Abstract:Molecular featurisation refers to the transformation of 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 techniques hold great 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). We systematically explore and further develop classical and graph-based molecular featurisation methods for two important tasks: 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 technical 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 and computationally evaluate a novel twin neural network model for AC-prediction. We further introduce an operation called substructure pooling for the vectorisation of structural fingerprints as a natural counterpart to graph pooling in GNN architectures. We go on to propose Sort & Slice, a simple substructure-pooling technique for ECFPs that robustly outperforms hash-based folding 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.

Evaluating the generalizability of graph neural networks for predicting collision cross section

Evaluating the generalizability of graph neural networks for predicting collision cross section

Predicting Collision Cross-Section Values for Small Molecules through Chemical Class-Based Multimodal Graph Attention Network

Deep learning the collisional cross sections of the peptide universe from a million experimental values

Using Graph Neural Networks for Mass Spectrometry Prediction

Predicting the Predicted: A Comparison of Machine Learning-Based Collision Cross-Section Prediction Models for Small Molecules

A data-driven machine learning approach for electron-molecule ionization cross sections

Benchmarking Accuracy and Generalizability of Four Graph Neural Networks Using Large In Vitro ADME Datasets from Different Chemical Spaces

Examining Generalizability of AI Models for Catalysis

Could graph neural networks learn better molecular representation for drug discovery? A comparison study of descriptor-based and graph-based models

General Graph Neural Network-Based Model To Accurately Predict Cocrystal Density and Insight from Data Quality and Feature Representation

High-Throughput Measurement and Machine Learning-Based Prediction of Collision Cross Sections for Drugs and Drug Metabolites

Interpreting Graph Neural Networks with Myerson Values for Cheminformatics Approaches

Graph Neural Networks-based Hybrid Framework For Predicting Particle Crushing Strength

QC-GN2oMS2: a Graph Neural Net for High Resolution Mass Spectra Prediction

Generating Molecular Fragmentation Graphs with Autoregressive Neural Networks

Beyond Independent Error Assumptions in Large GNN Atomistic Models

On the Scalability of GNNs for Molecular Graphs

Investigating Graph Neural Networks and Classical Feature-Extraction Techniques in Activity-Cliff and Molecular Property Prediction

Graph neural networks for materials science and chemistry