PROtein Analytics for Kinase Therapeutic Inhibitor Variants (PROAKTIV): A Machine learning approach to predict efficacy of tyrosine kinase inhibitors against mutations in EGFR, ALK and BRAF

PROtein Analytics for Kinase Therapeutic Inhibitor Variants (PROAKTIV): A Machine learning approach to predict efficacy of tyrosine kinase inhibitors against mutations in EGFR, ALK and BRAF | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Short Report PROtein Analytics for Kinase Therapeutic Inhibitor Variants (PROAKTIV): A Machine learning approach to predict efficacy of tyrosine kinase inhibitors against mutations in EGFR, ALK and BRAF Harold Mateo Mojica Urrego, Matthew Groves, Anthonie Van der Wekken, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8730878/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Introduction: The efficacy of targeted therapies in non-small cell lung cancer (NSCLC) is challenged by acquired resistance, driven by on-target mutations in kinases such as EGFR, ALK, and BRAF. Thus, predicting the functional impact of these mutations on tyrosine kinase inhibitor (TKI) sensitivity is critical for personalized treatment. This study presents a foundational machine learning framework for predicting ligand bioactivity against mutated kinases from sequence data, using EGFR, ALK, and BRAF as proof-of-concept models. The results demonstrate the feasibility of developing a generalized predictive model applicable across the kinase family. Methods An automated Python pipeline was developed to curate mutation and bioactivity data from a comprehensive dataset of 25,412 published in vitro pIC50 values for EGFR, ALK, and BRAF variants. Twelve deep learning architectures were trained and evaluated with different encoders for proteins and ligands. Model performance was assessed using Mean Squared Error (MSE), R-squared (R²), and Pearson Correlation Coefficient (PCC), with uncertainty quantified via Monte Carlo dropout. Results The best-performing model demonstrated robust predictive accuracy, providing a pearson correlation of 85% on the mutation/TKI pairs. Model predictions for clinically relevant drug-mutation pairs consistently aligned with established clinical outcomes, including EGFR T790M-mediated resistance to first-generation inhibitors, ALK G1202R resistance to crizotinib, and BRAF V600E sensitivity to selective inhibitors. A protein sequence language model (ESM2) offered improved predictions for complex, rare variants. Conclusions This study introduces a foundational machine learning framework for predicting the impact of kinase mutations on in vitro drug sensitivity. This methodology supports personalized therapy development in NSCLC and may enhance the efficiency of drug discovery pipelines. Biological sciences/Cancer Biological sciences/Computational biology and bioinformatics Biological sciences/Drug discovery Non-Small Cell Lung Cancer Targeted Therapy Machine Learning EGFR ALK BRAF Protein Language Model Figures Figure 1 Figure 2 Figure 3 Figure 4 INTRODUCTION Non-small cell lung cancer (NSCLC) accounts for approximately 85% of all lung cancer cases and remains the leading cause of cancer-related mortality worldwide.¹ Its profound molecular heterogeneity, characterized by a complex landscape of oncogenic driver mutations, poses a significant challenge to effective treatment.²,³ The paradigm of personalized medicine, which matches drug therapies to specific molecular drivers, has transformed outcomes for select patient populations.⁴ Key among these drivers are mutations in protein kinases that govern critical pro-survival signaling pathways, including the Epidermal Growth Factor Receptor (EGFR),⁵ Anaplastic Lymphoma Kinase (ALK),⁶ and B-Raf (BRAF).⁷ Activating mutations in EGFR, such as exon 19 deletions or the L858R point mutation, are found in 10–40% of NSCLC adenocarcinomas and confer sensitivity to first-generation tyrosine kinase inhibitors (TKIs) like gefitinib and erlotinib.⁸ Similarly, chromosomal rearrangements resulting in EML4-ALK fusion proteins define a distinct NSCLC subtype highly responsive to ALK inhibitors such as crizotinib.⁶,⁹ Downstream of these receptors, mutations in the BRAF serine/threonine kinase, particularly the V600E substitution, provide a direct and powerful engine for proliferation that can be targeted with selective inhibitors like dabrafenib and vemurafenib.⁷,¹⁰ Despite the initial success of these targeted therapies, their long-term efficacy is frequently limited by the emergence of on-target resistance mutations.¹¹ For example, the EGFR T790M "gatekeeper" mutation impacts the binding of first- and second-generation TKIs, leading to treatment failure.¹² The development of the third-generation inhibitor osimertinib successfully overcame T790M-mediated resistance,¹³ but its efficacy is now challenged by the subsequent acquisition of the C797S mutation at its covalent binding site.¹⁴,¹⁵ A similar evolutionary arms race is observed in ALK-positive NSCLC, where initial resistance to crizotinib via mutations like L1196M prompted the development of next-generation inhibitors, only to be met by further resistance from mutations such as G1202R. 12 , 16 , 17 This dynamic interplay highlights a critical need for predictive tools that can anticipate the effect of both well-characterized and rare kinase mutations on drug binding efficacy. While experimental and structure-based methods can offer deep mechanistic insights, 18 , significant time is required to provide robust in vitro pIC50 (-log10(IC50 M)) values that reflect the efficacy of a specific TKI on its target. This reliance provides a restriction in both the timely selection of a drug for rare mutations and in the creation of new drug-like molecules. To address this gap we propose a machine learning model (PROAKTIV; PROtein Analytics for Kinase Therapeutic Inhibitor Variants), a sequence-based machine learning approach to predict ligand bioactivity against mutated kinases. We finetuned the 3-billion-parameter ESM-2 protein language model to predict drug–target interactions in NSCLC-associated proteins (EGFR, ALK, BRAF) 23 . MATERIALS AND METHODS Data and Code Availability The complete code for the data curation pipeline, model training, and evaluation is publicly available on GitHub at https://github.com/HaroldMate1/proaktiv . Data Acquisition and Preprocessing An automated Python workflow was developed to retrieve data for human EGFR (UniProt ID: P00533), ALK (Q9UM72), and BRAF (P15056) from the UniProt database (Release 2025_02) 19 and ChEMBL (Release 33). 20 The workflow programmatically retrieved all available bioactivity data from the scientific literature, filtering for 'IC50' as the standard measurement type. Records with missing values or duplicates were removed, and IC50 values were converted to pIC50. Mutation Processing and Sequence Generation Mutation information was extracted from the 'assay description' and 'document title' fields in the ChEMBL metadata using regular expression algorithms. Extracted mutation strings were parsed and standardized to Human Genome Variation Society (HGVS) nomenclature. A custom function then programmatically applied these mutations, including single nucleotide polymorphisms (SNPs), insertions, and deletions, to the canonical wild-type FASTA sequences, generating a comprehensive library of variant protein sequences. Model Development and Evaluation The selection of deep learning architectures was guided by the benchmarking performed by Shi et al., who evaluated 80 encoder combinations to identify optimal configurations for EGFR-TKI sensitivity prediction 20 . Building upon their identification of high-performing encoder types, we constructed a comparative matrix of twelve models. However, this study diverged from their methodology in two key aspects: first, we utilized a custom-curated dataset rather than the clinical patient case database employed in the reference study; second, we substituted the standard Transformer protein encoder with a fine-tuned ESM2 protein language model to generate context-aware embeddings. Consequently, the final analysis evaluated pairings of four ligand representation strategies with three protein encoding architectures. Ligands were processed using either fixed representations (Morgan and Daylight fingerprints²²) or learnable architectures (Message Passing Neural Networks [MPNN] and Convolutional Neural Networks [CNN] applied to SMILES²¹). Simultaneously, protein sequences were encoded using three distinct architectures: a standard CNN, a hybrid CNN-RNN, and a Transformer model incorporating the pre-trained ESM2 language model. This resulted in a matrix of twelve model combinations. All models were implemented in PyTorch and randomly split into training (70%), test (20%), and validation (10%) sets. Training minimized Mean Squared Error (MSE) using the Adam optimizer²⁴, and performance was evaluated using MSE, R-squared (R²), and the Pearson Correlation Coefficient (PCC). Prediction uncertainty was quantified using Monte Carlo (MC) dropout²⁵ (see Supplementary Note 1 for detailed specifications of each encoder). RESULTS Automated Pipeline Curates a Large-Scale Bioactivity Dataset The automated pipeline successfully curated 25,412 high-quality bioactivity data points across EGFR (15,588), ALK (3,246), and BRAF (6,578). This dataset encompassed 10,074 unique ligands and dozens of distinct mutations, including clinically critical SNPs (T790M, G1202R, V600E), insertions, deletions (delE746_A750), and complex compound mutations (e.g., EGFR C797S/L858R/T790M). Exploratory data analysis confirmed that the dataset faithfully recapitulated known clinical patterns of sensitivity and resistance. CNN-RNN Benchmark Model Establishes Strong Predictive Performance To establish a robust performance benchmark, we systematically evaluated twelve different deep learning architectures. The combination of Morgan fingerprints for ligand encoding and a CNN-RNN for protein encoding consistently yielded the highest predictive accuracy. As shown in Fig. 1 , this benchmark configuration achieved a Pearson Correlation Coefficient of 0.85 on the EGFR test set, outperforming all other encoder combinations. Fine-Tuned ESM2 Language Model Enhances Predictions for Clinical Utility The final benchmark model, trained on the combined dataset, demonstrated robust generalization with a PCC of 0.82, an R² of 0.63, and an MSE of 0.78 on the test set (Fig. 2 ). To advance beyond this benchmark and create a tool better suited for clinical application on rare variants, we evaluated the fine-tuned 3B-parameter ESM2 model. This state-of-the-art approach achieved comparable overall performance (PCC = 0.7900 on the EGFR test set) and, crucially, demonstrated improved predictive capabilities for complex and less-common mutations not well-represented in the training data (Fig. 3 ). Model Predictions Align with Clinical Drug Sensitivity Profiles To assess the clinical relevance of the PROAKTIV framework, we used the advanced ESM2 model to predict pIC50 values for clinical EGFR inhibitors against a wide array of EGFR mutations. The predictions accurately mirrored established drug sensitivity and resistance profiles (Fig. 3 ). It correctly predicted that first- and second-generation TKIs would be potent against activating mutations but lose efficacy against T790M mutants. Furthermore, it predicted that osimertinib would overcome T790M-mediated resistance but fail against C797S mutations. Critically, the model predicted that the fourth-generation TKI candidate, BLU-945, would retain high potency against T790M and C797S-containing triple mutants, consistent with its design and preclinical data. This demonstrates the model's ability to generalize to complex scenarios. DISCUSSION In this study, we introduced the PROAKTIV framework, a machine learning approach to predict the bioactivity of inhibitors against mutated kinases. Our primary finding is that deep learning models, trained on systematically curated public data, can predict functional consequences of diverse mutations in key proteins implicated in NSCLC pathogenesis. This work moves beyond academic validation and provides a tangible framework with direct implications for clinical practice. For the clinician facing a patient with a rare or complex mutation profile for which treatment guidelines are non-existent, our approach offers a data-driven method to generate therapeutic hypotheses. Modern machine learning approaches have achieved the accuracy required to complement experimental studies by condensing a vast body of scientific knowledge into actionable analyses for matching drugs to specific mutations. The PROAKTIV framework exemplifies this leap forward, particularly in supporting the complex decisions faced by Molecular Tumor Boards (MTBs). When next-generation sequencing reveals a rare variant or a complex resistance mutation with no clear treatment path, our model can process extensive bioactivity data to generate a ranked list of inhibitors based on predicted in vitro efficacy (IC50). For instance, the model's ability to predict retained efficacy of the fourth generation TKIs against the EGFR C797S triple mutant provides a data-driven, actionable hypothesis where standard options are exhausted. 30 This shifts patient management from empirical trial-and-error towards a more rational, personalized strategy. The predicted pIC50 values function as a relative ranking system, designed to prioritize therapeutic options for a specific genetic context rather than serve as absolute measures of clinical potency, Fig. 4 . While standard architectures, such as our benchmark CNN-RNN, perform robustly on well-documented mutations, they often struggle to extrapolate beyond their training distribution. Literature has consistently shown that models relying solely on supervised learning with limited datasets tend to overfit to specific sequence identities, resulting in poor predictive power for 'out-of-distribution' variants. This limitation is addressed by the fine-tuned ESM2 model, which leverages embeddings derived from the evolutionary history of millions of protein sequences rather than simple sequence similarity. This capability is essential for characterizing rare mutations that lack experimental precedence. From a clinical perspective, the associated uncertainty quantification transforms these predictions into actionable insights: a high uncertainty score does not imply model failure, but rather identifies a unique biological context, signaling the need for confirmatory testing before clinical consideration. From a clinical standpoint, it is essential to frame these predictions within their limitations. The model’s analysis is based on in vitro bioactivity and does not account for the complexities of in vivo pharmacology, such as drug bioavailability, metabolism, or the influence of the tumor microenvironment. As such, the PROAKTIV framework is designed as a decision-support tool that complements, rather than replaces, clinical expertise and experimental validation. Its purpose is to initiate and inform the clinical conversation with data-driven hypotheses, not to provide definitive therapeutic answers. The path to clinical implementation requires a clear validation strategy. The immediate next step is a prospective study comparing PROAKTIV’s predictions in parallel with the recommendations of an active MTB. Establishing concordance with expert consensus and ultimately correlating predictions with patient outcomes are essential for building clinical trust. Future iterations will be enhanced by integrating 3D structural information, providing a more mechanistic understanding of how mutations alter drug binding.³¹ In conclusion, PROAKTIV provides a robust computational tool that condenses the vast landscape of genomic and pharmacological data into actionable insights. The advanced ESM2 model, in particular, holds considerable promise for realizing the vision of precision oncology, empowering clinicians to make more informed decisions for patients with both common and rare tumor mutations. Declarations Conflict of Interest Statement: The authors declare no potential conflicts of interest. Credit Author Statement Harold Mateo Mojica Urrego: Conceptualization, Resources, Data Curation, Formal Analysis, Investigation, Visualization, Methodology, Writing – Original Draft. Matthew Groves: Supervision, Formal Analysis, Writing – Review & Editing. Anthonie van der Wekken: Supervision, Formal Analysis, Writing – Review & Editing. Juvenal Yosa Reyes: Conceptualization, Investigation, Methodology, Formal Analysis, Writing – Review & Editing. Funding: This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Author Contribution Harold Mateo Mojica Urrego: Conceptualization, Resources, Data Curation, Formal Analysis, Investigation, Visualization, Methodology, Writing – Original Draft.Matthew Groves: Supervision, Formal Analysis, Writing – Review & Editing.Anthonie van der Wekken: Supervision, Formal Analysis, Writing – Review & Editing.Juvenal Yosa Reyes: Conceptualization, Investigation, Methodology, Formal Analysis, Writing – Review & Editing. Data Availability The complete code for the data curation pipeline, model training, and evaluation is publicly available on GitHub at https://github.com/HaroldMate1/proaktiv. References Bray F, Laversanne M, Weiderpass E, Soerjomataram I. The ever-increasing importance of cancer as a leading cause of premature death worldwide. Cancer . 2021;127(16):3029–3030. Pikor LA, Ramnarine VR, Lam S, Lam WL. Genetic alterations defining NSCLC subtypes and their therapeutic implications. Lung Cancer . 2013;82(2):179–189. Chen Z, Fillmore CM, Hammerman PS, Kim CF, Wong K-K. Non-small-cell lung cancers: a heterogeneous set of diseases. Nat Rev Cancer . 2014;14(8):535–546. Mok TS. Personalized medicine in lung cancer: what we need to know. Nat Rev Clin Oncol . 2011;8(11):661–668. Sharma SV, Bell DW, Settleman J, Haber DA. Epidermal growth factor receptor mutations in lung cancer. Nat Rev Cancer . 2007;7(3):169–181. Soda M, Choi YL, Enomoto M, et al. Identification of the transforming EML4-ALK fusion gene in non-small-cell lung cancer. Nature . 2007;448(7153):561–566. Davies H, Bignell GR, Cox C, et al. Mutations of the BRAF gene in human cancer. Nature . 2002;417(6892):949–954. Paez JG, Jänne PA, Lee JC, et al. EGFR mutations in lung cancer: correlation with clinical response to gefitinib therapy. Science . 2004;304(5676):1497–1500. Shaw AT, Yeap BY, Mino-Kenudson M, et al. Clinical features and outcome of patients with non-small-cell lung cancer who harbor EML4-ALK. J Clin Oncol . 2009;27(26):4247–4253. Planchard D, Smit EF, Groen HJ, et al. Dabrafenib plus trametinib in patients with previously untreated BRAFV600E-mutant metastatic non-small-cell lung cancer: an open-label, phase 2 trial. Lancet Oncol . 2017;18(10):1307–1316. Rotow J, Bivona TG. Understanding and targeting resistance mechanisms in NSCLC. Nat Rev Cancer . 2017;17(11):637–658. Pao W, Miller VA, Politi KA, et al. Acquired resistance of lung adenocarcinomas to gefitinib or erlotinib is associated with a second mutation in the EGFR kinase domain. PLoS Med . 2005;2(3):e73. Jänne PA, Yang JC-H, Kim D-W, et al. AZD9291 in EGFR inhibitor-resistant non-small-cell lung cancer. N Engl J Med . 2015;372(18):1689–1699. Thress KS, Paweletz CP, Felip E, et al. Acquired EGFR C797S mutation mediates resistance to AZD9291 in non-small cell lung cancer harboring EGFR T790M. Nat Med . 2015;21(6):560–562. Jia Y, Yun C-H, Park E, et al. Overcoming EGFR(T790M) and EGFR(C797S) resistance with mutant-selective allosteric inhibitors. Nature . 2016;534(7605):129–132. Katayama R, Shaw AT, Khan TM, et al. Mechanisms of acquired crizotinib resistance in ALK-rearranged lung cancers. Sci Transl Med . 2012;4(120):120ra17. Gainor JF, Dardaei L, Yoda S, et al. Molecular mechanisms of resistance to first-and second-generation ALK inhibitors in ALK-rearranged lung cancer. Cancer Discov . 2016;6(10):1118–1133. Robichaux JP, Le X, Vijayan R, et al. Structure-based classification predicts drug response in EGFR-mutant NSCLC. Nature . 2021;597(7878):732–737. UniProt Consortium. UniProt: the universal protein knowledgebase in 2023. Nucleic Acids Res . 2023;51(D1):D523-D531. Gaulton A, Hersey A, Nowotka M, et al. The ChEMBL database in 2017. Nucleic Acids Res . 2017;45(D1):D945-D954. Weininger D. SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules. J Chem Inf Comput Sci . 1988;28(1):31–36. Rogers D, Hahn M. Extended-connectivity fingerprints. J Chem Inf Model . 2010;50(5):742–754. Lin Z, Akin H, Rao R, et al. Language models of protein sequences at the scale of evolution. bioRxiv . 2022. doi: 10.1101/2022.07.20.500902 . Kingma DP, Ba J. Adam: A method for stochastic optimization. arXiv . 2014. arXiv:1412.6980. Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: a simple way to prevent neural networks from overfitting. J Mach Learn Res . 2014;15:1929–1958. Huang K, Fu T, Glass LM, Zitnik M, Xiao C, Sun J. DeepPurpose: a deep learning library for drug-target interaction prediction. Bioinformatics . 2020;36(22–23):5545–5547. Leonetti A, Sharma S, Minari R, Perego P, Giovannetti E, Tiseo M. Resistance mechanisms to osimertinib in EGFR-mutated non-small cell lung cancer. Br J Cancer . 2019;121(9):725–737. Addeo A, Tabbò F, Robinson T, Buffoni L, Novello S. Precision medicine in ALK rearranged NSCLC: A rapidly evolving scenario. Crit Rev Oncol Hematol . 2018;122:150–156. Leonetti A, Facchinetti F, Rossi G, et al. BRAF in non-small cell lung cancer (NSCLC): pickaxing another brick in the wall. Cancer Treat Rev . 2018;66:82–94. Koopman B, van der Wekken AJ, Ter Elst A, et al. Relevance and effectiveness of molecular tumor board recommendations for patients with non-small-cell lung cancer with rare or complex mutational profiles. JCO Precis Oncol . 2020;4:393–410. Galdadas I, Carlino L, Ward RA, Hughes SJ, Haider S, Gervasio FL. Structural basis of the effect of activating mutations on the EGF receptor. Elife . 2021;10:e65824. Meier J, Rao R, Verkuil R, et al. Language models enable zero-shot prediction of the effects of mutations on protein function. Adv Neural Inf Process Syst . 2021;34:29287–29303. Kris MG, Johnson BE, Berry LD, et al. Using multiplexed assays of oncogenic drivers in lung cancers to select targeted drugs. JAMA . 2014;311(19):1998–2006. Wong CH, Siah KW, Lo AW. Estimation of clinical trial success rates and related parameters. Biostatistics . 2019;20(2):273–286. Chiarle R, Voena C, Ambrogio C, Piva R, Inghirami G. The anaplastic lymphoma kinase in the pathogenesis of cancer. Nat Rev Cancer . 2008;8(1):11–23. Hallberg B, Palmer RH. Mechanistic insight into ALK receptor tyrosine kinase in human cancer biology. Nat Rev Cancer . 2013;13(10):685–700. Solomon BJ, Liu G, Felip E, et al. Lorlatinib vs crizotinib in treatment-naïve patients with advanced ALK + non-small cell lung cancer: 5-year progression-free survival and safety from the CROWN study. J Clin Oncol . 2024;42(suppl 17):LBA8507. Wan PT, Garnett MJ, Roe SM, et al. Mechanism of activation of the RAF-ERK signaling pathway by oncogenic mutations of B-RAF. Cell . 2004;116(6):855–867. Cross DA, Ashton SE, Ghiorghiu S, et al. AZD9291, an irreversible EGFR TKI, overcomes T790M-mediated resistance to EGFR inhibitors in lung cancer. Cancer Discov . 2014;4(9):1046–1061. Friboulet L, Li N, Katayama R, et al. The ALK inhibitor ceritinib overcomes crizotinib resistance in non-small cell lung cancer. Cancer Discov . 2014;4(6):662–673. Camidge DR, Kim HR, Ahn M-J, et al. Brigatinib versus crizotinib in ALK-positive non-small-cell lung cancer. N Engl J Med . 2018;379(21):2027–2039. Hyman DM, Puzanov I, Subbiah V, et al. Vemurafenib in multiple nonmelanoma cancers with BRAF V600 mutations. N Engl J Med . 2015;373(8):726–736. Soria JC, Ohe Y, Vansteenkiste J, et al. Osimertinib in untreated EGFR-mutated advanced non-small-cell lung cancer. N Engl J Med . 2018;378(2):113–125. Casaluce F, Sgambato A, Maione P, et al. ALK inhibitors: a new targeted therapy in the treatment of advanced NSCLC. Target Oncol . 2013;8(1):55–67. Roskoski Jr R. Anaplastic lymphoma kinase (ALK): structure, oncogenic activation, and pharmacological inhibition. Pharmacol Res . 2013;68(1):68–94. Downward J. Targeting RAS signalling pathways in cancer therapy. Nat Rev Cancer . 2003;3(1):11–22. Gerlinger M, Rowan AJ, Horswell S, et al. Intratumor heterogeneity and branched evolution revealed by multiregion sequencing. N Engl J Med . 2012;366(10):883–892. Bhullar KS, Lagarón NO, McGowan EM, et al. Kinase-targeted cancer therapies: progress, challenges and future directions. Mol Cancer . 2018;17(1):48. Yun C-H, Mengwasser KE, Toms AV, et al. The T790M mutation in EGFR kinase causes drug resistance by increasing the affinity for ATP. Proc Natl Acad Sci U S A . 2008;105(6):2070–2075. Hyman DM, Puzanov I, Subbiah V, et al. Vemurafenib in multiple nonmelanoma cancers with BRAF V600 mutations. N Engl J Med . 2015;373(8):726–736. Shi, Yulong, et al. "D3EGFR: a webserver for deep learning-guided drug sensitivity prediction and drug response information retrieval for EGFR mutation-driven lung cancer." Briefings in Bioinformatics 2024; 25.3. Additional Declarations No competing interests reported. Supplementary Files suppmaterial.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8730878","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Short Report","associatedPublications":[],"authors":[{"id":586098177,"identity":"dbb431e5-e5e9-416f-be6b-71cf46064b9c","order_by":0,"name":"Harold Mateo Mojica Urrego","email":"","orcid":"","institution":"University of Navarra","correspondingAuthor":false,"prefix":"","firstName":"Harold","middleName":"Mateo Mojica","lastName":"Urrego","suffix":""},{"id":586098179,"identity":"814ce738-cec0-4406-a7a0-2f7ccca04a6b","order_by":1,"name":"Matthew Groves","email":"","orcid":"","institution":"University of Groningen","correspondingAuthor":false,"prefix":"","firstName":"Matthew","middleName":"","lastName":"Groves","suffix":""},{"id":586098180,"identity":"98835b0f-946e-4dd2-828f-53b623d2b3de","order_by":2,"name":"Anthonie Van der Wekken","email":"","orcid":"","institution":"University Medical Center Groningen","correspondingAuthor":false,"prefix":"","firstName":"Anthonie","middleName":"Van der","lastName":"Wekken","suffix":""},{"id":586098181,"identity":"10ea0987-df91-4b3c-9dd0-ef725712f256","order_by":3,"name":"Juvenal Yosa Reyes","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAz0lEQVRIiWNgGAWjYLACCTDJ3PgAwmVuIFYLY7MBhMtIhBaoyjYJorQYHOB9wGDZdljOnP1gWzVvjp1dg3QjIS3sBgySbYeNLXsS227zbktObpA5SEgLGwNQS1rihgNgLQeSGSQSidNSv+H8w7ZiUrTYJBjcSGxjBmqxI6hF8jAbwwGJczaGG248bJacuy05gY2QFr7jbYyPJcok5A3OJx/88HabnT2/RPIBvFoUDjMwHJZAEkhsw6seCOSBjmD8gCRgT0jHKBgFo2AUjDwAAGxbRCQBt9nTAAAAAElFTkSuQmCC","orcid":"","institution":"University Medical Center Groningen","correspondingAuthor":true,"prefix":"","firstName":"Juvenal","middleName":"Yosa","lastName":"Reyes","suffix":""}],"badges":[],"createdAt":"2026-01-29 11:25:11","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8730878/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8730878/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102056647,"identity":"712f71f5-37b6-47b3-ab2a-a2fd0d4bd17d","added_by":"auto","created_at":"2026-02-06 15:59:02","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":286279,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eComparative Performance of Deep Learning Architectures.\u003c/strong\u003e Heatmap of Pearson Correlation Coefficients (PCC) for pIC50 prediction on the EGFR test set across twelve model configurations. The Morgan fingerprint drug encoder paired with a CNN-RNN protein encoder (bottom row, middle column) served as the best-performing benchmark model (PCC = 0.85).\u003c/p\u003e","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8730878/v1/2e93c4db48a8ec665d2cd8fe.jpeg"},{"id":102295818,"identity":"668982c8-cc99-4bc9-89bf-f3758d342ea1","added_by":"auto","created_at":"2026-02-10 10:15:11","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":105841,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTest Performance of the Final Morgan-CNN-RNN Benchmark Model.\u003c/strong\u003e \u003cstrong\u003e(A)\u003c/strong\u003e Scatter plot of predicted vs. actual pIC50 values (MSE = 0.6989,\u003cstrong\u003e \u003c/strong\u003ePCC = 0.82, R² = 0.63). \u003cstrong\u003e(B)\u003c/strong\u003eDistribution of prediction uncertainties. \u003cstrong\u003e(C)\u003c/strong\u003ePlot of absolute error vs. uncertainty.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8730878/v1/bf3a05ad5f52364d51b34d6a.png"},{"id":102295732,"identity":"f879c6a9-481a-48c9-8fee-52e2b861db41","added_by":"auto","created_at":"2026-02-10 10:14:27","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":116323,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTest Performance of the Morgan–ESM2 (3B parameters) model. (A)\u003c/strong\u003e Scatter plot of predicted vs. actual pIC50 values (MSE = 0.8583, PCC = 0.7900, R² = 0.6172). \u003cstrong\u003e(B)\u003c/strong\u003e Distribution of prediction uncertainties. \u003cstrong\u003e(C)\u003c/strong\u003ePlot of absolute error vs. uncertainty.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8730878/v1/31963f63b24a324102977609.png"},{"id":102056649,"identity":"088fa4a9-cb8f-4482-98e7-28d6f703a7bd","added_by":"auto","created_at":"2026-02-06 15:59:02","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":181004,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eESM2 Model-Predicted Bioactivity of Clinical EGFR Inhibitors.\u003c/strong\u003e Estimated pIC50 values for nine EGFR TKIs were predicted using the fine-tuned ESM2 model. The plot shows differential sensitivity patterns that align with known clinical outcomes, correctly identifying resistance mediated by T790M and C797S mutations and the high potency of next-generation inhibitors against these variants, showcasing its utility for complex mutational contexts.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8730878/v1/a793927e3aefbddb40d30faa.png"},{"id":102748801,"identity":"956f1d7a-dd37-4337-a703-5d939a9dbbd7","added_by":"auto","created_at":"2026-02-16 09:11:34","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1417458,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8730878/v1/ddcd615c-73bf-4b94-bc3e-5b16b597f1b5.pdf"},{"id":102056645,"identity":"169de842-1825-4fc2-b26e-d2b283acd8c9","added_by":"auto","created_at":"2026-02-06 15:59:02","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":12230,"visible":true,"origin":"","legend":"","description":"","filename":"suppmaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-8730878/v1/9030ffca98db24eef450a79c.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"PROtein Analytics for Kinase Therapeutic Inhibitor Variants (PROAKTIV): A Machine learning approach to predict efficacy of tyrosine kinase inhibitors against mutations in EGFR, ALK and BRAF","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eNon-small cell lung cancer (NSCLC) accounts for approximately 85% of all lung cancer cases and remains the leading cause of cancer-related mortality worldwide.\u0026sup1; Its profound molecular heterogeneity, characterized by a complex landscape of oncogenic driver mutations, poses a significant challenge to effective treatment.\u0026sup2;,\u0026sup3; The paradigm of personalized medicine, which matches drug therapies to specific molecular drivers, has transformed outcomes for select patient populations.⁴\u003c/p\u003e \u003cp\u003eKey among these drivers are mutations in protein kinases that govern critical pro-survival signaling pathways, including the Epidermal Growth Factor Receptor (EGFR),⁵ Anaplastic Lymphoma Kinase (ALK),⁶ and B-Raf (BRAF).⁷ Activating mutations in EGFR, such as exon 19 deletions or the L858R point mutation, are found in 10\u0026ndash;40% of NSCLC adenocarcinomas and confer sensitivity to first-generation tyrosine kinase inhibitors (TKIs) like gefitinib and erlotinib.⁸ Similarly, chromosomal rearrangements resulting in EML4-ALK fusion proteins define a distinct NSCLC subtype highly responsive to ALK inhibitors such as crizotinib.⁶,⁹ Downstream of these receptors, mutations in the BRAF serine/threonine kinase, particularly the V600E substitution, provide a direct and powerful engine for proliferation that can be targeted with selective inhibitors like dabrafenib and vemurafenib.⁷,\u0026sup1;⁰\u003c/p\u003e \u003cp\u003eDespite the initial success of these targeted therapies, their long-term efficacy is frequently limited by the emergence of on-target resistance mutations.\u0026sup1;\u0026sup1; For example, the EGFR T790M \"gatekeeper\" mutation impacts the binding of first- and second-generation TKIs, leading to treatment failure.\u0026sup1;\u0026sup2; The development of the third-generation inhibitor osimertinib successfully overcame T790M-mediated resistance,\u0026sup1;\u0026sup3; but its efficacy is now challenged by the subsequent acquisition of the C797S mutation at its covalent binding site.\u0026sup1;⁴,\u0026sup1;⁵ A similar evolutionary arms race is observed in ALK-positive NSCLC, where initial resistance to crizotinib via mutations like L1196M prompted the development of next-generation inhibitors, only to be met by further resistance from mutations such as G1202R.\u003csup\u003e12\u003c/sup\u003e,\u003csup\u003e16\u003c/sup\u003e,\u003csup\u003e17\u003c/sup\u003e This dynamic interplay highlights a critical need for predictive tools that can anticipate the effect of both well-characterized and rare kinase mutations on drug binding efficacy.\u003c/p\u003e \u003cp\u003eWhile experimental and structure-based methods can offer deep mechanistic insights,\u003csup\u003e18\u003c/sup\u003e, significant time is required to provide robust \u003cem\u003ein vitro\u003c/em\u003e pIC50 (-log10(IC50 M)) values that reflect the efficacy of a specific TKI on its target. This reliance provides a restriction in both the timely selection of a drug for rare mutations and in the creation of new drug-like molecules. To address this gap we propose a machine learning model (PROAKTIV; PROtein Analytics for Kinase Therapeutic Inhibitor Variants), a sequence-based machine learning approach to predict ligand bioactivity against mutated kinases. We finetuned the 3-billion-parameter ESM-2 protein language model to predict drug\u0026ndash;target interactions in NSCLC-associated proteins (EGFR, ALK, BRAF) \u003csup\u003e23\u003c/sup\u003e.\u003c/p\u003e"},{"header":"MATERIALS AND METHODS","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eData and Code Availability\u003c/h2\u003e \u003cp\u003eThe complete code for the data curation pipeline, model training, and evaluation is publicly available on GitHub at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/HaroldMate1/proaktiv\u003c/span\u003e\u003cspan address=\"https://github.com/HaroldMate1/proaktiv\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eData Acquisition and Preprocessing\u003c/h3\u003e\n\u003cp\u003eAn automated Python workflow was developed to retrieve data for human EGFR (UniProt ID: P00533), ALK (Q9UM72), and BRAF (P15056) from the UniProt database (Release 2025_02)\u003csup\u003e19\u003c/sup\u003e and ChEMBL (Release 33).\u003csup\u003e20\u003c/sup\u003e The workflow programmatically retrieved all available bioactivity data from the scientific literature, filtering for 'IC50' as the standard measurement type. Records with missing values or duplicates were removed, and IC50 values were converted to pIC50.\u003c/p\u003e\n\u003ch3\u003eMutation Processing and Sequence Generation\u003c/h3\u003e\n\u003cp\u003eMutation information was extracted from the 'assay description' and 'document title' fields in the ChEMBL metadata using regular expression algorithms. Extracted mutation strings were parsed and standardized to Human Genome Variation Society (HGVS) nomenclature. A custom function then programmatically applied these mutations, including single nucleotide polymorphisms (SNPs), insertions, and deletions, to the canonical wild-type FASTA sequences, generating a comprehensive library of variant protein sequences.\u003c/p\u003e\n\u003ch3\u003eModel Development and Evaluation\u003c/h3\u003e\n\u003cp\u003eThe selection of deep learning architectures was guided by the benchmarking performed by Shi et al., who evaluated 80 encoder combinations to identify optimal configurations for EGFR-TKI sensitivity prediction\u003csup\u003e20\u003c/sup\u003e. Building upon their identification of high-performing encoder types, we constructed a comparative matrix of twelve models. However, this study diverged from their methodology in two key aspects: first, we utilized a custom-curated dataset rather than the clinical patient case database employed in the reference study; second, we substituted the standard Transformer protein encoder with a fine-tuned ESM2 protein language model to generate context-aware embeddings.\u003c/p\u003e \u003cp\u003eConsequently, the final analysis evaluated pairings of four ligand representation strategies with three protein encoding architectures. Ligands were processed using either fixed representations (Morgan and Daylight fingerprints\u0026sup2;\u0026sup2;) or learnable architectures (Message Passing Neural Networks [MPNN] and Convolutional Neural Networks [CNN] applied to SMILES\u0026sup2;\u0026sup1;). Simultaneously, protein sequences were encoded using three distinct architectures: a standard CNN, a hybrid CNN-RNN, and a Transformer model incorporating the pre-trained ESM2 language model.\u003c/p\u003e \u003cp\u003eThis resulted in a matrix of twelve model combinations. All models were implemented in PyTorch and randomly split into training (70%), test (20%), and validation (10%) sets. Training minimized Mean Squared Error (MSE) using the Adam optimizer\u0026sup2;⁴, and performance was evaluated using MSE, R-squared (R\u0026sup2;), and the Pearson Correlation Coefficient (PCC). Prediction uncertainty was quantified using Monte Carlo (MC) dropout\u0026sup2;⁵ (see \u003cb\u003eSupplementary Note 1\u003c/b\u003e for detailed specifications of each encoder).\u003c/p\u003e"},{"header":"RESULTS","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\n \u003ch2\u003eAutomated Pipeline Curates a Large-Scale Bioactivity Dataset\u003c/h2\u003e\n \u003cp\u003eThe automated pipeline successfully curated 25,412 high-quality bioactivity data points across EGFR (15,588), ALK (3,246), and BRAF (6,578). This dataset encompassed 10,074 unique ligands and dozens of distinct mutations, including clinically critical SNPs (T790M, G1202R, V600E), insertions, deletions (delE746_A750), and complex compound mutations (e.g., EGFR C797S/L858R/T790M). Exploratory data analysis confirmed that the dataset faithfully recapitulated known clinical patterns of sensitivity and resistance.\u003c/p\u003e\n\u003c/div\u003e\n\u003ch3\u003eCNN-RNN Benchmark Model Establishes Strong Predictive Performance\u003c/h3\u003e\n\u003cp\u003eTo establish a robust performance benchmark, we systematically evaluated twelve different deep learning architectures. The combination of Morgan fingerprints for ligand encoding and a CNN-RNN for protein encoding consistently yielded the highest predictive accuracy. As shown in Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e, this benchmark configuration achieved a Pearson Correlation Coefficient of 0.85 on the EGFR test set, outperforming all other encoder combinations.\u003c/p\u003e\n\u003ch3\u003eFine-Tuned ESM2 Language Model Enhances Predictions for Clinical Utility\u003c/h3\u003e\n\u003cp\u003eThe final benchmark model, trained on the combined dataset, demonstrated robust generalization with a PCC of 0.82, an R\u0026sup2; of 0.63, and an MSE of 0.78 on the test set (Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e\n\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\n \u003cp\u003eTo advance beyond this benchmark and create a tool better suited for clinical application on rare variants, we evaluated the fine-tuned 3B-parameter ESM2 model. This state-of-the-art approach achieved comparable overall performance (PCC\u0026thinsp;=\u0026thinsp;0.7900 on the EGFR test set) and, crucially, demonstrated improved predictive capabilities for complex and less-common mutations not well-represented in the training data (Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\n \u003ch2\u003eModel Predictions Align with Clinical Drug Sensitivity Profiles\u003c/h2\u003e\n \u003cp\u003eTo assess the clinical relevance of the PROAKTIV framework, we used the advanced ESM2 model to predict pIC50 values for clinical EGFR inhibitors against a wide array of EGFR mutations. The predictions accurately mirrored established drug sensitivity and resistance profiles (Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e). It correctly predicted that first- and second-generation TKIs would be potent against activating mutations but lose efficacy against T790M mutants. Furthermore, it predicted that osimertinib would overcome T790M-mediated resistance but fail against C797S mutations. Critically, the model predicted that the fourth-generation TKI candidate, BLU-945, would retain high potency against T790M and C797S-containing triple mutants, consistent with its design and preclinical data. This demonstrates the model\u0026apos;s ability to generalize to complex scenarios.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eIn this study, we introduced the PROAKTIV framework, a machine learning approach to predict the bioactivity of inhibitors against mutated kinases. Our primary finding is that deep learning models, trained on systematically curated public data, can predict functional consequences of diverse mutations in key proteins implicated in NSCLC pathogenesis. This work moves beyond academic validation and provides a tangible framework with direct implications for clinical practice. For the clinician facing a patient with a rare or complex mutation profile for which treatment guidelines are non-existent, our approach offers a data-driven method to generate therapeutic hypotheses.\u003c/p\u003e \u003cp\u003eModern machine learning approaches have achieved the accuracy required to complement experimental studies by condensing a vast body of scientific knowledge into actionable analyses for matching drugs to specific mutations. The PROAKTIV framework exemplifies this leap forward, particularly in supporting the complex decisions faced by Molecular Tumor Boards (MTBs). When next-generation sequencing reveals a rare variant or a complex resistance mutation with no clear treatment path, our model can process extensive bioactivity data to generate a ranked list of inhibitors based on predicted in vitro efficacy (IC50). For instance, the model's ability to predict retained efficacy of the fourth generation TKIs against the EGFR C797S triple mutant provides a data-driven, actionable hypothesis where standard options are exhausted.\u003csup\u003e30\u003c/sup\u003e This shifts patient management from empirical trial-and-error towards a more rational, personalized strategy. The predicted pIC50 values function as a relative ranking system, designed to prioritize therapeutic options for a specific genetic context rather than serve as absolute measures of clinical potency, Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eWhile standard architectures, such as our benchmark CNN-RNN, perform robustly on well-documented mutations, they often struggle to extrapolate beyond their training distribution. Literature has consistently shown that models relying solely on supervised learning with limited datasets tend to overfit to specific sequence identities, resulting in poor predictive power for 'out-of-distribution' variants. This limitation is addressed by the fine-tuned ESM2 model, which leverages embeddings derived from the evolutionary history of millions of protein sequences rather than simple sequence similarity. This capability is essential for characterizing rare mutations that lack experimental precedence. From a clinical perspective, the associated uncertainty quantification transforms these predictions into actionable insights: a high uncertainty score does not imply model failure, but rather identifies a unique biological context, signaling the need for confirmatory testing before clinical consideration.\u003c/p\u003e \u003cp\u003eFrom a clinical standpoint, it is essential to frame these predictions within their limitations. The model\u0026rsquo;s analysis is based on \u003cem\u003ein vitro\u003c/em\u003e bioactivity and does not account for the complexities of \u003cem\u003ein vivo\u003c/em\u003e pharmacology, such as drug bioavailability, metabolism, or the influence of the tumor microenvironment. As such, the PROAKTIV framework is designed as a decision-support tool that complements, rather than replaces, clinical expertise and experimental validation. Its purpose is to initiate and inform the clinical conversation with data-driven hypotheses, not to provide definitive therapeutic answers.\u003c/p\u003e \u003cp\u003eThe path to clinical implementation requires a clear validation strategy. The immediate next step is a prospective study comparing PROAKTIV\u0026rsquo;s predictions in parallel with the recommendations of an active MTB. Establishing concordance with expert consensus and ultimately correlating predictions with patient outcomes are essential for building clinical trust. Future iterations will be enhanced by integrating 3D structural information, providing a more mechanistic understanding of how mutations alter drug binding.\u0026sup3;\u0026sup1;\u003c/p\u003e \u003cp\u003eIn conclusion, PROAKTIV provides a robust computational tool that condenses the vast landscape of genomic and pharmacological data into actionable insights. The advanced ESM2 model, in particular, holds considerable promise for realizing the vision of precision oncology, empowering clinicians to make more informed decisions for patients with both common and rare tumor mutations.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eConflict of Interest Statement:\u003c/h2\u003e \u003cp\u003eThe authors declare no potential conflicts of interest.\u003c/p\u003e\u003ch2\u003eCredit Author Statement\u003c/h2\u003e\n\u003cp\u003eHarold Mateo Mojica Urrego: Conceptualization, Resources, Data Curation, Formal Analysis, Investigation, Visualization, Methodology, Writing \u0026ndash; Original Draft.\u003c/p\u003e\n\u003cp\u003eMatthew Groves: Supervision, Formal Analysis, Writing \u0026ndash; Review \u0026amp; Editing.\u003c/p\u003e\n\u003cp\u003eAnthonie van der Wekken: Supervision, Formal Analysis, Writing \u0026ndash; Review \u0026amp; Editing.\u003c/p\u003e\n\u003cp\u003eJuvenal Yosa Reyes: Conceptualization, Investigation, Methodology, Formal Analysis, Writing \u0026ndash; Review \u0026amp; Editing.\u003c/p\u003e\u003ch2\u003eFunding:\u003c/h2\u003e \u003cp\u003eThis research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eHarold Mateo Mojica Urrego: Conceptualization, Resources, Data Curation, Formal Analysis, Investigation, Visualization, Methodology, Writing \u0026ndash; Original Draft.Matthew Groves: Supervision, Formal Analysis, Writing \u0026ndash; Review \u0026amp; Editing.Anthonie van der Wekken: Supervision, Formal Analysis, Writing \u0026ndash; Review \u0026amp; Editing.Juvenal Yosa Reyes: Conceptualization, Investigation, Methodology, Formal Analysis, Writing \u0026ndash; Review \u0026amp; Editing.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe complete code for the data curation pipeline, model training, and evaluation is publicly available on GitHub at https://github.com/HaroldMate1/proaktiv.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eBray F, Laversanne M, Weiderpass E, Soerjomataram I. The ever-increasing importance of cancer as a leading cause of premature death worldwide. \u003cem\u003eCancer\u003c/em\u003e. 2021;127(16):3029\u0026ndash;3030.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePikor LA, Ramnarine VR, Lam S, Lam WL. Genetic alterations defining NSCLC subtypes and their therapeutic implications. \u003cem\u003eLung Cancer\u003c/em\u003e. 2013;82(2):179\u0026ndash;189.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen Z, Fillmore CM, Hammerman PS, Kim CF, Wong K-K. Non-small-cell lung cancers: a heterogeneous set of diseases. \u003cem\u003eNat Rev Cancer\u003c/em\u003e. 2014;14(8):535\u0026ndash;546.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMok TS. Personalized medicine in lung cancer: what we need to know. \u003cem\u003eNat Rev Clin Oncol\u003c/em\u003e. 2011;8(11):661\u0026ndash;668.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSharma SV, Bell DW, Settleman J, Haber DA. Epidermal growth factor receptor mutations in lung cancer. \u003cem\u003eNat Rev Cancer\u003c/em\u003e. 2007;7(3):169\u0026ndash;181.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSoda M, Choi YL, Enomoto M, et al. Identification of the transforming EML4-ALK fusion gene in non-small-cell lung cancer. \u003cem\u003eNature\u003c/em\u003e. 2007;448(7153):561\u0026ndash;566.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDavies H, Bignell GR, Cox C, et al. Mutations of the BRAF gene in human cancer. \u003cem\u003eNature\u003c/em\u003e. 2002;417(6892):949\u0026ndash;954.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePaez JG, J\u0026auml;nne PA, Lee JC, et al. EGFR mutations in lung cancer: correlation with clinical response to gefitinib therapy. \u003cem\u003eScience\u003c/em\u003e. 2004;304(5676):1497\u0026ndash;1500.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShaw AT, Yeap BY, Mino-Kenudson M, et al. Clinical features and outcome of patients with non-small-cell lung cancer who harbor EML4-ALK. \u003cem\u003eJ Clin Oncol\u003c/em\u003e. 2009;27(26):4247\u0026ndash;4253.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePlanchard D, Smit EF, Groen HJ, et al. Dabrafenib plus trametinib in patients with previously untreated BRAFV600E-mutant metastatic non-small-cell lung cancer: an open-label, phase 2 trial. \u003cem\u003eLancet Oncol\u003c/em\u003e. 2017;18(10):1307\u0026ndash;1316.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRotow J, Bivona TG. Understanding and targeting resistance mechanisms in NSCLC. \u003cem\u003eNat Rev Cancer\u003c/em\u003e. 2017;17(11):637\u0026ndash;658.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePao W, Miller VA, Politi KA, et al. Acquired resistance of lung adenocarcinomas to gefitinib or erlotinib is associated with a second mutation in the EGFR kinase domain. \u003cem\u003ePLoS Med\u003c/em\u003e. 2005;2(3):e73.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJ\u0026auml;nne PA, Yang JC-H, Kim D-W, et al. AZD9291 in EGFR inhibitor-resistant non-small-cell lung cancer. \u003cem\u003eN Engl J Med\u003c/em\u003e. 2015;372(18):1689\u0026ndash;1699.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThress KS, Paweletz CP, Felip E, et al. Acquired EGFR C797S mutation mediates resistance to AZD9291 in non-small cell lung cancer harboring EGFR T790M. \u003cem\u003eNat Med\u003c/em\u003e. 2015;21(6):560\u0026ndash;562.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJia Y, Yun C-H, Park E, et al. Overcoming EGFR(T790M) and EGFR(C797S) resistance with mutant-selective allosteric inhibitors. \u003cem\u003eNature\u003c/em\u003e. 2016;534(7605):129\u0026ndash;132.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKatayama R, Shaw AT, Khan TM, et al. Mechanisms of acquired crizotinib resistance in ALK-rearranged lung cancers. \u003cem\u003eSci Transl Med\u003c/em\u003e. 2012;4(120):120ra17.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGainor JF, Dardaei L, Yoda S, et al. Molecular mechanisms of resistance to first-and second-generation ALK inhibitors in ALK-rearranged lung cancer. \u003cem\u003eCancer Discov\u003c/em\u003e. 2016;6(10):1118\u0026ndash;1133.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRobichaux JP, Le X, Vijayan R, et al. Structure-based classification predicts drug response in EGFR-mutant NSCLC. \u003cem\u003eNature\u003c/em\u003e. 2021;597(7878):732\u0026ndash;737.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eUniProt Consortium. UniProt: the universal protein knowledgebase in 2023. \u003cem\u003eNucleic Acids Res\u003c/em\u003e. 2023;51(D1):D523-D531.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGaulton A, Hersey A, Nowotka M, et al. The ChEMBL database in 2017. \u003cem\u003eNucleic Acids Res\u003c/em\u003e. 2017;45(D1):D945-D954.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWeininger D. SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules. \u003cem\u003eJ Chem Inf Comput Sci\u003c/em\u003e. 1988;28(1):31\u0026ndash;36.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRogers D, Hahn M. Extended-connectivity fingerprints. \u003cem\u003eJ Chem Inf Model\u003c/em\u003e. 2010;50(5):742\u0026ndash;754.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLin Z, Akin H, Rao R, et al. Language models of protein sequences at the scale of evolution. \u003cem\u003ebioRxiv\u003c/em\u003e. 2022. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1101/2022.07.20.500902\u003c/span\u003e\u003cspan address=\"10.1101/2022.07.20.500902\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKingma DP, Ba J. Adam: A method for stochastic optimization. \u003cem\u003earXiv\u003c/em\u003e. 2014. arXiv:1412.6980.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSrivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: a simple way to prevent neural networks from overfitting. \u003cem\u003eJ Mach Learn Res\u003c/em\u003e. 2014;15:1929\u0026ndash;1958.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang K, Fu T, Glass LM, Zitnik M, Xiao C, Sun J. DeepPurpose: a deep learning library for drug-target interaction prediction. \u003cem\u003eBioinformatics\u003c/em\u003e. 2020;36(22\u0026ndash;23):5545\u0026ndash;5547.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeonetti A, Sharma S, Minari R, Perego P, Giovannetti E, Tiseo M. Resistance mechanisms to osimertinib in EGFR-mutated non-small cell lung cancer. \u003cem\u003eBr J Cancer\u003c/em\u003e. 2019;121(9):725\u0026ndash;737.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAddeo A, Tabb\u0026ograve; F, Robinson T, Buffoni L, Novello S. Precision medicine in ALK rearranged NSCLC: A rapidly evolving scenario. \u003cem\u003eCrit Rev Oncol Hematol\u003c/em\u003e. 2018;122:150\u0026ndash;156.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeonetti A, Facchinetti F, Rossi G, et al. BRAF in non-small cell lung cancer (NSCLC): pickaxing another brick in the wall. \u003cem\u003eCancer Treat Rev\u003c/em\u003e. 2018;66:82\u0026ndash;94.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKoopman B, van der Wekken AJ, Ter Elst A, et al. Relevance and effectiveness of molecular tumor board recommendations for patients with non-small-cell lung cancer with rare or complex mutational profiles. \u003cem\u003eJCO Precis Oncol\u003c/em\u003e. 2020;4:393\u0026ndash;410.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGaldadas I, Carlino L, Ward RA, Hughes SJ, Haider S, Gervasio FL. Structural basis of the effect of activating mutations on the EGF receptor. \u003cem\u003eElife\u003c/em\u003e. 2021;10:e65824.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMeier J, Rao R, Verkuil R, et al. Language models enable zero-shot prediction of the effects of mutations on protein function. \u003cem\u003eAdv Neural Inf Process Syst\u003c/em\u003e. 2021;34:29287\u0026ndash;29303.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKris MG, Johnson BE, Berry LD, et al. Using multiplexed assays of oncogenic drivers in lung cancers to select targeted drugs. \u003cem\u003eJAMA\u003c/em\u003e. 2014;311(19):1998\u0026ndash;2006.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWong CH, Siah KW, Lo AW. Estimation of clinical trial success rates and related parameters. \u003cem\u003eBiostatistics\u003c/em\u003e. 2019;20(2):273\u0026ndash;286.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChiarle R, Voena C, Ambrogio C, Piva R, Inghirami G. The anaplastic lymphoma kinase in the pathogenesis of cancer. \u003cem\u003eNat Rev Cancer\u003c/em\u003e. 2008;8(1):11\u0026ndash;23.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHallberg B, Palmer RH. Mechanistic insight into ALK receptor tyrosine kinase in human cancer biology. \u003cem\u003eNat Rev Cancer\u003c/em\u003e. 2013;13(10):685\u0026ndash;700.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSolomon BJ, Liu G, Felip E, et al. Lorlatinib vs crizotinib in treatment-na\u0026iuml;ve patients with advanced ALK\u0026thinsp;+\u0026thinsp;non-small cell lung cancer: 5-year progression-free survival and safety from the CROWN study. \u003cem\u003eJ Clin Oncol\u003c/em\u003e. 2024;42(suppl 17):LBA8507.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWan PT, Garnett MJ, Roe SM, et al. Mechanism of activation of the RAF-ERK signaling pathway by oncogenic mutations of B-RAF. \u003cem\u003eCell\u003c/em\u003e. 2004;116(6):855\u0026ndash;867.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCross DA, Ashton SE, Ghiorghiu S, et al. AZD9291, an irreversible EGFR TKI, overcomes T790M-mediated resistance to EGFR inhibitors in lung cancer. \u003cem\u003eCancer Discov\u003c/em\u003e. 2014;4(9):1046\u0026ndash;1061.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFriboulet L, Li N, Katayama R, et al. The ALK inhibitor ceritinib overcomes crizotinib resistance in non-small cell lung cancer. \u003cem\u003eCancer Discov\u003c/em\u003e. 2014;4(6):662\u0026ndash;673.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCamidge DR, Kim HR, Ahn M-J, et al. Brigatinib versus crizotinib in ALK-positive non-small-cell lung cancer. \u003cem\u003eN Engl J Med\u003c/em\u003e. 2018;379(21):2027\u0026ndash;2039.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHyman DM, Puzanov I, Subbiah V, et al. Vemurafenib in multiple nonmelanoma cancers with BRAF V600 mutations. \u003cem\u003eN Engl J Med\u003c/em\u003e. 2015;373(8):726\u0026ndash;736.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSoria JC, Ohe Y, Vansteenkiste J, et al. Osimertinib in untreated EGFR-mutated advanced non-small-cell lung cancer. \u003cem\u003eN Engl J Med\u003c/em\u003e. 2018;378(2):113\u0026ndash;125.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCasaluce F, Sgambato A, Maione P, et al. ALK inhibitors: a new targeted therapy in the treatment of advanced NSCLC. \u003cem\u003eTarget Oncol\u003c/em\u003e. 2013;8(1):55\u0026ndash;67.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRoskoski Jr R. Anaplastic lymphoma kinase (ALK): structure, oncogenic activation, and pharmacological inhibition. \u003cem\u003ePharmacol Res\u003c/em\u003e. 2013;68(1):68\u0026ndash;94.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDownward J. Targeting RAS signalling pathways in cancer therapy. \u003cem\u003eNat Rev Cancer\u003c/em\u003e. 2003;3(1):11\u0026ndash;22.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGerlinger M, Rowan AJ, Horswell S, et al. Intratumor heterogeneity and branched evolution revealed by multiregion sequencing. \u003cem\u003eN Engl J Med\u003c/em\u003e. 2012;366(10):883\u0026ndash;892.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBhullar KS, Lagar\u0026oacute;n NO, McGowan EM, et al. Kinase-targeted cancer therapies: progress, challenges and future directions. \u003cem\u003eMol Cancer\u003c/em\u003e. 2018;17(1):48.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYun C-H, Mengwasser KE, Toms AV, et al. The T790M mutation in EGFR kinase causes drug resistance by increasing the affinity for ATP. \u003cem\u003eProc Natl Acad Sci U S A\u003c/em\u003e. 2008;105(6):2070\u0026ndash;2075.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHyman DM, Puzanov I, Subbiah V, et al. Vemurafenib in multiple nonmelanoma cancers with BRAF V600 mutations. \u003cem\u003eN Engl J Med\u003c/em\u003e. 2015;373(8):726\u0026ndash;736.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShi, Yulong, et al. \"D3EGFR: a webserver for deep learning-guided drug sensitivity prediction and drug response information retrieval for EGFR mutation-driven lung cancer.\" \u003cem\u003eBriefings in Bioinformatics\u003c/em\u003e 2024; 25.3.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Non-Small Cell Lung Cancer, Targeted Therapy, Machine Learning, EGFR, ALK, BRAF, Protein Language Model","lastPublishedDoi":"10.21203/rs.3.rs-8730878/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8730878/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eIntroduction:\u003c/h2\u003e \u003cp\u003eThe efficacy of targeted therapies in non-small cell lung cancer (NSCLC) is challenged by acquired resistance, driven by on-target mutations in kinases such as EGFR, ALK, and BRAF. Thus, predicting the functional impact of these mutations on tyrosine kinase inhibitor (TKI) sensitivity is critical for personalized treatment. This study presents a foundational machine learning framework for predicting ligand bioactivity against mutated kinases from sequence data, using EGFR, ALK, and BRAF as proof-of-concept models. The results demonstrate the feasibility of developing a generalized predictive model applicable across the kinase family.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eAn automated Python pipeline was developed to curate mutation and bioactivity data from a comprehensive dataset of 25,412 published \u003cem\u003ein vitro\u003c/em\u003e pIC50 values for EGFR, ALK, and BRAF variants. Twelve deep learning architectures were trained and evaluated with different encoders for proteins and ligands. Model performance was assessed using Mean Squared Error (MSE), R-squared (R\u0026sup2;), and Pearson Correlation Coefficient (PCC), with uncertainty quantified via Monte Carlo dropout.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThe best-performing model demonstrated robust predictive accuracy, providing a pearson correlation of 85% on the mutation/TKI pairs. Model predictions for clinically relevant drug-mutation pairs consistently aligned with established clinical outcomes, including EGFR T790M-mediated resistance to first-generation inhibitors, ALK G1202R resistance to crizotinib, and BRAF V600E sensitivity to selective inhibitors. A protein sequence language model (ESM2) offered improved predictions for complex, rare variants.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eThis study introduces a foundational machine learning framework for predicting the impact of kinase mutations on \u003cem\u003ein vitro\u003c/em\u003e drug sensitivity. This methodology supports personalized therapy development in NSCLC and may enhance the efficiency of drug discovery pipelines.\u003c/p\u003e","manuscriptTitle":"PROtein Analytics for Kinase Therapeutic Inhibitor Variants (PROAKTIV): A Machine learning approach to predict efficacy of tyrosine kinase inhibitors against mutations in EGFR, ALK and BRAF","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-06 15:58:57","doi":"10.21203/rs.3.rs-8730878/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"32b607d6-cdc6-4e63-b99d-d0397f7e4b79","owner":[],"postedDate":"February 6th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":62357107,"name":"Biological sciences/Cancer"},{"id":62357108,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":62357109,"name":"Biological sciences/Drug discovery"}],"tags":[],"updatedAt":"2026-02-14T22:08:49+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-06 15:58:57","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8730878","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8730878","identity":"rs-8730878","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}