapes, magnetic disk storage devices or other magnetic storage devices, or any other media that can be used to store desired information and can be accessed by a computer.

[0161] In one embodiment, the prediction module 2326 may be configured to perform method 2400 as shown in Figure 24. Method 2400 may be performed entirely or partially by a single computing device, multiple electronic devices, and so on. Method 2400 may include determining experimental data associated with one or more monoclonal antibodies (mAbs) in 2410. One or more mAbs may include one or more IgG1 antibodies or IgG4 antibodies. The experimental data may include experimental viscosity data. The experimental viscosity data may include one or more dynamic viscosity values ​​or kinematic viscosity values.

[0162] Determining experimental data related to one or more mAbs may involve measuring at least one of either dynamic viscosity values ​​or kinematic viscosity values ​​based on each solution and viscometer for one or more mAbs.

[0163] Experimental data may include experimental agglutination data. Experimental agglutination data may include high molecular weight (HMW) species formation data for each of one or more mAbs. Determining experimental data related to one or more mAbs may include measuring the amount of HMW species formation over time based on each of the one or more mAbs' solutions and size exclusion chromatography (SEC).

[0164] In 2420, computational derivation data related to one or more mAbs is determined, and the computational derivation data includes one or more computational parameters weighted based on the exposed area (ASA) of one or more residues of one or more mAbs. The computational derivation data may include charge data related to one or more regions related to the sequences of one or more mAbs, modified charge data related to one or more regions based on the solvent exposed area of ​​residues in the homology model of one or more mAbs, hydrophobicity index (HI), dipole moment, or isoelectric point (pI). Determining computational derivation data related to one or more mAbs may include complete antibody homology modeling of the sequences of one or more mAbs, or Fab region modeling of the antigen-binding fragment (Fab) sequences of one or more mAbs.

[0165] Determining computationally derived data associated with one or more mAbs may include: determining one or more charge values ​​associated with one or more residues in one or more regions of one or more mAbs based on a homology model of one or more mAbs; determining the solvent exposure area (SAS) of one or more residues in one or more regions based on a homology model of one or more mAbs; adjusting one or more charge values ​​associated with one or more residues to the total SAS associated with one or more mAbs based on weighting coefficients calculated using the SAS of one or more residues; and determining the charge values ​​associated with each of the one or more regions based on the homology model of one or more mAbs and the adjusted charge values ​​associated with one or more residues.

[0166] The computationally derived data may include charge data associated with one or more regions related to the sequence of one or more mAbs, modified charge data associated with one or more regions based on the solvent exposure area of ​​residues in the homology model of one or more mAbs, a hydrophobicity index (HI), dipole moment, isoelectric point (pI), aggregation tendency (AP), or a descriptor of steric stability. The steric stability descriptor may include the mean square deviation (RMSD) of the skeletal structure relative to the initial structure after rigid alignment. Determining the computationally derived data associated with one or more mAbs may include one or more molecular dynamics (MD) simulations associated with one or more mAbs.

[0167] In step 2430, multiple candidate predictive models are determined based on experimental and computationally derived data. Determining multiple candidate predictive models based on experimental and computationally derived data may include identifying one or more experimental parameters of the experimental data as dependent variables, identifying one or more computational parameters of the computationally derived data as independent variables, and determining multiple candidate predictive models based on the dependent and independent variables using a stepwise regression algorithm.

[0168] At 2440, an optimal prediction model is determined from a plurality of candidate prediction models. Determining an optimal prediction model from a plurality of candidate prediction models may include determining an Akaike Information Criterion (AIC) score for each candidate prediction model of the plurality of candidate prediction models, and determining, as the optimal prediction model, the candidate prediction model of the plurality of candidate prediction models associated with the highest AIC score.

[0169] Determining an optimal prediction model from a plurality of candidate prediction models may include determining, as the optimal prediction model, the candidate prediction model of the plurality of candidate prediction models associated with the minimum error when predicting the viscosity score of an mAb excluded from experimental data and computationally derived data.

[0170] Determining an optimal prediction model from a plurality of candidate prediction models may include determining, as the optimal prediction model, the candidate prediction model of the plurality of candidate prediction models associated with the minimum error when predicting the aggregation score of an mAb excluded from experimental data and computationally derived data.

[0171] At 2450, the optimal prediction model is output. Method 2400 may also include receiving computationally derived data related to a query mAb, providing the computationally derived data to the optimal prediction model, and determining a viscosity score related to the query mAb based on the optimal prediction model. Method 2400 may include adjusting an appropriate formulation composition or protein engineering strategy to mitigate specific issues of a drug candidate under development, such as adjusting the amount of a viscosity reducing agent in a solution related to the query mAb, based on the viscosity score.

[0172] Method 2400 may also include receiving computationally derived data related to a query mAb, providing the computationally derived data to the optimal prediction model, and determining an aggregation score based on the optimal prediction model.

[0173] In one embodiment, the prediction module 2326 may be configured to execute a method 2500 as shown in FIG. 25. The method 2500 may be implemented in whole or in part by a single computing device, multiple electronic devices, and the like. The method 2500 may include, at 2510, receiving computationally derived data related to a monoclonal antibody (mAb). The computationally derived data may include viscosity data derived from calculations. The viscosity data derived from calculations may include one or more of a dynamic viscosity value or a kinematic viscosity value.

[0174] At 2520, provide the computationally derived data to a prediction model. At 2530, determine a viscosity score related to the mAb based on the prediction model. The method 2500 may also include adjusting an appropriate formulation composition or protein engineering strategy to reduce specific challenges of a drug candidate under development based on the viscosity score, for example, adjusting the amount of a viscosity reducing agent in a solution related to the query mAb.

[0175] The method 2500 may also include receiving sequence data related to the mAb and determining computationally derived data based on the sequence data. The method 2500 may also include receiving computationally derived data related to a query mAb, providing the computationally derived data to an optimal prediction model, and determining a viscosity score related to the query mAb based on the optimal prediction model.

[0176] In one embodiment, the prediction module 2326 may be configured to perform method 2600 as shown in Figure 26. Method 2600 may be performed entirely or partially by a single computing device, multiple electronic devices, and so on. Method 2600 may include receiving computationally derived data related to a monoclonal antibody (mAb) in 2610. The computationally derived data may include computationally derived agglutination data. The computationally derived agglutination data may include high molecular weight (HMW) species formation data for the mAb. The computationally derived data may include charge data related to one or more regions related to the sequence of the mAb, modified charge data related to one or more regions based on the solvent exposure area of ​​residues in the homology model of the mAb, a descriptor of hydrophobicity index (HI), dipole moment, isoelectric point (pI), agglutination tendency (AP), or stereostructural stability.

[0177] In 2620, the predictive model is provided with computationally derived data. At 2630, the aggregation score associated with mAb is determined based on the predictive model. Method 2600 may also include receiving sequence data related to the mAb and determining computationally derived data based on the sequence data.

[0178] Method 2600 may also include determining the optimal predictive model from among several candidate predictive models related to the minimum error in predicting the aggregation score associated with mAbs. Method 2600 may also include receiving computationally derived data related to a query mAb, providing the computationally derived data related to the query mAb to an optimal predictive model, and determining an agglomeration score related to the query mAb based on the optimal predictive model.

[0179] In light of the apparatus, systems, and methods described herein and their variations thereof, specific embodiments of the invention described more specifically below. However, these particularly enumerated embodiments should not be construed as having any limiting effect on any different claims, including different or more general teachings described herein, nor should the “specific” embodiments be construed as being limited in any way other than the inherent meaning of the language used therein.

[0180] Embodiment 1: A method comprising determining experimental data associated with one or more monoclonal antibodies (mAbs), determining computationally derived data associated with one or more mAbs, wherein the computationally derived data includes one or more computational parameters weighted based on the exposed area (ASA) of one or more residues of one or more mAbs, determining a plurality of candidate predictive models based on the experimental data and the computationally derived data, determining the optimal predictive model from the plurality of candidate predictive models, and outputting the optimal predictive model.

[0181] Embodiment 2: The embodiment according to Embodiment 1, wherein one or more mAbs contain one or more IgG1 antibodies or IgG4 antibodies. Embodiment 3: An embodiment according to any one of Embodiments 1 to 2, wherein the experimental data includes experimental viscosity data.

[0182] Embodiment 4: An embodiment according to any one of Embodiments 1 to 3, wherein the experimental viscosity data includes one or more of the dynamic viscosity values ​​or kinematic viscosity values. Embodiment 5: An embodiment according to any one of Embodiments 1 to 4, wherein determining experimental data related to one or more mAbs includes measuring at least one of a dynamic viscosity value or a kinematic viscosity value based on each solution and viscometer of one or more mAbs.

[0183] Embodiment 6: An embodiment according to any one of Embodiments 1 to 5, wherein the calculated derived data includes charge data associated with one or more regions associated with the sequence of one or more mAbs, modified charge data associated with one or more regions based on the solvent exposure area of ​​residues in the homology model of one or more mAbs, a hydrophobicity index (HI), a dipole moment, or an isoelectric point (pI).

[0184] Embodiment 7: An embodiment according to any one of Embodiments 1 to 6, wherein determining the computationally derived data relating to one or more mAbs includes complete antibody homology modeling of the sequences of one or more mAbs, or Fab region modeling of the antigen-binding fragment (Fab) sequences of one or more mAbs.

[0185] Embodiment 8: An embodiment according to any one of Embodiments 1 to 7, wherein determining computationally derived data related to one or more mAbs includes: determining one or more charge values ​​related to one or more residues in one or more regions of one or more mAbs based on a homology model of one or more mAbs; determining the solvent exposure area (SAS) of one or more residues in one or more regions based on a homology model of one or more mAbs; adjusting one or more charge values ​​related to one or more residues based on weighting coefficients calculated using the SAS of one or more residues to the total SAS related to one or more mAbs; and determining the charge value related to each region of one or more regions based on a homology model of one or more mAbs and the adjusted charge values ​​related to one or more residues.

[0186] Embodiment 9: An embodiment according to any one of Embodiments 1 to 8, wherein determining multiple candidate predictive models based on experimental data and computationally derived data includes identifying one or more experimental parameters of the experimental data as dependent variables, identifying one or more computational parameters of the computationally derived data as independent variables, and determining multiple candidate predictive models based on a stepwise regression algorithm, based on the dependent variables and based on the independent variables.

[0187] Embodiment 10: An embodiment according to any one of Embodiments 1 to 9, wherein determining the optimal prediction model from a plurality of candidate prediction models includes determining an Akaike Information Criterion (AIC) score for each of the plurality of candidate prediction models, and determining the candidate prediction model of the plurality of candidate prediction models that is associated with the highest AIC score as the optimal prediction model.

[0188] Embodiment 11: An embodiment according to any one of Embodiments 1 to 10, wherein determining the optimal prediction model from a plurality of candidate prediction models includes determining a candidate prediction model among a plurality of candidate prediction models that is associated with the minimum error when predicting the viscosity score of mAbs excluded from experimental data and computationally derived data as the optimal prediction model.

[0189] Embodiment 12: An embodiment according to any one of Embodiments 1 to 11, further comprising receiving computationally derived data related to a query mAb, providing the computationally derived data to an optimal predictive model, and determining a viscosity score related to the query mAb based on the optimal predictive model.

[0190] Embodiment 13: An embodiment of Embodiment 12, further comprising adjusting an appropriate formulation composition or protein engineering strategy to mitigate specific challenges of a drug candidate under development, such as adjusting the amount of viscosity-reducing agent in a solution related to a query mAb, based on a viscosity score.

[0191] Embodiment 14: An embodiment according to any one of Embodiments 1 to 13, wherein the experimental data includes experimental aggregation data. Embodiment 15: The embodiment according to Embodiment 14, wherein the experimental aggregation data includes high molecular weight (HMW) species formation data for each of the one or more mAbs.

[0192] Embodiment 16: The embodiment according to any one of Embodiments 1 to 15, wherein determining experimental data related to one or more mAbs includes measuring the amount of HMW species formation over time based on each solution of the one or more mAbs and size exclusion chromatography (SEC).

[0193] Embodiment 17: The embodiment according to any one of Embodiments 1 to 16, wherein determining experimental data related to one or more mAbs includes measuring the amount of HMW species formation over time based on each solution of the one or more mAbs and size exclusion chromatography (SEC).

[0194] Embodiment 18: The embodiment according to any one of Embodiments 1 to 17, wherein the descriptor of conformational stability includes the root mean square deviation (RMSD) of the backbone of the conformational structure with respect to the initial structure after rigid body alignment.

[0195] Embodiment 19: The embodiment according to any one of Embodiments 1 to 18, wherein determining computationally derived data related to one or more mAbs includes performing one or more molecular dynamics (MD) simulations related to the one or more mAbs.

[0196] Embodiment 20: The embodiment according to any one of Embodiments 1 to 19, wherein determining the optimal prediction model from a plurality of candidate prediction models includes determining, as the optimal prediction model, a candidate prediction model of the plurality of candidate prediction models that is related to the minimum error when predicting the aggregation score of an mAb excluded from the experimental data and the computationally derived data.

[0197] Embodiment 21: The embodiment according to any one of Embodiments 1 to 20, further comprising receiving computationally derived data related to the query mAb, providing the computationally derived data to the optimal prediction model, and determining an aggregation score based on the optimal prediction model.

[0198] Embodiment 22: An embodiment of any one of Embodiments 1 to 21, further comprising adjusting an appropriate formulation composition or protein engineering strategy to mitigate specific challenges of a drug candidate under development, such as adjusting the amount of anti-aggregation agent in a solution associated with a query mAb, based on the aggregation score.

[0199] Embodiment 23: A method comprising receiving computationally derived data related to a monoclonal antibody (mAb), providing the computationally derived data to a predictive model, and determining a viscosity score related to the mAb based on the predictive model.

[0200] Embodiment 24: The embodiment according to Embodiment 23, further comprising adjusting an appropriate formulation composition or protein engineering strategy to mitigate specific challenges of a drug candidate under development, such as adjusting the amount of viscosity-reducing agent in a solution related to a query mAb, based on a viscosity score.

[0201] Embodiment 25: An embodiment according to any one of Embodiments 23 to 24, further comprising receiving sequence data related to an mAb and determining computationally derived data based on the sequence data.

[0202] Embodiment 26: An embodiment according to any one of Embodiments 23 to 25, wherein the calculated derived data includes viscosity data derived from calculations. Embodiment 27: An embodiment according to any one of Embodiments 23 to 26, wherein the calculated derived data includes viscosity data derived from calculations.

[0203] Embodiment 28: An embodiment according to any one of embodiments 23 to 27, further comprising receiving computationally derived data related to a query mAb, providing the computationally derived data to an optimal predictive model, and determining a viscosity score related to the query mAb based on the optimal predictive model.

[0204] Embodiment 29: A method comprising receiving computationally derived data related to a monoclonal antibody (mAb), providing the computationally derived data to a predictive model, and determining an agglutination score related to the mAb based on the predictive model.

[0205] Embodiment 30: An embodiment of Embodiment 29, further comprising adjusting an appropriate formulation composition or protein engineering strategy to mitigate specific challenges of a drug candidate under development, such as adjusting the amount of anti-aggregation agent in a solution associated with a query mAb, based on the aggregation score.

[0206] Embodiment 31: An embodiment according to any one of embodiments 29 to 30, further comprising receiving sequence data related to an mAb and determining computationally derived data based on the sequence data.

[0207] Embodiment 32: An embodiment according to any one of Embodiments 29 to 31, wherein the calculated derived data includes aggregated data derived from the calculation. Embodiment 33: An embodiment according to any one of claims 29 to 32, wherein the aggregation data derived from calculations includes high molecular weight (HMW) species formation data for mAbs.

[0208] Embodiment 34: An embodiment according to any one of Embodiments 29 to 33, wherein the calculated derived data includes charge data related to one or more regions related to the sequence of the mAb, modified charge data related to one or more regions based on the solvent exposure area of ​​the residues in the homology model of the mAb, a hydrophobicity index (HI), dipole moment, isoelectric point (pI), aggregation tendency (AP), or a descriptor of stereostructural stability.

[0209] Embodiment 35: An embodiment according to any one of embodiments 29 to 34, further comprising determining the optimal predictive model from a plurality of candidate predictive models related to the minimum error in predicting the aggregation score associated with mAbs.

[0210] Embodiment 36: An embodiment according to any one of embodiments 29 to 35, further comprising receiving computationally derived data related to a query mAb, providing the computationally derived data related to the query mAb to an optimal predictive model, and determining an agglomeration score related to the query mAb based on the optimal predictive model.

[0211] While methods and systems are described in relation to preferred embodiments and specific examples, their scope is not intended to be limited to the specific embodiments described herein. This is because the embodiments described herein are intended to be illustrative rather than restrictive in all respects.

[0212] Unless otherwise specified, no method described herein is intended to be construed as requiring its steps to be performed in a specific order. Therefore, if a claim of a method does not enumerate the order in which its steps should actually be followed, or unless otherwise specified in the claims or specification, no order should be presumed in any respect. This applies to all possible implicit basis for interpretation, including issues of logic regarding the arrangement of steps or the sequence of operations, the obvious meaning derived from grammatical organization or punctuation, and the number or type of embodiments described herein.

[0213] Those skilled in the art will be able to recognize or confirm numerous equivalents of the methods and specific embodiments of the compositions described herein by means of ordinary experiments. Such equivalents are intended to be covered by the following claims.

Claims

1. A computer implementation method for generating a predictive model for predicting the physical properties of one or more monoclonal antibodies (mAbs), comprising: To determine experimental agglutination data associated with one or more monoclonal antibodies (mAbs), Determining computationally derived data related to one or more mAbs, wherein the computationally derived data includes one or more computational parameters, and the one or more computational parameters include one or more charge values ​​related to one or more residues in one or more regions of the one or more mAbs, The weighting coefficient is calculated based on the solvent exposure area (SAS) of the one or more residues within the one or more regions of the one or more mAb, Based on the weighting coefficient, adjust the one or more charge values ​​associated with the one or more residues, Determining the charge value associated with each of the one or more regions based on the adjusted charge values ​​associated with the one or more residues, A method comprising generating a predictive model for predicting the physical properties of one or more mAb based on the experimental aggregation data and the computationally derived data, wherein the computationally derived data includes the charge values ​​related to each of the one or more regions.

2. The method according to claim 1, wherein the one or more mAbs comprises one or more IgG1 antibodies or IgG4 antibodies.

3. The method according to claim 1, wherein the one or more regions relate to the sequence of the one or more mAbs.

4. The method according to claim 1, wherein the experimental aggregation data includes high molecular weight (HMW) species formation data for each of the one or more mAbs.

5. The method according to claim 1, wherein determining the experimental agglutination data associated with one or more mAbs comprises measuring the amount of HMW species formation over time based on each of the one or more mAbs' solutions and size exclusion chromatography (SEC).

6. The method according to claim 1, wherein the calculated derived data includes charge data relating to one or more regions relating to the sequence of one or more mAbs, modified charge data relating to one or more regions based on the solvent exposure area of ​​residues in the homology model of one or more mAbs, a hydrophobicity index (HI), a dipole moment, an isoelectric point (pI), an aggregation tendency (AP), or a descriptor of stereostructural stability.

7. The method according to claim 6, wherein the descriptor for three-dimensional structural stability includes the mean squared deviation (RMSD) of the skeletal configuration of the structure relative to the initial structure after rigid alignment.

8. The method according to claim 1, wherein determining the computationally derived data relating to the one or more mAbs comprises one or more molecular dynamics (MD) simulations relating to the one or more mAbs.

9. The method according to claim 1, wherein determining the computationally derived data relating to the one or more mAbs includes complete antibody homology modeling of the sequences of the one or more mAbs, or Fab region modeling of the antigen-binding fragment (Fab) sequences of the one or more mAbs.

10. The weighting coefficient is calculated based on the solvent exposure area (SAS) of the one or more residues within the one or more regions of the one or more mAb. Based on the homology model of the one or more mAbs, determine the one or more charge values ​​associated with the one or more residues within the one or more regions of the one or more mAbs, The method according to claim 1, comprising determining the solvent exposure area (SAS) of one or more residues within one or more regions based on the homology model of one or more mAbs.

11. Based on the experimental aggregation data and the computationally derived data, the predictive model is generated. Identifying one or more experimental parameters of the aforementioned experimental aggregate data as the dependent variable, Identifying one or more computational parameters of the aforementioned derived data as independent variables, The method according to claim 1, comprising generating the predictive model based on a stepwise regression algorithm, based on the dependent variable and based on the independent variable.

12. The method according to claim 1, further comprising determining the Akaike Information Criterion (AIC) score for the prediction model.

13. Receiving computationally derived data related to query mAb, To provide the aforementioned predictive model with the aforementioned calculated derived data, The method according to claim 1, further comprising determining an agglutination score related to the query mAb based on the prediction model.

14. The method according to claim 13, further comprising adjusting the amount of flocculation-reducing agent in the solution related to the query mAb based on the flocculation score.

15. Based on the experimental aggregation data and the computationally derived data, the predictive model is generated. To generate a training dataset that includes at least a portion of the experimental agglutination data and at least a portion of the computationally derived data, wherein at least a portion of the experimental agglutination data includes direct measurement of agglutination of an antibody solution, and at least a portion of the computationally derived data includes one or more charge values ​​determined through computational modeling for residues on the antibody in the antibody solution. Based on the aforementioned training dataset, multiple parameters are extracted, The method according to claim 1, comprising training a machine learning-based classification model configured to predict antibody aggregation based on the training dataset and the plurality of parameters.

16. It is a device, One or more processors, A device comprising: a memory for storing processor-executable instructions that, when executed by one or more processors, cause the device to perform the method described in any one of claims 1 to 14.

17. One or more non-transient computer-readable media for storing processor-executable instructions that, when executed by the processor, cause the processor to perform the method according to any one of claims 1 to 14.

18. It is a system, A computing device configured to carry out the method described in any one of claims 1 to 14, A system comprising: a user device configured to display the output of the prediction model;

19. Receiving query calculation derivation data related to a query monoclonal antibody (mAb), wherein the calculation derivation data includes one or more calculation parameters, and the one or more calculation parameters include one or more charge values ​​related to one or more residues within one or more regions of the query mAb. A weighting coefficient is calculated based on the solvent exposure area (SAS) of one or more residues within one or more regions of the query mAb. Based on the weighting coefficient, adjust the one or more charge values ​​associated with the one or more residues, Determining the charge value associated with each of the one or more regions based on the adjusted charge values ​​associated with the one or more residues, The prediction model is provided with the computationally derived data, wherein the query computationally derived data includes the charge values ​​related to each of the one or more regions. A method comprising determining one or more aggregation scores or viscosity scores associated with the query mAb based on the prediction model.

20. Based on the aforementioned aggregation score, adjust the amount of aggregation-reducing agent in the solution related to the query mAb, or The method according to claim 19, further comprising one or more of the following: adjusting the amount of viscosity-reducing agent in the solution related to the query mAb based on the viscosity score.

21. Receiving sequence data related to the aforementioned query mAb, The method according to claim 19, further comprising determining the query calculation derived data based on the array data.

22. The method according to claim 19, wherein the query calculation derived data includes calculation derived aggregate data.

23. The method according to claim 22, wherein the calculated derived aggregation data includes high molecular weight (HMW) species formation data for the mAb.

24. The method according to claim 19, wherein the query calculation derived data includes calculated derived viscosity data.

25. The method according to claim 24, wherein the query calculation derived viscosity data includes one or more of dynamic viscosity values ​​or kinematic viscosity values.

26. The method according to claim 19, wherein the calculated derived data includes charge data relating to one or more regions relating to the sequence of the mAb, modified charge data relating to the one or more regions based on the solvent exposure area of ​​the residues in the homology model of the mAb, a descriptor of hydrophobicity index (HI), dipole moment, isoelectric point (pI), aggregation tendency (AP), or stereostructural stability.

27. The weighting coefficient is calculated based on the solvent exposure area (SAS) of the one or more residues within the one or more regions of the one or more mAb. Based on the homology model of the one or more mAbs, determine the one or more charge values ​​associated with the one or more residues within the one or more regions of the one or more mAbs, The method according to claim 19, comprising determining the solvent exposure area (SAS) of one or more residues within one or more regions based on the homology model of one or more mAbs.

28. The aforementioned prediction model is based on experimental aggregation data and computationally derived data. Identifying one or more experimental parameters of the aforementioned experimental aggregate data as the dependent variable, Identifying one or more computational parameters of the aforementioned derived data as independent variables, The predictive model is generated based on a stepwise regression algorithm, based on the dependent variable and based on the independent variable. The method according to claim 19, which is produced by a method including the following:

29. The aforementioned prediction model is based on experimental aggregation data and computationally derived data. To generate a training dataset that includes at least a portion of the experimental agglutination data and at least a portion of the computationally derived data, wherein at least a portion of the experimental agglutination data includes direct measurement of agglutination of an antibody solution, and at least a portion of the computationally derived data includes one or more charge values ​​determined through computational modeling for residues on the antibody in the antibody solution. Based on the aforementioned training dataset, multiple parameters are extracted, Training a machine learning-based classification model configured to predict antibody aggregation based on the training dataset and the multiple parameters, The method according to claim 19, which is produced by a method including the following:

30. The aforementioned prediction model is based on experimental viscosity data and computationally derived data. Identifying one or more experimental parameters of the aforementioned experimental viscosity data as the dependent variable, Identifying one or more computational parameters of the aforementioned derived data as independent variables, The predictive model is generated based on a stepwise regression algorithm, based on the dependent variable and based on the independent variable. The method according to claim 19, which is produced by a method including the following:

31. The aforementioned prediction model is based on experimental viscosity data and computationally derived data. To generate a training dataset that includes at least a portion of the experimental viscosity data and at least a portion of the computationally derived data, wherein at least a portion of the experimental viscosity data includes direct measurement of the viscosity of the antibody solution, and at least a portion of the computationally derived data includes one or more charge values ​​determined through computational modeling for residues on the antibody in the antibody solution. Based on the aforementioned training dataset, multiple parameters are extracted, Training a machine learning-based classification model configured to predict antibody viscosity based on the training dataset and the multiple parameters, The method according to claim 19, which is produced by a method including the following:

32. It is a device, One or more processors, A device comprising: a memory for storing processor-executable instructions that, when executed by one or more processors, cause the device to perform the method according to any one of claims 19 to 31.

33. One or more non-transient computer-readable media for storing processor-executable instructions that, when executed by a processor, cause the processor to perform the method according to any one of claims 19 to 31.

34. It is a system, A computing device configured to carry out the method described in any one of claims 19 to 31, A system comprising: a user device configured to display the output of the prediction model;

Images