Predicting drug targets: How I came to learn and love machine learning + Making data ready for an AI-driven world

Predicting drug targets: How I came to learn and love machine learning Making data ready for an AI-driven world Illimar Rekand, Ph.D. (He/Him) Department of Informatics+Chemistry, University of Bergen 2025-10-15 Læringsdagen for Teknikere, UiB 1 This presentation is available on Zenodo under a CC license: https://doi.org/10.5281/zenodo.17349124 Background ● M.Sc. in organic chemistry (UiB) ● Ph.D. in Biomedicine + Chemistry (UiB) ● Librarian (Open Science, Data management) ● Elixir Norway ● Last 3 years: Software Carpentry Course ● Currently: Senior Engineer, Chemistry+Inf iCat4Bio project 2 iCat4Bio “Closed loop” Automation of chemical experiments Automated HPLC analysis Direct Injection Generate data Machine learning / “AI” method Propose new experiment Images: ● HPLC-diagram -Lukke - Own work, CC BY-SA 3.0, Wikimedia ● Robot: Unchained Labs ● Agilent HPLC 3 4 Illimar Rekand. (2024). kjemist/TotalDeaths20thCentury: Minimum Working Example (v0.1). Zenodo. https://doi.org/10.5281/zenodo.12620698 Today’s situation [US, Annual Numbers] CDC. Antibiotic Resistance Threats in the United States, 2019 http://dx.doi.org/10.15620/cdc:82532 5 Background - Antimicrobial resistance (AMR) ● By 2050; more deaths caused by AMR then cancer (amr-review.org) ● Antibiotics with new mechanisms ● Most antibiotic classes were discovered a long time ago ● New antibiotics are not interesting for pharma companies 6 “Pills”, the Javorac, Flickr, CC 2.0 Paracetamol 2D representation 3D representation 7 These molecules need to go somewhere: “Drug target” Mostly proteins 8 “Lock” or Target/ Binding pocket “Key” Drug molecule “Lock/Keycomplex” 9 Volume? ● Cubic Ångström ● Surface area of whole pocket ● How many atoms fit inside ● +++ Form & shape? ● How big is the 1st, 2nd principal moment of inertia? ● What is the spherocity index? ● +++ Chemical properties? ● What is the area of polar atoms? ● What is the ratio between polar/nonpolar atoms? ● How many oxygens/nitrogens are in the pocket? ● +++ Accessibility? ● What is the ratio between open/closed surface area? ● How many atoms are exposed? ● How much of each atom is exposed? ● +++ Binding pocket attributes: 16 The “druggability” spectrum Poor targets: Generally poor attributes Good targets: Generally good attributes ??? Cutoff-line between good and bad targets 17 “Druggability” is an abstract concept 18 ✅ 19 By Nosferattus - Own work, CC0, https://commons.wikimedia.org/w/index.php?curid=102323072 Why use machine learning? Pocket: Druggable? Property 1 Property 2 Property 3 Property n A Yes ✅ ✅ ❌ … B Yes ❌ ✅ ✅ … CNo ❌ ✅ ❌ … Decision tree approach - “xgboost” Prediction model ● Many data points; complexity increases ● Druggability classification is not trivial ● Difficult for human beings to keep track of data, while a machine can do this easier 20 How to create a druggability predictor? Identify previously known good + bad drug targets Calculate pocket properties Split into training/testing set Feed properties to machine learning method Evaluate model precision & accuracy Apply model to unknown drug targets 21 ● Garbage in → Garbage out ● Curating a good dataset is key ● Negative instances are as important as positive ● Identifying negatives more demanding High quality models are born from high quality data 22 All possible data Quality data: Less noise / redundancy Positive instances / Good drug targets Negative instances / Bad drug targets How to evaluate a prediction model? ● Accuracy & Precision; misclassification of known drug targets ●Training set: properties used for training the models ●Testing set: apply model on “unseen” instances ● Arrows shows so-called misclassifications Training Testing 23 ML models are created iteratively ● Models “learn” from themselves iteratively ● First iteration: higher error rate ● Next iterations: progressively better ● Avoid “overtraining” 24 Peeking into the black box: SHAP-values ● “Black box”: Data in ➡”Magic happens”➡Data out ● Shapley Additive Explanations (SHAP) give insight into what data contributes to magic ● A: Properties with overall largest contribution to model, ranked highest to lowest ● B: Cutoff-values for top 6 properties 25 Documentation ● “Documentation is a loveletter you write to yourself” ● Electronic > physical ● All documentation and metadata (“data about data”) that is needed to understand & reproduce your data: - Methodology - Analytical & procedural information - Variable/value/unit definitions - Data quality measures - Confidentiality, conditions for use - File structure A ReadMe file can be used for documentation purposes 32 Ostrop, J. (2025, April 3). Hands-On Data Management Planning - University of Bergen Library webinar 04/04/2025. Zenodo. CC-BY 4.0 https://doi.org/10.5281/zenodo.15130737 Meaningful file names ● Meaningful file names help both the author and the recipient ● Length 20-30 characters ● Avoid special character and whitespace ● Date format YYYY-MM-DD ● DO ✅: 2025-02-07_thesis_intro_v3.tex ● DON’T ❌: thesis FINAL 2.docx XKCD # 459 (CC 2.5) 33 Conclusion ● Machine learning can be a useful predictive tool when used with good data ● AI models do not have to be a “black box” ● Good data is created when the data is generated through good data management ● Publication: DrugPred_RNA—A Tool for Structure-Based Druggability Predictions for RNA Binding Sites Illimar Hugo Rekand and Ruth Brenk Journal of Chemical Information and Modeling 2021 DOI: https://doi.org/10.1021/acs.jcim.1c00155 34 Thank you for your attention Brenk Lab Bsky: illimar.bsky.social OrcID: 0000-0002-0457-8024 35 This presentation is available on Zenodo under a CC license: https://doi.org/10.5281/zenodo.17349124