[Paper Review] Explaining Chemical Toxicity using Missing Features

Kar Wai Lim, Bhanushee Sharma+3 more|arXiv (Cornell University)|Sep 23, 2020

Computational Drug Discovery Methods25 references4 citations

TL;DR

This paper proposes using the Contrastive Explanations Method (CEM) to explain chemical toxicity predictions by identifying both toxicophores (pertinent positives) and missing structural features (pertinent negatives) that influence model predictions. By optimizing minimal perturbations via FISTA, CEM reveals substructures whose presence or absence flips toxicity labels, producing explanations consistent with known toxicophores and improving interpretability in deep learning-based toxicity models.

ABSTRACT

Chemical toxicity prediction using machine learning is important in drug development to reduce repeated animal and human testing, thus saving cost and time. It is highly recommended that the predictions of computational toxicology models are mechanistically explainable. Current state of the art machine learning classifiers are based on deep neural networks, which tend to be complex and harder to interpret. In this paper, we apply a recently developed method named contrastive explanations method (CEM) to explain why a chemical or molecule is predicted to be toxic or not. In contrast to popular methods that provide explanations based on what features are present in the molecule, the CEM provides additional explanation on what features are missing from the molecule that is crucial for the prediction, known as the pertinent negative. The CEM does this by optimizing for the minimum perturbation to the model using a projected fast iterative shrinkage-thresholding algorithm (FISTA). We verified that the explanation from CEM matches known toxicophores and findings from other work.

Motivation & Objective

• To improve interpretability of deep learning models in computational toxicology by identifying not only features that cause toxicity but also those whose absence prevents toxicity.
• To address the limitation of existing explainable AI methods that focus only on present features, by introducing pertinent negatives—missing substructures crucial for prediction.
• To validate that CEM-generated explanations align with known toxicophores and experimental findings in chemical toxicity prediction.
• To demonstrate that CEM can guide molecular design by showing how to modify toxic molecules into non-toxic ones through minimal structural changes.
• To extend the scope of toxicity explanation beyond structural alerts to include actionable insights on molecular transformation.

Proposed method

• CEM is applied to a deep neural network (DNN) model trained on molecular structures for predicting toxicity in the SR-MMP (mitochondrial membrane potential disruption) task.
• The method identifies Pertinent Positives (PPs) as the minimal set of features whose presence leads to a toxic prediction.
• The method identifies Pertinent Negatives (PNs) as the minimal set of features whose absence prevents a toxic prediction, representing missing toxicophores.
• Optimization is performed using the projected fast iterative shrinkage-thresholding algorithm (FISTA) to find the smallest perturbation to the input molecule that flips the model’s prediction.
• Feature importance is derived by analyzing the minimal changes required to alter the model’s output, focusing on structural fragments like functional groups and substructures.
• Explanations are validated against known toxicophores from literature, including phenolic OH, aliphatic halides, polycyclic aromatic systems, and aromatic nitro groups.

Experimental results

Research questions

• RQ1 Can CEM identify both present and missing structural features that explain why a molecule is predicted as toxic or non-toxic?
• RQ2 Do the pertinent negatives (missing features) identified by CEM correspond to known toxicophores or structural alerts in chemical toxicity?
• RQ3 How do the CEM-generated explanations compare to established experimental findings on molecular toxicity?
• RQ4 Can CEM provide actionable insights for molecular design by suggesting minimal modifications to convert a toxic molecule into a non-toxic one?
• RQ5 To what extent do CEM explanations align with known mechanisms of toxicity, such as the role of phenolic OH in mitochondrial membrane disruption?

Key findings

• CEM successfully identified phenolic OH as a pertinent positive in molecules like tyrphostin 9 and 23, consistent with experimental findings that this group contributes to toxicity.
• The method detected known toxicophores such as aliphatic halides, polycyclic aromatic systems, and aromatic nitro groups as both pertinent positives and negatives, confirming alignment with established toxicological knowledge.
• For 2,6-dichlorophenol, CEM highlighted the phenolic OH and chlorine substituents as pertinent positives, explaining its incorrect toxic prediction.
• Pertinent negatives were found to represent missing toxicophores—e.g., the absence of a phenolic OH group in non-toxic molecules—providing insight into why some molecules are predicted as non-toxic.
• The explanations from CEM were consistent with the mechanism of weakly acidic compounds disrupting mitochondrial membrane potential, particularly through phenolic OH groups.
• The model demonstrated that CEM can generate both positive and negative explanations in a single framework, enabling reverse design of safer molecules by identifying missing toxicophores.