System and method for predicting JAK mutations and their impact on drug efficacy

An Al-based framework using a hybrid transformer model with 3D convolutional neural networks addresses the challenge of predicting JAK mutations by capturing global features and long-range interactions, improving drug efficacy predictions and personalized medicine.

WO2026241282A1PCT designated stage Publication Date: 2026-11-19UNIV OF FLORIDA RESEARCH FOUNDATION INC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/US2026/027828
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-15
Filing Date
2026-05-12
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

Current methods struggle to accurately predict the impact of JAK mutations on drug efficacy due to limitations in capturing global features and long-range interactions, leading to insufficient accuracy in identifying stabilizing or destabilizing mutations, especially in complex proteins like JAK, which are crucial for understanding drug resistance in diseases.

Method used

An integrative framework combining Al techniques with structural biology and quantum chemistry to predict JAK mutations by using a hybrid transformer model with 3D convolutional neural networks, capturing global features and long-range interactions, and integrating multi-head attention mechanisms to enhance prediction accuracy.

Benefits of technology

The framework systematically characterizes JAK mutations, providing high-quality hypotheses for drug efficacy predictions and potential therapeutic insights, enhancing personalized medicine approaches.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 00000045_0000
    Figure 00000045_0000
  • Figure 00000046_0000
    Figure 00000046_0000
  • Figure 00000047_0000
    Figure 00000047_0000
Patent Text Reader

Abstract

Various embodiments of the present disclosure provide methods, apparatus, systems, computing devices, computing entities, and / or the like for receiving one or more protein sequences and structural inputs associated with the one or more protein sequences; iteratively masking residues to train a model to predict missing amino acids associated with the one or more protein sequences; capturing one or more global features and associated protein structure of the one or more protein sequences; integrating, using a hybrid transformer model architecture, the one or more global features and the associated protein structure; capturing, using a multi-head attention mechanism, at least one relationship between a set of parts of the associated protein structure; determining probability scores associated with the one or more protein sequences, the probability scores indicating a likelihood of mutation impacts on stability of the associated protein structure; and re-training the model based on specific subsets of data.
Need to check novelty before this filing date? Find Prior Art

Description

SYSTEM AND METHOD FOR PREDICTING JAK MUTATIONS AND THEIR IMPACT ON DRUG EFFICACYCROSS-REFERENCES TO RELATED APPLICATION(S)

[0001] The present application claims priority to U. S. Patent Application No. 63 / 806,088 (filed on May 15, 2025), which is incorporated by reference herein in its entirety.BACKGROUND

[0002] Various embodiments of the present disclosure address technical challenges related to predicting JAK mutations and their impact(s) on drug efficacy.

[0003] Protein misfolding and dysfunction are fundamental processes that underline numerous human diseases. Misfolded proteins can aggregate or exhibit abnormal activity, leading to cellular dysfunction and tissue damage in various disorders, including neurodegenerative diseases, cancer, metabolic, and inflammatory conditions. Abnormal protein function, such as constitutive activation of kinases, can lead to dysregulated signaling pathways as well, potentially leading to uncontrolled cell proliferation and survival. The Janus kinase (JAK 1-3 and TYK2) protein family is an example where single point mutations or multiple synergistic mutations can lead to dysfunction directly implicated in pathogenesis. JAK proteins are crucial components of the JAK-STAT signaling pathway, which regulates numerous biological processes including cell growth, differentiation, and immune responses. JAK2’s V617F mutation, for example, is an acquired, somatic mutation present in the majority of patients with myeloproliferative neoplasms (MPNs). The mutation disrupts normal JAK2 suppressive functions, causing its abnormal, constitutive activation.BRIEF SUMMARY

[0004] In general, various embodiments of the present disclosure provide methods, apparatus, systems, computing devices, computing entities, and / or the like for predicting JAK mutations and / or SOCS mutations and their impact(s) on drug efficacy.

[0005] In some embodiments, a computer-implemented method comprises: receiving, by one or more processors, one or more protein sequences and one or more structural inputs associated with the one or more protein sequences; iteratively masking, by the one or more processors,1LEGAL / 51273753vlresidues to train a model to predict missing amino acids associated with the one or more protein sequences; capturing, by the one or more processors, one or more global features and associated protein structure of the one or more protein sequences; integrating, by the one or more processors and using a hybrid transformer model architecture, the one or more global features and the associated protein structure; capturing, by the one or more processors and using a multi-head attention mechanism, at least one relationship between a set of parts of the associated protein structure; determining, by the one or more processors, one or more probability scores associated with the one or more protein sequences, the one or more probability scores indicating a likelihood of mutation impacts on stability of the associated protein structure; and re-training, by the one or more processors, the model based on one or more specific subsets of data.

[0006] In some embodiments, re-training the model based on one or more specific subsets of data includes fine-tuning the model using a first subset of data comprising one or more stable proteins.

[0007] In some embodiments, re-training the model based on one or more specific subsets of data includes fine-tuning the model using a second subset of data comprising one or more unstable proteins.

[0008] In some embodiments, the hybrid transformer model architecture comprises a 3-dimensional convolutional neural network (3DCNN).

[0009] In some embodiments, the hybrid transformer model architecture is configured to generate one or more global feature embeddings.

[0010] In some embodiments, the computer-implemented method further comprises conducting, by the one or more processors, energy minimization and molecular dynamics (MD) simulation on the one or more protein sequences.

[0011] In some embodiments, the computer-implemented method further comprises generating, by the one or more processors, a ranking of the one or more protein sequences based at least in part on the one or more probability scores.

[0012] In some embodiments, a computing apparatus comprises memory and one or more processors communicatively coupled to the memory, the one or more processors configured to perform operations comprising: receiving a multi-omics dataset corresponding to a set of compounds, identifying, using real-time parallel screening techniques, and based at least in part on a multi-omics dataset, a subset of target compounds from the set of compounds associated with2LEGAL / 51273753vlone or more biomarkers of a biological target, generating, using one or more machine learning models, at least one predicted compound efficacy associated with the subset of target compounds and the biological target, and validating, using a regulatory-ready data component, the at least one predicted compound efficacy associated with the subset of target compounds and the biological target.

[0013] In some embodiments, one or more non-transitory computer-readable storage media including instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising: receiving one or more protein sequences and one or more structural inputs associated with the one or more protein sequences; iteratively masking residues to train a model to predict missing amino acids associated with the one or more protein sequences; capturing one or more global features and associated protein structure of the one or more protein sequences; integrating, using a hybrid transformer model architecture, the one or more global features and the associated protein structure; capturing, using a multi-head attention mechanism, at least one relationship between a set of parts of the associated protein structure; determining one or more probability scores associated with the one or more protein sequences, the one or more probability scores indicating a likelihood of mutation impacts on stability of the associated protein structure; and re-training the model based on one or more specific subsets of data.BRIEF DESCRIPTION OF THE DRAWINGS

[0014] FIG. 1 provides an example overview of a framework in accordance with some embodiments of the present disclosure.

[0015] FIG. 2 provides an example computing entity in accordance with some embodiments of the present disclosure.

[0016] FIG. 3 provides an example user computing entity in accordance with some embodiments of the present disclosure.

[0017] FIG. 4 depicts an example cell mutation progression as described with respect to at least some embodiments of the present disclosure.

[0018] FIG. 5 depicts an example hybrid transformer model in accordance with at least some embodiments of the present disclosure.3LEGAL / 51273753vl

[0019] FIG. 6 depicts example JAK2-inhibitor MD results in accordance with at least some embodiments of the present disclosure.

[0020] FIG. 7 depicts an example table of in silico validation results in accordance with at least some example embodiments of the present disclosure.

[0021] FIG. 8A depicts an example Al-source code portion in accordance with at least some example embodiments of the present disclosure.

[0022] FIG. 8B depicts an example Al-source code portion in accordance with at least some example embodiments of the present disclosure.

[0023] FIG. 8C depicts an example Al-source code portion in accordance with at least some example embodiments of the present disclosure.

[0024] FIG. 8D depicts an example Al-source code portion in accordance with at least some example embodiments of the present disclosure.

[0025] FIG. 8E depicts an example Al-source code portion in accordance with at least some example embodiments of the present disclosure.

[0026] FIG. 8F depicts an example Al-source code portion in accordance with at least some example embodiments of the present disclosure.

[0027] FIG. 8G depicts an example Al-source code portion in accordance with at least some example embodiments of the present disclosure.

[0028] FIG. 8H depicts an example Al-source code portion in accordance with at least some example embodiments of the present disclosure.

[0029] FIG. 81 depicts an example Al-source code portion in accordance with at least some example embodiments of the present disclosure.

[0030] FIG. 8J depicts an example Al-source code portion in accordance with at least some example embodiments of the present disclosure.

[0031] FIG. 9 depicts an example process for systematically characterizing structural impact(s) known and new mutations in JAK2.

[0032] FIG. 10 depicts an example process for systematically characterizing structural impact(s) of known and new protein mutations in accordance with at least some embodiments of the present disclosure.4LEGAL / 51273753vlDETAILED DESCRIPTION

[0033] Various embodiments of the present disclosure are described more fully hereinafter with reference to the accompanying drawings, in which some, but not all embodiments of the present disclosure are shown. Indeed, the present disclosure may be embodied in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will satisfy applicable legal requirements. The term “or” is used herein in both the alternative and conjunctive sense, unless otherwise indicated. The terms “illustrative” and “example” are used to be examples with no indication of quality level. Terms such as “computing,” “determining,” “generating,” and / or similar words are used herein interchangeably to refer to the creation, modification, or identification of data. Further, “based on,” “based at least in part on,” “based at least on,” “based upon,” and / or similar words are used herein interchangeably in an open-ended manner such that they do not necessarily indicate being based only on or based solely on the referenced element or elements unless so indicated. Like numbers refer to like elements throughout.

[0034] Embodiments of the present disclosure include systems, methods, computer program products, and techniques for predicting JAK mutations and their impact(s) on drug efficacy.I. Overview, Technical Improvements, and Technical Advantages

[0035] Various embodiments of the present disclosure make important technical contributions to fields and / or environments for predicting JAK mutations and their impact(s) on drug efficacy.

[0036] Many relevant JAK mutations and their impacts on folding stability and activity remain unknown. The relationship between uncharacterized missense mutations and abnormal phosphorylation, as well as changes in JAK structural conformation, solubility, and activity, requires further exploration. While JAK inhibitors have shown promise in treating blood cancers, certain acquired mutations can confer resistance to these therapies. Secondary mutations in the JAK2 can emerge during malignant cell proliferation, altering its structure in a way that reduces inhibitor binding affinity, turning the cancer therapy ineffective.

[0037] With respect to existing solutions, it has not been feasible to experimentally probe each residue across all JAK proteins and their isoforms with each potential substitution to determine mutations that may impact the folding, stability, catalytic activity, and interactions with inhibitors. Protein engineering is a powerful approach with demonstrated impact in medicine, molecular5LEGAL / 51273753vlbiology, and translational sciences, but requires lengthy laboratory experimentation to discover critical mutation targets. In silico mutation screening methods are a faster alternative to assess the impact of mutations. However, most mutation screening tools are based on supervised learning models, which predict protein structural stability relying on small, labeled datasets (e.g., experimentally determined changes in free energy from single point mutations). Due to a major bias for knockout mutants in the labeled data available, most tools perform better at identifying destabilizing mutations (mutations that reduce the thermodynamic stability of a protein, making it less soluble, and more prone to misfolding or aggregation) and exhibit limited accuracy at stabilizing mutations. These tools would potentially miss pathogenic mutations that stabilize constitutively activated conformation of JAKs. Moreover, typical methods for predicting the impact of mutations are unable to consider the effects of interdomain interactions that are essential for JAK folding, stability, and function. They also utilize oversimplified energy functions that may fail to capture the complexity of protein folding stability and impact of long-range interactions within the structure.

[0038] Embodiments of the present disclosure describe Al-based methods that can directly propose candidate mutants from protein structure. Current solutions can predict mutations responsible for protein stability and function by learning structural-evolutionary patterns across all protein families from protein structure databases. These tools, however, are generalist models, which are trained on vast amounts of data from various protein families, thereby allowing them to have broad knowledge and capture general evolutionary patterns across proteins. Predicting the impact of mutations on a specific protein (like JAK, for example) requires more specific knowledge of its structure and function, including integrative approaches to assess the impact of mutations on posttranslational modifications (e.g., phosphorylation) and interactions with other molecules (e.g., inhibitors). Also, current Al-powered mutation screening models are based on 3D convolutional neural networks (3D CNN), which can only capture spatial patterns within a limited neighborhood around each residue in the protein, ignoring global features and long-range interactions, as well as how they are affected by the proposed mutations. Such a local view may be insufficient for predicting the effects of mutations in larger and more complex proteins like JAKs, which require multimodal sources of data, detection of complex long-range interactions and global conformational changes.6LEGAL / 51273753vl

[0039] Embodiments of the present disclosure provide an integrative framework for predicting JAK mutations and their impact(s) on drug efficacy and providing subsequent novel therapeutics that are mutation specific. Embodiments of the present disclosure integrate Al techniques with structural biology, biochemistry, Newtonian mechanics, and quantum chemistry principles to identify mutations in JAK2, as well as characterize their structural, functional, and clinical implications. Certain mutations in JAK2 may alter local protein structure, affecting folding intermediates and ultimately leading to misfolding or dysfunction. These structural changes may have far-reaching effects on JAK2-inhibitor interactions, potentially explaining the emergence of drug resistance, ultimately hindering the efficacy of JAK inhibitors in the treatment of JAK-associated diseases. Embodiments of the present disclosure will enhance predictions of JAK misfolding and dysfunction and will also provide valuable insights generating high quality hypotheses that can be further clinically tested in the future. By focusing on fundamental mechanisms underlying these processes, embodiments of the present disclosure provide a strong foundation for biomedical research and personalized medicine, with potential implications across a wide range of human diseases and biological phenomena.

[0040] Embodiments of the present disclosure systemically characterize JAK2 mutant proteins using Al, structural biochemistry, phylogenetics, and experimental validation. In at least some embodiments, a robust methodology is provided for characterizing the impact of mutations across the JAK family and other proteins contributing to personalized medicine solutions.II. Computer Program Products, Methods, and Computing Entities

[0041] Embodiments of the present disclosure may be implemented in various ways, including as computer program products that comprise articles of manufacture. Such computer program products may include one or more software components including, for example, software objects, methods, data structures, or the like. A software component may be coded in any of a variety of programming languages. An illustrative programming language may be a lower-level programming language such as an assembly language associated with a particular hardware architecture and / or operating system platform. A software component comprising assembly language instructions may require conversion into executable machine code by an assembler prior to execution by the hardware architecture and / or platform. Another example programming language may be a higher-level programming language that may be portable across multiple7LEGAL / 51273753vlarchitectures. A software component comprising higher-level programming language instructions may require conversion to an intermediate representation by an interpreter or a compiler prior to execution.

[0042] Other examples of programming languages include, but are not limited to, a macro language, a shell or command language, a job control language, a script language, a database query or search language, and / or a report writing language. In one or more example embodiments, a software component comprising instructions in one of the foregoing examples of programming languages may be executed directly by an operating system or other software component without having to be first transformed into another form. A software component may be stored as a file or other data storage construct. Software components of a similar type or functionally related may be stored together such as, for example, in a particular directory, folder, or library. Software components may be static (e.g., pre-established or fixed) or dynamic (e.g., created or modified at the time of execution).

[0043] A computer program product may include a non-transitory computer-readable storage medium storing applications, programs, program modules, scripts, source code, program code, object code, byte code, compiled code, interpreted code, machine code, executable instructions, and / or the like (also referred to herein as executable instructions, instructions for execution, computer program products, program code, and / or similar terms used herein interchangeably). Such non-transitory computer-readable storage media include all computer-readable media (including volatile and non-volatile media).

[0044] In some embodiments, a non-volatile computer-readable storage medium may include a floppy disk, flexible disk, hard disk, solid-state storage (SSS) (e.g., a solid state drive (SSD), solid state card (SSC), solid state module (SSM), enterprise flash drive, magnetic tape, or any other non-transitory magnetic medium, and / or the like. A non-volatile computer-readable storage medium may also include a punch card, paper tape, optical mark sheet (or any other physical medium with patterns of holes or other optically recognizable indicia), compact disc read only memory (CD-ROM), compact disc-rewritable (CD-RW), digital versatile disc (DVD), Blu-ray disc (BD), any other non-transitory optical medium, and / or the like. Such a non-volatile computer-readable storage medium may also include read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory (e.g., Serial, NAND, NOR, and / or8LEGAL / 51273753vlthe like), multimedia memory cards (MMC), secure digital (SD) memory cards, SmartMedia cards, CompactFlash (CF) cards, Memory Sticks, and / or the like. Further, a non-volatile computer-readable storage medium may also include conductive-bridging random access memory (CBRAM), phase-change random access memory (PRAM), ferroelectric random-access memory (FeRAM), non-volatile random-access memory (NVRAM), magnetoresistive random-access memory (MRAM), resistive random-access memory (RRAM), Silicon-Oxide-Nitride-Oxide-Silicon memory (SONOS), floating junction gate random access memory (FJGRAM), Millipede memory, racetrack memory, and / or the like.

[0045] In some embodiments, a volatile computer-readable storage medium may include random access memory (RAM), dynamic random access memory (DRAM), static random access memory (SRAM), fast page mode dynamic random access memory (FPM DRAM), extended data-out dynamic random access memory (EDO DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), double data rate type two synchronous dynamic random access memory (DDR2 SDRAM), double data rate type three synchronous dynamic random access memory (DDR3 SDRAM), Rambus dynamic random access memory (RDRAM), Twin Transistor RAM (TTRAM), Thyristor RAM (T-RAM), Zero-capacitor (Z-RAM), Rambus in-line memory module (RIMM), dual in-line memory module (DIMM), single in-line memory module (SIMM), video random access memory (VRAM), cache memory (including various levels), flash memory, register memory, and / or the like. It will be appreciated that where embodiments are described to use a computer-readable storage medium, other types of computer-readable storage media may be substituted for or used in addition to the computer-readable storage media described above.

[0046] As should be appreciated, various embodiments of the present disclosure may also be implemented as methods, apparatus, systems, computing devices, computing entities, and / or the like. As such, embodiments of the present disclosure may take the form of an apparatus, system, computing device, computing entity, and / or the like executing instructions stored on a computer-readable storage medium to perform certain steps or operations. Thus, embodiments of the present disclosure may also take the form of an entirely hardware embodiment, an entirely computer program product embodiment, and / or an embodiment that comprises combination of computer program products and hardware performing certain steps or operations.9LEGAL / 51273753vl

[0047] Embodiments of the present disclosure are described below with reference to block diagrams and flowchart illustrations. Thus, it should be understood that each block of the block diagrams and flowchart illustrations may be implemented in the form of a computer program product, an entirely hardware embodiment, a combination of hardware and computer program products, and / or apparatus, systems, computing devices, computing entities, and / or the like carrying out instructions, operations, steps, and similar words used interchangeably (e.g., the executable instructions, instructions for execution, program code, and / or the like) on a computer-readable storage medium for execution. For example, retrieval, loading, and execution of code may be performed sequentially such that one instruction is retrieved, loaded, and executed at a time. In some example embodiments, retrieval, loading, and / or execution may be performed in parallel such that multiple instructions are retrieved, loaded, and / or executed together. Thus, such embodiments can produce specifically configured machines performing the steps or operations specified in the block diagrams and flowchart illustrations. Accordingly, the block diagrams and flowchart illustrations support various combinations of embodiments for performing the specified instructions, operations, or steps.III. Example Framework

[0048] FIG. 1 provides an example overview of a framework 100 in accordance with some embodiments of the present disclosure. The framework 100 may correspond to a system for predicting JAK mutations and their impact(s) on drug efficacy. As depicted, framework 100 includes data layer 110, pre-processing layer 120, transformer layer 130, multi-head attention mechanism 140, output layer 150, and task layer 160. Framework 100 may provide a hybrid transformer model that combines feature embeddings with a 3D convolutional neural network to predict stabilizing and destabilizing mutations in proteins in a self-supervised learning framework. Framework 100 may be configured to mask wild type residue iteratively at each given position and train the subject model(s) to predict the masked residue. In doing so, the subject model(s) can learn representations that capture the general evolutionary patterns of protein stability, even in the absence of direct labels for all possible mutations.

[0049] As depicted, framework 100 includes a data layer 110. Data layer 110 may include a general dataset comprising a plurality of protein structures and sequences. In at least some embodiments, data layer 110 may include a database or other datastore configured to provide data10LEGAL / 51273753vlcorresponding to protein structures and sequences. Data layer 110 may be configured to provide labeled data or unlabeled data. In general, data layer 110 includes any application, component, apparatus, device, and the like configured to store and / or provide data corresponding to subject protein structures and sequences.

[0050] Data layer 110 may be configured to store and provide one or more protein sequences and one or more structural inputs associated with the one or more protein sequences. In at least some embodiments, data layer 110 may include a database or other datastore configured to provide data corresponding to protein structures and sequences. Data layer 110 may be configured to provide labeled data or unlabeled data. In at least some embodiments, data layer 110 may provide unlabeled Protein Data Bank (PDB) structures for training in a self-supervised framework. In at least some embodiments, data layer 110 may provide one or more specific subsets of data for retraining the model, including subsets comprising one or more stable proteins and subsets comprising one or more unstable proteins. In general, data layer 110 includes any application, component, apparatus, device, and the like configured to store and / or provide data corresponding to subject protein structures and sequences.

[0051] As depicted, framework 100 includes pre-processing layer 120. Pre-processing layer 120 may include one or more mechanisms configured to extract relevant features from the protein sequences and structures as provided via the data layer 110. Pre-processing layer 120 may be configured to identify and extract characteristics or features from raw data corresponding to the protein sequences and structure data to create a more informative dataset for subsequent tasks. Preprocessing layer 120 may be configured to transform raw data into a format more readily usable and understood by algorithms as leveraged by machine learning models. In at least some embodiments, pre-processing layer 120 is configured to encode global feature vectors, such as 1D / 2D features (such as ASA / HB) into embeddings and capturing 3D structural information.

[0052] Pre-processing layer 120 may be configured to identify and extract characteristics or features from raw data corresponding to the protein sequences and structure data to create a more informative dataset for subsequent tasks. Pre-processing layer 120 may be configured to transform raw data into a format more readily usable and understood by algorithms as leveraged by machine learning models. In at least some embodiments, pre-processing layer 120 is configured to encode global feature vectors, such as 1D / 2D features (such as ASA / HB) into embeddings and capturing 3D structural information. This may include encoding global feature vectors into embeddings and11LEGAL / 51273753vlcapturing structural information using 3DCNN. Tn at least some embodiments, pre-processing layer 120 may leverage a residue-aligned voxelization framework that replaces traditional global Euclidean grids with local spherical voxel maps centered on each amino acid, creating a rotationally consistent coordinate system. In at least some embodiments, pre-processing layer 120 may extract detailed atomic-level interaction features from 3D protein-ligand complexes, capturing hydrogen bonds, hydrophobic contacts, and van der Waals forces, along with intramolecular stability -related features. In at least some embodiments, pre-processing layer 120 may be configured to capture one or more global features and associated protein structure of the one or more protein sequences. In general, pre-processing layer 120 includes any application, component, apparatus, device, and the like configured to extract and transform features from the one or more protein sequences and one or more structural inputs associated with the one or more protein sequences.

[0053] As depicted, framework 100 includes transformer layer 130. Transformer layer 130 may include a combination of feature embeddings and a 3D convolutional neural network. Transformer layer 130 may be configured to capture long-range dependencies and contextual information within sequences. Transformer layer 130 may be configured, via the 3D convolutional neural network (3DCNN), to detect spatial and hierarchical patterns from 3D data such as protein structures. In general, transformer layer 130 may be configured to integrate extracted multimodal features.

[0054] Transformer layer 130 may include a hybrid transformer model architecture comprising a combination of feature embeddings and a 3-dimensional convolutional neural network (3DCNN). Transformer layer 130 may be configured to capture long-range dependencies and contextual information within sequences. Transformer layer 130 may be configured, via the 3DCNN, to detect spatial and hierarchical patterns from 3D data such as protein structures. In at least some embodiments, transformer layer 130 may be configured to integrate extracted multimodal features into an early-fusion attention head mechanism. In at least some embodiments, transformer layer 130 may be configured to generate one or more global feature embeddings. In at least some embodiments, transformer layer 130 may capture evolutionary, structural, and physicochemical patterns from the one or more protein sequences and one or more structural inputs. In general, transformer layer 130 may be configured to integrate the one or more global12LEGAL / 51273753vlfeatures and the associated protein structure to enable prediction of stabilizing and destabilizing mutations in proteins.

[0055] As depicted, framework 100 includes multi-head attention mechanism 140. Multi-head attention mechanism 140 may be configured to enhance contextual understanding with respect to the protein structures and sequences. In at least some embodiments, multi-head attention mechanism 140 may be configured to be applied multiple times in parallel, each time with a different set of learnable parameters, to capture different types of relationships between the protein structures and sequences. By doing so, multi-head attention mechanism 140 allows the subject model(s) to attend to various aspects of the protein structures and sequences simultaneously, leading to richer representations and better overall performance. In at least some embodiments, multi-head attention mechanism determines how distant residues in a subject protein interact, while also boosting the model’s capacity to identify dependencies between interactions.

[0056] Multi-head attention mechanism 140 may be configured to capture at least one relationship between a set of parts of the associated protein structure. Multi-head attention mechanism 140 may be configured to enhance contextual understanding with respect to the protein structures and sequences. In at least some embodiments, multi-head attention mechanism 140 may be configured to be applied multiple times in parallel, each time with a different set of learnable parameters, to capture different types of relationships between the protein structures and sequences. By doing so, multi-head attention mechanism 140 allows the subject model(s) to attend to various aspects of the protein structures and sequences simultaneously, leading to richer representations and better overall performance.

[0057] In at least some embodiments, multi-head attention mechanism 140 may be configured to capture complex relationships between different parts of the protein structure. In at least some embodiments, multi-head attention mechanism 140 determines how distant residues in a subject protein interact, while also boosting the model’s capacity to identify dependencies between interactions. Multi-head attention mechanism 140 may be configured to capture long-range residue interactions associated with the one or more protein sequences. In at least some embodiments, extracted multimodal features may be integrated into an early-fusion attention head mechanism. In general, multi-head attention mechanism 140 may be configured to detect complex long-range interactions and global conformational changes, which may be particularly relevant for predicting the effects of mutations in larger and more complex proteins.13LEGAL / 51273753vl

[0058] As depicted, framework 100 includes output layer 150. Output layer 150 is configured to determine probability scores for each one of the 20 possible amino acids (AA) for each residue in a subject protein sequence. As depicted, framework 100 includes task layer 160. Task layer 160 may be configured to fine-tune the subject model for a plurality of specific tasks. For example, task layer 160 may be configured to predict which mutations enhance the stability and activity of a protein, and which mutations will destabilize the protein structure and function. In at least some embodiments, task layer 160 is configured to train the model with specific subsets of data for stable and unstable proteins, which allows the model to specialize in predicting mutants with these different outcomes. The final output comprises of a list of proposed mutations and probabilities indicating whether these mutations will stabilize or destabilize the protein. The list can then be further refined via molecular dynamics (MD) simulations and phylogenetic analysis before experimental validation.

[0059] Output layer 150 may be configured to determine one or more probability scores associated with the one or more protein sequences, the one or more probability scores indicating a likelihood of mutation impacts on stability of the associated protein structure. In at least some embodiments, output layer 150 may be configured to determine probability scores for each one of the 20 possible amino acids (AA) for each residue in a subject protein sequence. Output layer 150 may be configured to generate a list of proposed mutations and probabilities indicating whether these mutations will stabilize or destabilize the protein. In at least some embodiments, output layer 150 is configured to rank predicted mutants according to a confidence score, which may be represented by the model’s predicted probability for a given mutation. In at least some embodiments, output layer 150 may be configured to generate a ranking of the one or more protein sequences based at least in part on the one or more probability scores. In general, output layer 150 may be configured to provide probability scores that can be further refined via molecular dynamics (MD) simulations and phylogenetic analysis before experimental validation.

[0060] It should be appreciated that, though FIG. 1 depicts each of the various components collocated, many embodiments exist in which the various components are distributed amongst a plurality of physically distinct systems, which may be connected via a network connection, for example.Example User Device14LEGAL / 51273753vl

[0061] FIG. 2 provides an example computing entity 200 in accordance with some embodiments of the present disclosure. The computing entity 200 may be an example of the framework 100 of FIG. 1, for example. In general, the terms computing entity, computer, entity, device, system, and / or similar words used herein interchangeably may refer to, for example, one or more computers, computing entities, desktops, mobile phones, tablets, phablets, notebooks, laptops, distributed systems, kiosks, input terminals, servers or server networks, blades, gateways, switches, processing devices, processing entities, set-top boxes, relays, routers, network access points, base stations, the like, and / or any combination of devices or entities adapted to perform the functions, operations, and / or processes described herein.

[0062] As shown in FIG. 2, in some embodiments, the computing entity 200 may include, or be in communication with, one or more processing elements 205 (also referred to as processors, processing circuitry, and / or similar terms used herein interchangeably) that communicate with other elements within the computing entity 200 via a bus, for example. As will be understood, the processing element 205 may be embodied in a number of different ways.

[0063] For example, the processing element 205 may be embodied as one or more complex programmable logic devices (CPLDs), microprocessors, multi-core processors, coprocessing entities, application-specific instruction-set processors (ASIPs), microcontrollers, and / or controllers. Further, the processing element 205 may be embodied as one or more other processing devices or circuitry. The term circuitry may refer to an entirely hardware embodiment or a combination of hardware and computer program products. Thus, the processing element 205 may be embodied as integrated circuits, application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), programmable logic arrays (PLAs), hardware accelerators, other circuitry, and / or the like.

[0064] As will therefore be understood, the processing element 205 may be configured for a particular use or configured to execute instructions stored in volatile or non-volatile media or otherwise accessible to the processing element 205. As such, whether configured by hardware or computer program products, or by a combination thereof, the processing element 205 may be capable of performing steps or operations according to embodiments of the present disclosure when configured accordingly.

[0065] In some embodiments, the computing entity 200 may further include, or be in communication with, non-volatile media (also referred to as non-volatile storage, memory,15LEGAL / 51273753vlmemory storage, memory circuitry, and / or similar terms used herein interchangeably). In some embodiments, the non-volatile media may include one or more non-volatile memory 210, including, but not limited to, hard disks, ROM, PROM, EPROM, EEPROM, flash memory, MMCs, SD memory cards, Memory Sticks, CBRAM, PRAM, FeRAM, NVRAM, MRAM, RRAM, SONOS, FJG RAM, Millipede memory, racetrack memory, and / or the like.

[0066] As will be recognized, the non-volatile media may store databases, database instances, database management systems, data, applications, programs, program modules, scripts, code (e.g., source code, object code, byte code, compiled code, interpreted code, machine code, etc.) that embodies one or more computer functions described herein, executable instructions, and / or the like. The term database, database instance, database management system, and / or similar terms used herein interchangeably, may refer to a collection of records or data that is stored in a computer-readable storage medium using one or more database models; such as a hierarchical database model, network model, relational model, entity-relationship model, object model, document model, semantic model, graph model, and / or the like.

[0067] In some embodiments, the computing entity 200 may further include, or be in communication with, volatile media (also referred to as volatile storage, memory, memory storage, memory circuitry, and / or similar terms used herein interchangeably). In some embodiments, the volatile media may also include one or more volatile memory 215, including, but not limited to, RAM, DRAM, SRAM, FPM DRAM, EDO DRAM, SDRAM, DDR SDRAM, DDR2 SDRAM, DDR3 SDRAM, RDRAM, TTRAM, T-RAM, Z-RAM, RIMM, DIMM, SIMM, VRAM, cache memory, register memory, and / or the like.

[0068] As will be recognized, the volatile storage or memory media may be used to store at least portions of the databases, database instances, database management systems, data, applications, programs, program modules, code (source code, object code, byte code, compiled code, interpreted code, machine code) that embodies one or more other computer functions described herein, executable instructions, and / or the like being executed by, for example, the processing element 205. Thus, the databases, database instances, database management systems, data, applications, programs, program modules, code (source code, object code, byte code, compiled code, interpreted code, machine code) that embodies one or more computer functions described herein, executable instructions, and / or the like may be used to control certain aspects of16LEGAL / 51273753vlthe operation of the computing entity 200 with the assistance of the processing element 205 and operating system.

[0069] As indicated, in some embodiments, the computing entity 200 may also include one or more network interfaces 220 for communicating between various computing entities such as by communicating data, code, content, information, and / or similar terms used herein interchangeably that may be transmitted, received, operated on, processed, displayed, stored, and / or the like. Such communication may be executed using a wired data transmission protocol, such as fiber distributed data interface (FDDI), digital subscriber line (DSL), Ethernet, asynchronous transfer mode (ATM), frame relay, data over cable service interface specification (DOCSIS), or any other wired transmission protocol. In some embodiments, the computing entity 200 communicates with another computing entity for uploading or downloading data or code. Similarly, the computing entity 200 may be configured to communicate via wireless external communication networks using any of a variety of protocols, such as general packet radio service (GPRS), Universal Mobile Telecommunications System (UMTS), Code Division Multiple Access 2000 (CDMA2000), CDMA2000 IX (IxRTT), Wideband Code Division Multiple Access (WCDMA), Global System for Mobile Communications (GSM), Enhanced Data rates for GSM Evolution (EDGE), Time Division-Synchronous Code Division Multiple Access (TD-SCDMA), Long Term Evolution (LTE), Evolved Universal Terrestrial Radio Access Network (E-UTRAN), Evolution-Data Optimized (EVDO), High Speed Packet Access (HSPA), High-Speed Downlink Packet Access (HSDPA), IEEE 802.11 (Wi-Fi), Wi-Fi Direct, 802.16 (WiMAX), ultra-wideband (UWB), infrared (IR) protocols, near field communication (NFC) protocols, Wibree, Bluetooth protocols, wireless universal serial bus (USB) protocols, and / or any other wireless protocol.

[0070] Although not shown, the computing entity 200 may include, or be in communication with, one or more input elements, such as a keyboard input, a mouse input, a touch screen / display input, motion input, movement input, audio input, pointing device input, joystick input, keypad input, and / or the like. The computing entity 200 may also include, or be in communication with, one or more output elements (not shown), such as audio output, video output, screen / display output, motion output, movement output, and / or the like.Example External Computing Entity17LEGAL / 51273753vl

[0071] FIG. 3 provides an example computing entity 300 in accordance with some embodiments of the present disclosure. In general, FIG. 3 may be representative of components of any of the modules / systems described herein generally enabling processing within each component and communication between the components. In general, the terms device, system, computing entity, entity, and / or similar words used herein interchangeably may refer to, for example, one or more computers, computing entities, desktops, mobile phones, tablets, phablets, notebooks, laptops, distributed systems, kiosks, input terminals, servers or server networks, blades, gateways, switches, processing devices, processing entities, set-top boxes, relays, routers, network access points, base stations, the like, and / or any combination of devices or entities adapted to perform the functions, operations, and / or processes described herein. As shown in FIG. 3, the computing entity 300 may include an antenna 312, a transmitter 304 (e.g., radio), a receiver 306 (e.g., radio), and a processing element 308 (e.g., CPLDs, microprocessors, multi-core processors, coprocessing entities, ASIPs, microcontrollers, and / or controllers) that provides signals to and receives signals from the transmitter 304 and receiver 306, correspondingly. Generally, computing entity 300 may be an example of the framework 100 of FIG. 1.

[0072] The signals provided to and received from the transmitter 304 and the receiver 306, correspondingly, may include signaling information / data in accordance with air interface standards of applicable wireless systems. In this regard, the computing entity 300 may be capable of operating with one or more air interface standards, communication protocols, modulation types, and access types. More particularly, the computing entity 300 may operate in accordance with any of a number of wireless communication standards and protocols, such as those described above with regard to the computing entity 200. In some embodiments, the computing entity 300 may operate in accordance with multiple wireless communication standards and protocols, such as UMTS, CDMA2000, IxRTT, WCDMA, GSM, EDGE, TD-SCDMA, LTE, E-UTRAN, EVDO, HSPA, HSDPA, Wi-Fi, Wi-Fi Direct, WiMAX, UWB, IR, NFC, Bluetooth, USB, and / or the like. Similarly, the computing entity 300 may operate in accordance with multiple wired communication standards and protocols, such as those described above with regard to the computing entity 200 via a network interface 320.

[0073] Via these communication standards and protocols, the computing entity 300 may communicate with various other entities using mechanisms such as Unstructured Supplementary Service Data (USSD), Short Message Service (SMS), Multimedia Messaging Service (MMS),18LEGAL / 51273753vlDual-Tone Multi -Frequency Signaling (DTMF), and / or Subscriber Identity Module Dialer (SIM dialer). The computing entity 300 may also download code, changes, add-ons, and updates, for instance, to its firmware, software (e.g., including executable instructions, applications, program modules), and operating system.

[0074] The computing entity 300 may also comprise a user interface (that may include an output device 316 (e.g., display, speaker, tactile instrument, etc.) coupled to a processing element 308) and / or a user input interface (coupled to a processing element 308). For example, the user interface may be a user application, browser, user interface, and / or similar words used herein interchangeably executing on and / or accessible via the computing entity 300 to interact with and / or cause display of information / data from the computing entity 200, as described herein. The user input interface may comprise any of a plurality of input devices 318 (or interfaces) allowing the computing entity 300 to receive code and / or data, such as a keypad (hard or soft), a touch display, voice / speech or motion interfaces, or other input device. In some embodiments including a keypad, the keypad may include (or cause display of) the conventional numeric (0-9) and related keys (#, *), and other keys used for operating the computing entity 300 and may include a full set of alphabetic keys or set of keys that may be activated to provide a full set of alphanumeric keys. In addition to providing input, the user input interface may be used, for example, to activate or deactivate certain functions, such as screen savers and / or sleep modes.

[0075] The computing entity 300 may also include volatile memory 322 and / or non-volatile memory 324, which may be embedded and / or may be removable. For example, the non-volatile memory 324 may be ROM, PROM, EPROM, EEPROM, flash memory, MMCs, SD memory cards, Memory Sticks, CBRAM, PRAM, FeRAM, NVRAM, MRAM, RRAM, SONOS, FJG RAM, Millipede memory, racetrack memory, and / or the like. The volatile memory 322 may be RAM, DRAM, SRAM, FPM DRAM, EDO DRAM, SDRAM, DDR SDRAM, DDR2 SDRAM, DDR3 SDRAM, RDRAM, TTRAM, T-RAM, Z-RAM, RIMM, DIMM, SIMM, VRAM, cache memory, register memory, and / or the like. The volatile and non-volatile memory may store databases, database instances, database management systems, data, applications, programs, program modules, scripts, code (source code, object code, byte code, compiled code, interpreted code, machine code, etc.) that embodies one or more computer functions described herein, executable instructions, and / or the like to implement the functions of the computing entity 300. As indicated, this may include a user application that is resident on the computing entity 300 or19LEGAL / 51273753vlaccessible through a browser or other user interface for communicating with the computing entity 200 and / or various other computing entities.

[0076] In another embodiment, the computing entity 300 may include one or more components or functionalities that are the same or similar to those of the computing entity 200, as described in greater detail above. As will be recognized, these architectures and descriptions are provided for example purposes only and are not limited to the various embodiments.

[0077] In various embodiments, the computing entity 300 may be embodied as an artificial intelligence (Al) computing entity. Accordingly, the computing entity 300 may be configured to provide and / or receive information / data from a user via an input / output mechanism, such as a display, a camera, a speaker, a voice-activated input, and / or the like. In certain embodiments, an Al computing entity may comprise one or more predefined and executable program algorithms stored within an onboard memory storage module, and / or accessible over a network. In various embodiments, the Al computing entity may be configured to retrieve and / or execute one or more of the predefined program algorithms upon the occurrence of a predefined trigger event.Example Embodiments

[0078] FIG. 4 depicts an example cell mutation progression 400 as described with respect to at least some embodiments of the present disclosure. As depicted, cell mutation progression 400 includes the progression from a normal cell 410 (e.g., a cell exhibiting typical JAK-STAT signaling pathway function, with properly regulated cell growth, differentiation, and immune responses) to a malignant cell 420 (e.g., a cell that has undergone transformation due to the primary mutation, exhibiting dysregulated signaling pathways that may lead to uncontrolled cell proliferation and survival) via a primary mutation 414. The primary mutation 414 may comprise an initial somatic mutation, such as JAK2’s V617F, that disrupts normal JAK suppressive functions and causes abnormal, constitutive activation of kinase.

[0079] As depicted, the malignant cell 420 includes a JAK-inhibitor sensitive mutant 422. The JAK-inhibitor sensitive mutant 422 may comprise a mutant protein within the malignant cell that remains responsive to JAK inhibitor therapies, maintaining sufficient inhibitor binding affinity. The malignant cell 420 proceeds to undergo proliferation 424, thereby altering the cell’s structure in a manner which reduces inhibitor binding affinity and leads to the generation of a JAK inhibitorresistant mutant 432. Proliferation 424 may comprise any process during which malignant cells20LEGAL / 51273753vlmultiply, during which secondary mutations can emerge that alter JAK2 structure. Subsequent to a dose escalation 434 (e.g., an increase in therapeutic dosage administered), the cell mutation progression 400 as depicted ultimately leads to an acquired resistance cell cluster 440, which ultimately renders the previously administered therapy (or therapies) ineffective. The acquired resistance cell cluster 440 may comprise a population of cells harboring JAK inhibitor-resistant mutations that render previously administered JAK inhibitor therapies ineffective.

[0080] FIG. 5 depicts an example hybrid transformer model 500 in accordance with at least some embodiments of the present disclosure. As depicted, the transformer model includes a data layer 510, pre-processing layer 520, transformer layer 530, multi-head attention mechanism 540, output layer 550, and task layer 560.

[0081] Data layer 510 may include a general dataset comprising a plurality of protein structures and sequences. In at least some embodiments, data layer 510 may include a database or other datastore configured to provide data corresponding to protein structures and sequences. Data layer 510 may be configured to provide labeled data or unlabeled data. In general, data layer 510 includes any application, component, apparatus, device, and the like configured to store and / or provide data corresponding to subject protein structures and sequences.

[0082] Pre-processing layer 520 may include one or more mechanisms configured to extract relevant features from the protein sequences and structures as provided via the data layer 510. Preprocessing layer 520 may be configured to identify and extract characteristics or features from raw data corresponding to the protein sequences and structure data to create a more informative dataset for subsequent tasks. Pre-processing layer 520 may be configured to transform raw data into a format more readily usable and understood by algorithms as leveraged by machine learning models. In at least some embodiments, pre-processing layer 520 is configured to encode global feature vectors, such as 1D / 2D features (such as ASA / HB) into embeddings and capturing 3D structural information.

[0083] Transformer layer 530 may include a combination of feature embeddings and a 3D convolutional neural network. Transformer layer 530 may be configured to capture long-range dependencies and contextual information within sequences. Transformer layer 530 may be configured, via the 3D convolutional neural network (3DCNN), to detect spatial and hierarchical patterns from 3D data such as protein structures. In general, transformer layer 530 may be configured to integrate extracted multimodal features.21LEGAL / 51273753vl

[0084] Multi-head attention mechanism 540 may be configured to enhance contextual understanding with respect to the protein structures and sequences. In at least some embodiments, multi-head attention mechanism 540 may be configured to be applied multiple times in parallel, each time with a different set of learnable parameters, to capture different types of relationships between the protein structures and sequences. By doing so, multi-head attention mechanism 540 allows the subject model(s) to attend to various aspects of the protein structures and sequences simultaneously, leading to richer representations and better overall performance. In at least some embodiments, multi-head attention mechanism determines how distant residues in a subject protein interact, while also boosting the model’s capacity to identify dependencies between interactions.

[0085] Output layer 550 is configured to determine probability scores for each one of the 20 possible amino acids (AA) for each residue in a subject protein sequence. Task layer 560 may be configured to fine-tune the subject model for a plurality of specific tasks. For example, task layer 560 may be configured to predict which mutations enhance the stability of a protein, and which mutations will destabilize the protein structure. In at least some embodiments, task layer 560 is configured to train the model with specific subsets of data for stable and unstable proteins, which allows the model to specialize in predicting mutants with these different outcomes. The final output comprises of a list of proposed mutations and probabilities indicating whether these mutations will stabilize or destabilize the protein. The list can then be further refined via molecular dynamics (MD) simulations and phylogenetic analysis before experimental validation.

[0086] Hybrid transformer model 500 contrasts with conventional supervised models, which require a vast number of experimentally validated labels for training. In practice, such labeled datasets are limited due to the cost and effort required to experimentally determine protein stability changes for every mutation and combinations of mutations. Hybrid transformer model 500 instead circumvents this limitation by training on a large unlabeled dataset (137,399 PDB structures stratified into 90-10% train-test split, for example), rather than relying exclusively on experimentally determined labels. After fine tuning the model by training and validating it in target subsets, the model can then predict new stabilizing and destabilizing mutations in a much more scalable and specialized approach. Hybrid transformer model 500 is able to predict combinations of mutations, which is particularly challenging for supervised models. In addition, both models can be applied to assess the probability of structural impact for known missense variations found22LEGAL / 51273753vlacross human populations. These variations, although widely documented, may not have been tested in terms of their effects on protein stability. Further, the results provided by hybrid transformer model 500 comprise of a fine-tuned set of candidate mutations that can be further validated in silico and in vitro, ultimately helping the fundamental understanding of key mutations that affect protein structural dynamics.

[0087] FIG. 6 depicts example JAK2-inhibitor MD results in accordance with at least some embodiments of the present disclosure. As depicted, the example JAK2-inhibitor MD results are depicted in four panels. Panel 610 depicts a root mean square deviation (RMSD) plot calculated for mutants showing significant changes. Panel 620 depicts that the wildtype structure was aligned with the M865K structure. Panel 630 depicts the binding site of the wildtype protein, whereas panel 640 depicts the binding site of M865K, both with the ligand and K865.

[0088] FIG. 7 depicts an example table 700 of in silico validation results in accordance with at least some example embodiments of the present disclosure. As depicted, table 1 ranks the predicted mutants according to their confidence score, which is represented by the model’s predicted probability for a given mutation. The mutations were further evaluated in silico by building the mutant structures using a protein modeling software. The mutant 3D models were utilized for calculating changes in Gibbs free energy relative to the wildtype protein, which quantifies the impact of the mutation on the protein’s structural stability. 5 of the 12 candidate mutants possess a strong impact on structural stability, whereas 4 were destabilizing. The top 12 hits also included mutations at residues V617, M865, and Q885. While Gibbs free energy calculations provide an initial assessment of protein stability changes, they fundamentally rely on machine learning models. Molecular dynamics (MD) simulations offer a physics-based approach grounded in Newtonian mechanics and quantum chemistry principles providing a refinement method for filtering potentially biased predictions. MD simulations capture subtle conformational changes, solvent effects, and time-dependent phenomena that are often overlooked in simplified predictive models, yielding a more mechanistically insightful result for evaluating each candidate mutant. The resultant table 700 identifies M865 as a potential target for developing selective inhibitors.

[0089] FIG. 8A depicts an example Al-source code portion in accordance with at least some example embodiments of the present disclosure.

[0090] FIG. 8B depicts an example Al-source code portion in accordance with at least some example embodiments of the present disclosure.23LEGAL / 51273753vl

[0091] FIG. 8C depicts an example Al-source code portion in accordance with at least some example embodiments of the present disclosure.

[0092] FIG. 8D depicts an example Al-source code portion in accordance with at least some example embodiments of the present disclosure.

[0093] FIG. 8E depicts an example Al-source code portion in accordance with at least some example embodiments of the present disclosure.

[0094] FIG. 8F depicts an example Al-source code portion in accordance with at least some example embodiments of the present disclosure.

[0095] FIG. 8G depicts an example Al-source code portion in accordance with at least some example embodiments of the present disclosure.

[0096] FIG. 8H depicts an example Al-source code portion in accordance with at least some example embodiments of the present disclosure.

[0097] FIG. 81 depicts an example Al-source code portion in accordance with at least some example embodiments of the present disclosure.

[0098] FIG. 8J depicts an example Al-source code portion in accordance with at least some example embodiments of the present disclosure.

[0099] FIGS 8A-8J depict a source code configured for a plurality of steps associated with a hybrid transformer model for predicting the probability of mutations. In at least some embodiments, the depicted source code is configured to create a list of box features (atom presence / absence, for example) and their respective labels. Such data may be extracted and readily available. In at least some embodiments, the depicted source code is configured to split the file list into train / validation sets, and loads and pre-processes them (for example, one hot encoding of labels, a subsampling of a random box from each protein, and the like). In at least some embodiments, the depicted source code is configured to create a ready-to-train batch in each step or iteration. Once an instance is created, then an Al model is implemented, compiled, and fit for each batch that is pulled from the instance. In at least some embodiments, the depicted source code is configured to calculate validation accuracy after a few epochs.IV. Example System Operations

[0100] FIG. 9 depicts an example process 900 for systemically characterizing structural impact(s) of known and new mutations in JAK2. As depicted, the process 900 includes receiving24LEGAL / 51273753vl(902) protein sequences and mutants, analyzing (904) the received protein sequences, generating (906) 3D models of the received mutants, filtering (908) sequences based on strong positive selection, filtering (910) mutants using energy minimization and MD simulations, conducting (912) in vitro validation on final candidates, and applying (914) the validated candidates to one or more precision medicine techniques. Example process 900 may be executed as part of an integrative framework to systemically characterize the structural impact of known and new mutations of a set of clinically relevant protein families.

[0101] At step 902, framework 100 is configured to receive protein sequences and mutants. Receiving (902) protein sequences and mutants may include gathering a defined number of protein sequences and a defined number of candidate mutants derived using a hybrid transformer model, such as hybrid transformer model 500. The hybrid transformer model combines feature embeddings with a 3DCNN to predict stabilizing and destabilizing mutations in proteins in a selfsupervised learning framework. In at least some embodiments, the model iteratively masks residues to train the model to predict missing amino acids associated with the one or more protein sequences, thereby learning representations that capture the general evolutionary patterns of protein stability, even without direct labels for all possible mutations.

[0102] At step 904, framework 100 is configured to analyze the received protein sequences. Analyzing (904) the received protein sequences may include identifying mutations from a phylogenetics perspective using MSA, MK tests, and phylogenetic analysis to assess conserved and variable residues / motifs across evolutionary lines. Through these analyses, the framework may identify critical differences caused by the mutations, ultimately providing a clearer picture of how these changes might influence the protein’s role in disease processes.

[0103] Analyzing (904) the received protein sequences to identify critical differences caused by the mutations may include a multi-faceted approach combining phylogenetic analysis, structural feature extraction, and binding affinity assessment. In at least some embodiments, analyzing (904) the received protein sequences includes utilizing the protein sequences to identify and analyze mutations from a phylogenetics perspective. The sequences may undergo multiple sequence alignment (MSA), McDonald-Kreitman (MK) tests, and phylogenetic analysis to assess conserved and variable residues / motifs across evolutionary lines. Through these analyses, the framework may identify critical differences caused by the mutations, ultimately providing a clearer picture of how these changes might influence the protein’s role in disease processes.25LEGAL / 51273753vl

[0104] In at least some embodiments, analyzing (904) the received protein sequences includes extracting and analyzing comprehensive structural features, including interatomic interactions and physicochemical properties that may change due to the candidate mutations. In at least some embodiments, the framework may leverage a toolkit configured to extract detailed atomic-level interaction features from 3D protein-ligand complexes, capturing hydrogen bonds, hydrophobic contacts, and van der Waals forces, along with intramolecular stability-related features. The toolkit may integrate explainable Al (XAI) to deliver interpretable insights into model predictions, revealing mechanistic understanding of the molecular determinants driving protein stability, function, and binding. In at least some embodiments, analyzing the received protein sequences includes performing energy minimization followed by equilibration, and extracting root mean square deviation (RMSD) to quantify structural stability and identify conformational changes.

[0105] In at least some embodiments, the framework may apply a binding affinity prediction tool to predict protein-inhibitor binding affinity (pKD) after molecular dynamics simulation, quantifying the effects caused by each mutation on the protein-inhibitor binding. Such analysis may comprise comparing the wildtype protein-inhibitor interaction with that of the mutant proteininhibitor complexes, which describes how mutations affect the binding and overall stability of the protein structure. In at least some embodiments, the features identified as significantly associated with protein structure-function-binding serve as the basis for modeling mutations that disrupt protein function and protein-inhibitor binding.

[0106] At step 906, framework 100 is configured to generate 3D models of the received mutants. Generating (906) 3D models of the received mutants may include using a protein structure prediction application, such as AlphaFold, to generate high-quality structural representations of the received mutants. In at least some embodiments, generating 3D models of the received mutants includes ranking the predicted mutants by their confidence score, which is represented by the model’s predicted probability for a given mutation, and selecting a defined number of top candidate mutants for further evaluation. For each predicted mutation, the framework may generate 3D models using a protein structure prediction application, ensuring high-quality structural representations. In at least some embodiments, generating 3D models of the received mutants includes generating a 3D model of a protein dimer in complex with a clinically validated inhibitor. The protein structure prediction application may yield a high-quality structure with over 95% of the protein showing a confidence score greater than 99%. In at least some embodiments,26LEGAL / 51273753vlthe mutant 3D models may be utilized for calculating changes in Gibbs free energy (AAG) relative to the wildtype protein, which quantifies the impact of the mutation on the protein’s structural stability.

[0107] Molecular dynamics (MD) simulations offer a physics-based approach grounded in Newtonian mechanics and quantum chemistry principles, providing a refinement method for filtering potentially biased predictions. In at least some embodiments, the framework may conduct energy minimization and MD simulations on both wildtype and mutant protein structures, including complexes with clinically relevant inhibitors. These simulations provide insights into conformational changes, protein stability, and drug interactions. MD simulations capture subtle conformational changes, solvent effects, and time-dependent phenomena that are often overlooked in simplified predictive models, yielding a more mechanistically insightful result for evaluating each candidate mutant. In at least some embodiments, the focus of generating 3D models comprises comparing the wildtype protein-inhibitor interaction with that of the mutant proteininhibitor complexes, which describes how mutations affect the binding and overall stability of the protein structure.

[0108] At step 908, framework 100 is configured to filter sequences based on strong positive selection. Filtering (908) sequences based on strong positive selection may include identifying candidate mutants under strong positive selection, since they are likely to destabilize the function of the protein(s), constitutively inhibiting the protein. Through these analyses, framework 100 is configured to identify critical differences caused by the mutations, thereby providing a clearer picture of how these changes might influence the protein’s role in disease processes.

[0109] In at least some embodiments, filtering (908) sequences based on strong positive selection includes subjecting the protein sequences to multiple sequence alignment (MSA), McDonald-Kreitman (MK) tests, and phylogenetic analysis to assess conserved and variable residues / motifs across evolutionary lines. The MSA may be configured to align homologous sequences across species to identify regions of conservation and variability. The MK test may be configured to compare the ratio of nonsynonymous to synonymous substitutions within and between species to detect signatures of selection pressure. In at least some embodiments, filtering sequences based on strong positive selection includes identifying and eliminating candidate mutants that are under strong positive selection, since these mutations are likely to destabilize the function of the protein, constitutively inhibiting the protein. Mutations under strong positive27LEGAL / 51273753vlselection may indicate evolutionary pressure that has driven functional changes, which may correlate with loss of normal protein function. In at least some embodiments, filtering sequences based on strong positive selection includes identifying family-wide conserved and divergent patterns of mutation impacts. Through these analyses, the framework may identify critical differences caused by the mutations, ultimately providing a clearer picture of how these changes might influence the protein’s role in disease processes.

[0110] In at least some embodiments, the list of proposed mutations and probabilities indicating whether these mutations will stabilize or destabilize the protein is refined via phylogenetic analysis before experimental validation. The most promising mutant candidates may be selected for in vitro testing from a consensus result combining structural / dynamics and sequence / phylogenetic based analyses extracted from the proposed workflow. In at least some embodiments, filtering (908) sequences based on strong positive selection may help identify pathogenic mutations that stabilize the constitutively activated conformation of proteins, which typical mutation screening tools may miss due to bias for knockout mutants in labeled training data.[OHl] At step 910, framework 100 is configured to filter mutants using energy minimization and MD simulations. Filtering (910) mutants using energy minimization and MD simulations may include conducting simulations on both wildtype structures and mutant protein structures, including complex structures with clinically relevant inhibitors. In at least some embodiments, filtering (910) mutants using energy minimization and MD simulations includes performing energy minimization to relieve any steric clashes and optimize the geometry of the protein-ligand complex. Energy minimization helps achieve a more realistic and stable conformation that reflects physiological conditions. In at least some embodiments, the energy minimization is followed by equilibration simulations, which allow the system to reach a stable state that mimics physiological conditions better than static docking poses.

[0112] In at least some embodiments, filtering (910) mutants using energy minimization and MD simulations includes performing energy minimization followed by equilibration, and extracting root mean square deviation (RMSD) to quantify structural stability and identify conformational changes. The dynamic nature of MD simulations can help refine the docking pose by allowing the ligand and protein to explore conformational space beyond what is captured in docking alone, leading to more accurate predictions of binding affinity (pKD). In at least some embodiments, filtering (910) mutants using energy minimization and MD simulations includes28LEGAL / 51273753vlgenerating plots of key properties from the simulation, including potential energy (showing the system’s energy minimization and stability over time), pressure (indicating system stability and proper pressure coupling), root mean square fluctuation (RMSF) (illustrating the flexibility of different regions in the protein or system), and radius of gyration (Rg) (showing the compactness of the protein structure over time).

[0113] In at least some embodiments, filtering (910) mutants using energy minimization and MD simulations includes applying a binding affinity prediction tool to predict protein-inhibitor binding affinity (pKD) after molecular dynamics simulation, quantifying the effects caused by each mutation on the protein-inhibitor binding. MD simulations capture subtle conformational changes, solvent effects, and time-dependent phenomena that are often overlooked in simplified predictive models, yielding a more mechanistically insightful result for evaluating each candidate mutant. In at least some embodiments, filtering ( 10) mutants using energy minimization and MD simulations includes extracting and analyzing comprehensive structural features from the MD trajectories, including interatomic interactions and physicochemical properties that may change due to the candidate mutations. The focus of such analysis comprises comparing the wildtype protein-inhibitor interaction with that of the mutant protein-inhibitor complexes, which describes how mutations affect the binding and overall stability of the protein structure.

[0114] In at least some embodiments, RMSD results that significantly differ from wildtype (such as those identified via Wilcoxon test FDR corrected p-value) may be used to identify mutations that potentially cause conformational changes in the protein’s structure. Such conformational changes may reduce the protein’s structural flexibility and disrupt protein-inhibitor contacts. The subject simulations will provide insights into conformational changes, protein stability, and drug interactions. Further yet, framework 100 may be configured to apply a GLM-score tool to detect changes in protein-inhibitor binding affinities. Framework 100 may be configured to extract and analyze comprehensive structural features from the MD trajectories, including interatomic interactions and physicochemical properties that may change due to the candidate mutations.

[0115] At step 912, framework 100 is configured to conduct in vitro validation on a set of final mutant candidates. Conducting (912) in vitro validation on the set of final mutant candidates may include experimentally verifying the predictions from both the baseline and the aforementioned candidate determinations. In at least some embodiments, conducting (912) in vitro validation29LEGAL / 51273753vlincludes selecting the most promising mutant candidates for testing from a consensus result combining structural / dynamics and sequence / phylogenetic based analyses extracted from the proposed workflow. The primary outcome of the in vitro validation is to identify functional variants that may significantly change protein stability and activity. Conducting (912) in vitro validation may include assessing the stability of the mutated proteins in comparison to the wildtype proteins through changes in solubility and assessing changes in kinase function by measuring changes in the phosphorylation of established protein substrates.

[0116] In at least some embodiments, wildtype and mutated proteins including relevant domains may be generated by a protein synthesis service. Conducting (912) in vitro validation may include utilizing basic biochemical techniques of Coomassie blue stain, BCA protein analysis, and ammonium sulfate precipitation to assess the stability of the mutated proteins compared to wildtype. Western blotting may also be incorporated for the detection of the subject protein. In at least some embodiments, conducting (912) in vitro validation includes subjecting a solution of wildtype or mutant protein to increasing concentrations of ammonium sulfate for a defined period of time. Precipitate and solute fractions may be separated by centrifugation, followed by dissolution of precipitate into PBS. Protein concentration within resuspended fraction and solute may be determined by BCA analysis subjecting the samples to gel electrophoresis. The fraction of wildtype or mutant protein precipitate generated at varied concentrations of ammonium sulfate may be determined by Coomassie blue stain, or western blot specific for the subject protein followed by quantitation.

[0117] In at least some embodiments, conducting (912) in vitro validation includes assessing the functionality of mutated proteins through changes in their kinase function via in vitro kinase assays. In at least some embodiments, conducting (912) in vitro validation includes performing an in vitro kinase assay, such as an ADP-Glo Kinase assay, to assess activity. The ADP-Glo kinase assay measures the generated ADP after the phosphate is transferred from ATP to the substrate. Controls may include an irrelevant peptide substrate and an inhibitor of protein phosphorylation as a negative control. The ADP-Glo assay works by two catalytic processes: (1) degradation of all existing ATP, and (2) all ADP arising from the use of ATP to phosphorylate the substrate is converted back to ATP. The ATP is then measured through a luciferase assay.

[0118] In at least some embodiments, conducting (912) in vitro validation includes validating the results of the kinase assay through an in vitro assay where the phosphorylation of a well-30LEGAL / 51273753vlestablished biological substrate (such as STAT1 for JAK2) is assessed using a phosphorylated-substrate antibody. In at least some embodiments, conducting (912) in vitro validation includes complementing biochemical assays with protein-dependent cell lines to assess the functional impact of mutations on downstream signaling pathways.

[0119] At step 914, framework 100 is configured to apply the validated candidates to one or more precision medicine techniques. Applying (914) the validated candidates to one or more precision medicine techniques may include designing therapeutic strategies targeting specific protein mutations. The insights gained from the integrative framework may have applications in precision medicine, allowing the design of more effective therapeutic strategies targeting specific mutations.

[0120] In at least some embodiments, designing therapeutic strategies targeting specific protein mutations includes identifying unique residues within a protein family that may serve as potential targets for developing selective inhibitors. For example, the framework may identify residues that are unique to a specific protein within a protein family around a given region, making such residues potential targets for developing selective inhibitors that distinguish between closely related family members.

[0121] In at least some embodiments, designing therapeutic strategies targeting specific protein mutations includes characterizing the structural, functional, and clinical implications of identified mutations. Certain mutations may alter local protein structure, affecting folding intermediates and ultimately leading to misfolding or dysfunction. These structural changes may have far-reaching effects on protein-inhibitor interactions, potentially explaining the emergence of drug resistance, ultimately hindering the efficacy of inhibitors in the treatment of associated diseases. By understanding the mechanistic impact of mutations on the protein and guiding future experiments for understanding the link between protein structure, function, and associated pathogenesis, the framework may inform the design of next-generation therapeutics. In at least some embodiments, designing therapeutic strategies targeting specific protein mutations includes quantitatively linking sequence, structure, dynamics, and inhibitor binding to predict novel dysfunction-driving mutations and systematically prioritize among existing variants of unknown significance (VUS) which variants should be escalated for functional and clinical follow-up.

[0122] In at least some embodiments, designing therapeutic strategies targeting specific protein mutations includes expanding the framework to include all members of a protein family, allowing31LEGAL / 51273753vlidentification of family-wide conserved and divergent patterns of mutation impacts. The framework may be configured to identify mutations that confer resistance to existing inhibitor therapies, thereby informing the development of alternative therapeutic approaches that may overcome such resistance mechanisms. In at least some embodiments, the framework may be configured to generate high-quality hypotheses that can be further clinically tested, building a strong foundation for biomedical research and personalized medicine with potential implications across a wide range of human diseases and biological phenomena.

[0123] In at least some embodiments, framework 100 may additionally be configured to execute one or more prediction-based actions based at least in part on the validated candidates. Such prediction-based actions may include automated alerts, automated instructions to insulin delivery devices, automated adjustments to medical equipment, automated adjustments to allocations of medical, computing, hospital, facility, and / or human resources. Further, such automated actions may include automated physician notification actions, automated patient notification actions, automated appointment scheduling actions, automated prescription recommendation actions, automated drug prescription generation actions, automated implementation of precautionary actions, automated record updating actions, automated datastore updating actions, automated hospital preparation actions, automated workforce management operational management actions, automated server load balancing actions, automated resource allocation actions, automated call center preparation actions, automated hospital preparation actions, automated pricing actions, automated plan update actions, automated alert generation actions, generating a diagnostic report, generating action scripts, generating one or more electronic communications, and / or the like. The one or more prediction-based actions may further include displaying visual renderings of the aforementioned examples of prediction-based actions in addition to values, charts, and representations associated with the one or more policy scores and the prediction output using a prediction output user interface.

[0124]

[0125] FIG. 10 depicts an example process 1000 for systemically characterizing structural impact(s) of known and new protein mutations in accordance with at least some embodiments of the present disclosure. As depicted, process 1000 includes receiving (1002) protein sequences and structural input, training (1004) a model to predict missing amino acids, capturing (1006) global features and structural information, integrating (1008) multimodal extracted features, capturing32LEGAL / 51273753vl(1010) relationships between parts of protein structure(s), determining (1012) probability for mutation impacts on stability, and re-training (1014) the model based on specific subsets of data. Example process 1000 may be executed as part of an integrative framework to systemically characterize the structural impact of known and new mutations of a set of clinically relevant protein families.

[0126] At step 1002, framework 100 is configured to receive protein sequences and structural input. Receiving (1002) protein sequences and structural input may include receiving and / or retrieving a subject number of protein sequences and corresponding structural inputs and / or structural information. In at least some embodiments, receiving (1002) protein sequences and structural input includes receiving a protein sequence and structural input for JAK2.

[0127] At step 1004, framework 100 is configured to train a model to predict missing amino acids. Training (1004) a model to predict missing amino acids may include iteratively masking residues at each given position and subsequently training the model to predict the masked residue. By training the model accordingly, the model can learn representations that capture the general evolutionary patterns of protein stability, even in the absence of direct labels for all possible mutations.

[0128] Training (1004) a model to predict missing amino acids may include iteratively masking residues at each given position and subsequently training the model to predict the masked residue. By training the model accordingly, the model can learn representations that capture the general evolutionary patterns of protein stability, even in the absence of direct labels for all possible mutations. In at least some embodiments, training (1004) a model to predict missing amino acids leverages a self-supervised learning framework in which the wildtype (WT) residue is masked iteratively at each given position and the model is trained to predict the masked residue. This strategy allows the model to learn representations that capture the general evolutionary patterns of protein stability, even without direct labels for all possible mutations. Unlike traditional supervised learning models that predict protein structural stability relying on labeled datasets (e.g., experimentally determined changes in free energy from single point mutations), the model can predict key mutations by capturing evolutionary, structural, and physicochemical patterns from a much broader dataset.

[0129] In at least some embodiments, training (1004) a model to predict missing amino acids includes training on a large unlabeled dataset of protein structures (such as 137,399 PDB structures33LEGAL / 51273753vlstratified into a 90-10% train-test split), rather than relying exclusively on experimentally determined labels. This self-supervised learning framework circumvents label scarcity while capturing evolutionary folding patterns. In at least some embodiments, training (1004) a model to predict missing amino acids includes a data generation process configured to create a list of box features (such as atom presence / absence) and their respective labels. The data generation process may split the file list into train / validation sets, and load and pre-process them (for example, one hot encoding of labels, subsampling a random box from each protein, and the like). The data generation process may yield a ready-to-train batch in each step or iteration. Once an instance is created, an Al model is implemented, compiled, and fit for each batch that is pulled from the instance. Validation accuracy may then be calculated after a defined number of epochs. In at least some embodiments, training (1004) a model to predict missing amino acids includes extracting relevant features from protein sequences and structures after masking. This includes encoding global feature vectors (such as 1D / 2D features like ASA / HB) into embeddings and capturing structural information using 3DCNN.

[0130] In at least some embodiments, training (1004) a model to predict missing amino acids includes, after initial training on a general dataset, fine-tuning the model for two specific tasks: predicting which mutations will enhance the stability of the protein and predicting mutations that will destabilize the protein structure. This fine-tuning is done by training the model with specific subsets of data for stable and unstable proteins, which allows the model to specialize in predicting mutants with these different outcomes. After fine-tuning the model by training and validating it in target subsets, the model can then predict new stabilizing and destabilizing mutations in a much more scalable and specialized approach, as well as assess the impact of known missense variations found across human populations. The model may also be configured to predict combinations of mutations.

[0131] At step 1006, framework 100 is configured to capture global features and structural information associated with the received protein sequences and structural input. Capturing (1006) global features and structural information may include leveraging one or more feature extraction techniques to identify the subject global features. In at least some embodiments, capturing global features may include encoding global feature vectors, such as 1 -dimensional and 2-dimensional features, into embeddings. In at least some embodiments, capturing structural information34LEGAL / 51273753vlassociated with the received protein sequences and structural input includes leveraging a 3DCNN to identify a plurality of 3D features.

[0132] In at least some embodiments, capturing (1006) global features and structural information includes extracting relevant features from protein sequences and structures after masking. This may include encoding global feature vectors, such as 1 -dimensional and 2-dimensional features (for example, accessible surface area (ASA) and hydrogen bonds (HB)), into embeddings. In at least some embodiments, capturing structural information includes leveraging a 3-dimensional convolutional neural network (3DCNN) to identify a plurality of 3D features from the protein structures. The 3DCNN may be configured to detect spatial and hierarchical patterns from 3D data such as protein structures.

[0133] In at least some embodiments, capturing (1006) global features and structural information includes processing the data through two parallel branches. A first branch may comprise a 3DCNN branch configured to process 3D structural input through a series of 3D convolutional layers with batch normalization and activation functions, followed by max pooling operations to progressively extract spatial features at multiple scales. The 3DCNN branch may include multiple convolutional layers with progressively varying kernel sizes (such as 9x9x9, 5x5x5, and 3x3x3) to capture structural patterns at different resolutions. A second branch may comprise a dense branch configured to process global features through normalization and dense layers with batch normalization and activation functions to encode the global feature vectors into embeddings.

[0134] In at least some embodiments, capturing (1006) global features and structural information includes leveraging a residue-aligned voxelization framework that replaces traditional global Euclidean grids with local spherical voxel maps centered on each amino acid, creating a rotationally consistent coordinate system. This residue-centric representation bridges global volumetric encoders and graph-based models, providing a versatile foundation for downstream tasks such as mutation effect prediction. In at least some embodiments, capturing (1006) global features and structural information includes extracting detailed atomic-level interaction features from 3D protein-ligand complexes, capturing hydrogen bonds, hydrophobic contacts, and van der Waals forces, along with intramolecular stability-related features. The features identified as significantly associated with protein structure-function-binding may serve as the basis for modeling mutations that disrupt protein function and protein-inhibitor binding.35LEGAL / 51273753vl

[0135] Engineering of knowledge-based interaction features may boost Al performance in the profding of protein structure and function. In at least some embodiments, the hybrid transformer model combines local and global structural features, where transformers excel at capturing long-range dependencies and contextual information within sequences, while 3DCNNs excel in detecting spatial and hierarchical patterns from 3D data such as protein structures. The extracted multimodal features from both branches may then be integrated into an early-fusion attention head mechanism for subsequent processing by the transformer architecture.

[0136] At step 1008, framework 100 is configured to integrate multimodal extracted features. Integrating (1008) multimodal extracted features may include leveraging a hybrid transformer architecture / hybrid transformer model to integrate all multimodal features into a multi-head attention mechanism. In at least some embodiments, integrating (1008) multimodal extracted features includes combining the outputs of two parallel processing branches. A first branch comprises a 3DCNN branch that processes 3D structural input through a series of 3D convolutional layers to extract spatial features. The output of the 3DCNN branch is passed through a transformer encoder block, which applies layer normalization and multi-head attention with a defined head size and number of heads (such as a head size of 32 and 4 attention heads), followed by a feedforward network with dropout for regularization. The transformer encoder includes residual connections and layer normalization to stabilize training. The output of the transformer encoder is then flattened. A second branch comprises a dense branch that processes global features through normalization and dense layers to encode the global feature vectors into embeddings. The outputs of both branches are then concatenated to produce a combined feature representation.

[0137] In at least some embodiments, integrating (1008) multimodal extracted features includes integrating the extracted multimodal features into an early-fusion attention mechanism. The early-fusion approach allows the model to jointly process and learn from both the 3D structural features and the global feature embeddings from the earliest stages of the attention computation, rather than processing them independently and combining them at a later stage. In at least some embodiments, integrating (1008) multimodal extracted features includes applying the multi-head attention mechanism to capture complex relationships between different parts of the protein structure. The multi-head attention mechanism captures long-range residue interactions, which is relevant for understanding how distant residues in the protein interact and boosts the transformer’s capacity to understand dependencies between interactions.36LEGAL / 51273753vl

[0138] In at least some embodiments, the concatenated feature representation is passed through one or more additional dense layers for further processing before being provided to the output layer. The model may be compiled and trained using adaptive learning rate scheduling, early stopping, and checkpoint callbacks to optimize performance and prevent overfitting. In at least some embodiments, the hybrid transformer model combines local and global structural features, where transformers excel at capturing long-range dependencies and contextual information within sequences, while 3DCNNs excel in detecting spatial and hierarchical patterns from 3D data such as protein structures. Iteratively replacing the 3DCNN architecture with the hybrid transformer may further enhance predictive performance.

[0139] At step 1010, framework 100 is configured to capture relationships between parts of the protein structure(s). Capturing (1010) relationships between parts of protein structure(s) may include leveraging a multi-head attention mechanism to capture complex relationships between different parts of the protein structures. In at least some embodiments, capturing (1010) relationships between parts of the protein structure(s) includes determining how distant residues in the protein interact, thereby boosting the model’s ability to understand dependencies between said interactions.

[0140] In at least some embodiments, capturing (1010) relationships between parts of protein structure(s) includes applying a multi-head attention mechanism multiple times in parallel, each time with a different set of learnable parameters, to capture different types of relationships between the protein structures and sequences. By doing so, the model attends to various aspects of the protein structures and sequences simultaneously, leading to richer representations and better overall performance. The multi-head attention mechanism captures long-range residue interactions, which is relevant for understanding how distant residues in the protein interact and boosts the transformer’s capacity to understand dependencies between interactions.

[0141] In at least some embodiments, capturing (1010) relationships between parts of protein structure(s) addresses a limitation of existing 3D convolutional neural network (3D CNN) approaches, which can only capture spatial patterns within a limited neighborhood around each residue in the protein, ignoring global features, long-range interactions, as well as how they are affected by the proposed mutations. Such a local view may be insufficient for predicting the effects of mutations in larger and more complex proteins, which require multimodal sources of data, detection of complex long-range interactions, and global conformational changes.37LEGAL / 51273753vl

[0142] In at least some embodiments, capturing (1010) relationships between parts of protein structure(s) includes processing the integrated multimodal features through a transformer encoder block. The transformer encoder applies layer normalization followed by multi-head attention with a defined head size and number of heads (such as a head size of 32 and 4 attention heads). The multi-head attention computes self-attention over the input, allowing each position to attend to all other positions. The output of the multi-head attention is combined with the input via a residual connection and further normalized. A feed-forward network comprising one-dimensional convolutional layers with activation functions and dropout is then applied, followed by another residual connection and layer normalization.

[0143] In at least some embodiments, capturing (1010) relationships between parts of protein structure(s) enables the framework to quantitatively link sequence, structure, dynamics, and inhibitor binding, thereby predicting novel dysfunction-driving mutations and systematically prioritizing among existing variants of unknown significance (VUS) which variants should be escalated for functional and clinical follow-up. In at least some embodiments, the effects of interdomain interactions that are relevant for protein folding, stability, and function may be considered, which typical methods for predicting the impact of mutations are unable to capture.

[0144] At step 1012, framework 100 is configured to determine probability for mutation impacts on stability. Determining (1012) probability for mutation impacts on stability may include assessing the stability of the mutated proteins in comparison to the wildtype proteins through changes in solubility and assessing changes in kinase function by measuring changes in the phosphorylation of established protein substrates. The fraction of wildtype or mutant protein precipitate generated at varied concentrations of ammonium sulfate is determined by Coomassie blue stain or western blot followed by quantification. Determining (1012) probability for mutation impacts on stability may further include assessing functionality of mutated proteins through changes in their kinase function via in vitro kinase assays.

[0145] Determining (1012) probability for mutation impacts on stability may include generating one or more probability scores associated with the one or more protein sequences, the one or more probability scores indicating a likelihood of mutation impacts on stability of the associated protein structure. In at least some embodiments, determining (1012) probability for mutation impacts on stability includes generating, via the output layer, probability scores for each of the possible amino acids (AA) for each residue in a subject protein sequence. The output layer38LEGAL / 51273753vlrepresents the probability scores for each of the possible residues for each position in the protein sequence.

[0146] In at least some embodiments, determining (1012) probability for mutation impacts on stability includes generating a list of proposed mutations and probabilities indicating whether these mutations will stabilize or destabilize the protein. In at least some embodiments, determining (1012) probability for mutation impacts on stability includes ranking the predicted mutants by their confidence score, which is represented by the model’s predicted probability for a given mutation, and selecting a defined number of top candidate mutants for further evaluation. In at least some embodiments, the mutant 3D models may be utilized for calculating changes in Gibbs free energy (A A G) relative to the wildtype protein, which quantifies the impact of the mutation on the protein’s structural stability. Stabilizing mutations may be characterized by A A G < - 0.5 kcal / mol, whereas destabilizing mutations may be characterized by A A G > 0.5 kcal / mol.

[0147] In at least some embodiments, determining (1012) probability for mutation impacts on stability includes quantitatively linking sequence, structure, dynamics, and inhibitor binding to predict novel dysfunction-driving mutations and systematically prioritize among existing variants of unknown significance (VUS) which variants should be escalated for functional and clinical follow-up. In at least some embodiments, the results comprise of a fine-tuned set of candidate mutations that can be further validated in silico and in vitro, ultimately helping the fundamental understanding of key mutations that affect protein structural dynamics. The list of proposed mutations and probabilities can be further refined via MD simulations and phylogenetic analysis before experimental validation.

[0148] At step 1014, framework 100 is configured to re-train the model based on specific subsets of data. Re-training (1014) the model based on specific subsets of data may include fine-tuning the model for two specific tasks: predicting which mutations will enhance the stability of the protein and predicting which mutations will destabilize the protein structure. In at least some embodiments, re-training (1012) the model based on specific subsets of data includes training the model with specific subsets of data for stable and unstable proteins, which allows the model to specialize in predicting mutants with these different outcomes. The final output may comprise of a list of proposed mutations and probabilities indicating whether these mutations will stabilize or destabilize the protein. The list can be further refined via MD simulations and phylogenetic analysis before experimental validation.39LEGAL / 51273753vl

[0149] In some embodiments, the framework 100 is configured to execute one or more prediction-based actions responsive to the determined probability for mutation impacts on stability. Such prediction-based actions may include automated alerts, automated instructions to insulin delivery devices, automated adjustments to medical equipment, automated adjustments to allocations of medical, computing, hospital, facility, and / or human resources. Further, such automated actions may include automated physician notification actions, automated patient notification actions, automated appointment scheduling actions, automated prescription recommendation actions, automated drug prescription generation actions, automated implementation of precautionary actions, automated record updating actions, automated datastore updating actions, automated hospital preparation actions, automated workforce management operational management actions, automated server load balancing actions, automated resource allocation actions, automated call center preparation actions, automated hospital preparation actions, automated pricing actions, automated plan update actions, automated alert generation actions, generating a diagnostic report, generating action scripts, generating one or more electronic communications, and / or the like. The one or more prediction-based actions may further include displaying visual renderings of the aforementioned examples of prediction-based actions in addition to values, charts, and representations associated with the one or more policy scores and the prediction output using a prediction output user interface.V. Conclusion

[0150] Many modifications and other embodiments will come to mind to one skilled in the art to which this disclosure pertains having the benefit of the teachings presented in the foregoing descriptions and the associated drawings. Therefore, it is to be understood that the disclosure is not to be limited to the specific embodiments disclosed and that modifications and other embodiments are intended to be included within the scope of the appended claims. Although specific terms are employed herein, they are used in a generic and descriptive sense only and not for purposes of limitation.40LEGAL / 51273753vl

Claims

CLAIMS1. A computer-implemented method comprising:receiving, by one or more processors, one or more protein sequences and one or more structural inputs associated with the one or more protein sequences;iteratively masking, by the one or more processors, residues to train a model to predict missing amino acids associated with the one or more protein sequences;capturing, by the one or more processors, one or more global features and associated protein structure of the one or more protein sequences;integrating, by the one or more processors and using a hybrid transformer model architecture, the one or more global features and the associated protein structure;capturing, by the one or more processors and using a multi-head attention mechanism, at least one relationship between a set of parts of the associated protein structure;determining, by the one or more processors, one or more probability scores associated with the one or more protein sequences, the one or more probability scores indicating a likelihood of mutation impacts on stability of the associated protein structure; andre-training, by the one or more processors, the model based on one or more specific subsets of data.

2. The computer-implemented method of claim 1, wherein re-training the model based on one or more specific subsets of data includes fine-tuning the model using a first subset of data comprising one or more stable proteins.

3. The computer-implemented method of claim 1, wherein re-training the model based on one or more specific subsets of data includes fine-tuning the model using a second subset of data comprising one or more unstable proteins.

4. The computer-implemented method of claim 1, wherein the hybrid transformer model architecture comprises a 3-dimensional convolutional neural network (3DCNN).41LEGAL / 51273753vl5. The computer-implemented method of claim 1, wherein the hybrid transformer model architecture is configured to generate one or more global feature embeddings.

6. The computer-implemented method of claim 1, further comprising conducting, by the one or more processors, energy minimization and molecular dynamics (MD) simulation on the one or more protein sequences.

7. The computer-implemented method of claim 1, further comprising generating, by the one or more processors, a ranking of the one or more protein sequences based at least in part on the one or more probability scores.

8. A computing apparatus comprising memory and one or more processors communicatively coupled to the memory, the one or more processors configured to perform operations comprising:receiving one or more protein sequences and one or more structural inputs associated with the one or more protein sequences;iteratively masking residues to train a model to predict missing amino acids associated with the one or more protein sequences;capturing one or more global features and associated protein structure of the one or more protein sequences;integrating, using a hybrid transformer model architecture, the one or more global features and the associated protein structure;capturing, using a multi-head attention mechanism, at least one relationship between a set of parts of the associated protein structure;determining one or more probability scores associated with the one or more protein sequences, the one or more probability scores indicating a likelihood of mutation impacts on stability of the associated protein structure; andre-training the model based on one or more specific subsets of data.42LEGAL / 51273753vl9. The computing apparatus of claim 8, wherein re-training the model based on one or more specific subsets of data includes fine-tuning the model using a first subset of data comprising one or more stable proteins.

10. The computing apparatus of claim 8, wherein re-training the model based on one or more specific subsets of data includes fine-tuning the model using a second subset of data comprising one or more unstable proteins.

11. The computing apparatus of claim 8, wherein the hybrid transformer model architecture comprises a 3-dimensional convolutional neural network (3DCNN).

12. The computing apparatus of claim 8, wherein the hybrid transformer model architecture is configured to generate one or more global feature embeddings.

13. The computing apparatus of claim 8, the operations further comprising conducting energy minimization and molecular dynamics (MD) simulation on the one or more protein sequences.

14. The computing apparatus of claim 8, the operations further comprising generating a ranking of the one or more protein sequences based at least in part on the one or more probability scores.

15. One or more non-transitory computer-readable storage media including instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising:receiving one or more protein sequences and one or more structural inputs associated with the one or more protein sequences;iteratively masking residues to train a model to predict missing amino acids associated with the one or more protein sequences;capturing one or more global features and associated protein structure of the one or more protein sequences;43LEGAL / 51273753vlintegrating, using a hybrid transformer model architecture, the one or more global features and the associated protein structure;capturing, using a multi-head attention mechanism, at least one relationship between a set of parts of the associated protein structure;determining one or more probability scores associated with the one or more protein sequences, the one or more probability scores indicating a likelihood of mutation impacts on stability of the associated protein structure; andre-training the model based on one or more specific subsets of data.

16. The one or more non-transitory computer-readable storage media of claim 15, wherein re-training the model based on one or more specific subsets of data includes fine-tuning the model using a first subset of data comprising one or more stable proteins.

17. The one or more non-transitory computer-readable storage media of claim 15, wherein re-training the model based on one or more specific subsets of data includes fine-tuning the model using a second subset of data comprising one or more unstable proteins.

18. The one or more non-transitory computer-readable storage media of claim 15, wherein the hybrid transformer model architecture comprises a 3-dimensional convolutional neural network (3DCNN).

19. The one or more non-transitory computer-readable storage media of claim 15, wherein the hybrid transformer model architecture is configured to generate one or more global feature embeddings.

20. The one or more non-transitory computer-readable storage media of claim 15, the operations further comprising generating a ranking of the one or more protein sequences based at least in part on the one or more probability scores.44LEGAL / 51273753vl