MSc Defence: Predicting Amine pKa Values Using Machine Learning Models (Kamal Aslam)

CANDIDATE: Kamal Aslam

ABSTRACT:  Accurate prediction of amine pK_a is central to carbon-capture solvent design, where the protonated-amine pK_a governs the trade-off between fast CO₂ absorption and low-energy solvent regeneration. This thesis develops machine-learning pK_a predictors in a three-part progression: standard machine-learning baselines, a transfer-learning model built on universal-potential embeddings, and a physics-anchored prediction head refined by evolutionary architecture search.  The first benchmarks five standard architectures (CNN, MLP, Random Forest, SVR, XGBoost) across five molecular representations on curated ChEMBL and IUPAC amine datasets: twenty-one trainable combinations, SVR being tractable only as a single-representation spot-check (the ORCA sigma profile) at this dataset size; the best, XGBoost on Benson-group features, reproduces the ChemAxon-computed CX Basic pK_a labels used as its training target with a Test R² of 0.92 and an MAE of 0.60. The second introduces a new input feature: the penultimate-layer embeddings of the FairChem uma-s-1p2 universal machine-learning interatomic potential, consumed by a three-stage transfer-learning network trained on a purpose-built, quality-controlled pKahub corpus. Under a strict leak-free protocol, this model attains an MAE of 0.27 pK_a units on an 86-amine carbon-capture (CCUS) test set (the best standard model scores 0.58 on the same molecules, though that baseline is out-of-domain on CCUS and was trained on computed rather than experimental labels) and supports temperature-dependent and multiprotic (pK_aH2/pK_aH3) prediction. The same protocol delineates the model’s domain of validity: it generalises within the carbon-capture amine domain but not to arbitrary drug-like amines (Novartis MAE 1.54, R² ≈ 0.1), and is therefore a domain predictor rather than a universal one. The progression culminates in a unified physics-informed head (coupling a linear free-energy anchor on the gas-phase protonation energy, a van’t Hoff temperature law, and a hard multiprotic ordering, with its neural components refined by a large-language-model evolutionary architecture search), which attains an MAE of 0.25 on the same CCUS set and 0.88 on out-of-domain drug-like amines, the best out-of-domain result of the promoted production models developed in this work.