Using Attribution to Decode Dataset Bias in Neural Network Models for Chemistry

Abstract

Deep neural networks have achieved state of the art accuracy at classifying molecules with respect to whether they bind to specific protein targets. A key breakthrough would occur if these models could reveal the fragment pharmacophores that are causally involved in binding. Extracting chemical details of binding from the networks could potentially lead to scientific discoveries about the mechanisms of drug actions. But doing so requires shining light into the black box that is the trained neural network model, a task that has proved difficult across many domains. Here we show how the binding mechanism learned by deep neural network models can be interrogated, using a recently described attribution method. We first work with carefully constructed synthetic datasets, in which the 'fragment logic' of binding is fully known. We find that networks that achieve perfect accuracy on held out test datasets still learn spurious correlations due to biases in the datasets, and we are able to exploit this non-robustness to construct adversarial examples that fool the model. The dataset bias makes these models unreliable for accurately revealing information about the mechanisms of protein-ligand binding. In light of our findings, we prescribe a test that checks for dataset bias given a hypothesis. If the test fails, it indicates that either the model must be simplified or regularized and/or that the training dataset requires augmentation.

References in corpus (12)

• Scikit-learn: Machine Learning in Python
• Neural Message Passing for Quantum Chemistry
• Axiomatic Attribution for Deep Networks
• Molecular Graph Convolutions: Moving Beyond Fingerprints
• Quantum-Chemical Insights from Deep Tensor Neural Networks
• Adversarial Attacks on Neural Networks for Graph Data
• Generating Natural Adversarial Examples
• Optimization of Molecules via Deep Reinforcement Learning
• Massively Multitask Networks for Drug Discovery
• Multi-task Neural Networks for QSAR Predictions
• Chemception: A Deep Neural Network with Minimal Chemistry Knowledge Matches the Performance of Expert-developed QSAR/QSPR Models
• Deep Neural Network Computes Electron Densities and Energies of a Large Set of Organic Molecules Faster than Density Functional Theory (DFT)

Cited by in corpus (15)

• Bayesian semi-supervised learning for uncertainty-calibrated prediction of molecular properties and active learning
• Impossibility Theorems for Feature Attribution
• Multi-Objective Molecule Generation using Interpretable Substructures
• Analyzing Atomic Interactions in Molecules as Learned by Neural Networks
• Towards Symbolic XAI -- Explanation Through Human Understandable Logical Relationships Between Features
• MEGAN: Multi-Explanation Graph Attention Network
• Machine-Learning Driven Drug Repurposing for COVID-19
• Explainable Deep Relational Networks for Predicting Compound-Protein Affinities and Contacts
• Physical machine learning outperforms "human learning" in Quantum Chemistry
• Node-Level Differentially Private Graph Neural Networks
• Releasing Graph Neural Networks with Differential Privacy Guarantees
• Quantitative Evaluation of Explainable Graph Neural Networks for Molecular Property Prediction
• Attribution Methods Reveal Flaws in Fingerprint-Based Virtual Screening
• Marginal Contribution Feature Importance -- an Axiomatic Approach for The Natural Case
• Counterfactual Explanation with Multi-Agent Reinforcement Learning for Drug Target Prediction