Investigating graph neural networks and classical feature-extraction techniques in activity-cliff and molecular property prediction
Markus Dablander · arXiv (Cornell University) · 2023
Molecular featurisation refers to the process of transforming molecular data into numerical feature vectors. It is one of the key research areas in molecular machine learning and computational drug discovery. Recently, message-passing graph neural networks (GNNs) have emerged as a novel method to learn differentiable features directly from molecular graphs. While such graph-based techniques hold great theoretical promise, further investigations are needed to clarify if and when they indeed manage to definitively outcompete classical molecular featurisations such as extended-connectivity fingerprints (ECFPs) and physicochemical-descriptor vectors (PDVs). In this thesis, we systematically explore and further develop classical as well as graph-based molecular featurisation methods for two important tasks: the well-studied problem of molecular property prediction, in particular quantitative structure-activity relationship (QSAR) prediction, and the largely unexplored challenge of activity-cliff (AC) prediction. We first give a mathematical description and critical analysis of PDVs, ECFPs and message-passing GNNs, with a focus on graph isomorphism networks (GINs). We then conduct a rigorous computational study to compare the performance of PDVs, ECFPs and GINs for QSAR and AC-prediction. Following this, we mathematically describe a novel twin neural network model for AC-prediction and experimentally evaluate ECFP-based and GIN-based versions of this dual architecture. In an additional project, we introduce substructure pooling as a general mathematical operation for the vectorisation of structural fingerprints that represents a natural counterpart to graph-pooling in GNN-architectures. We propose Sort & Slice as a simple substructure-pooling technique for ECFPs that robustly outperforms hashing at molecular property prediction. Finally, we outline two ideas for future research: (i) a graph-based self-supervised learning strategy to make classical molecular featurisations trainable, and (ii) trainable substructure-pooling via differentiable self-attention.