XAI4Chem: Contrastive explanations

Previous posts have discussed explanations and model explainability.

One area of interest, both in academia and the pharmaceutical industries, is the explainability of neural networks predicting structures or properties.

Here is an example, from the paper "Enhancing preclinical drug discovery with artificial intelligence" (the references were edited out, and can be found in the original article):

(...) many successful applications of QSAR-based VS workflows for hit identification have been reported. Zhang et al. described the successful implementation of an ML-based QSAR workflow for VS that led to the discovery of novel antimalarial agents.

The ML algorithms were trained on a small dataset of antimalarial drugs. The authors continue:

The QSAR models were used to carry out VS against the ChemBridge database and resulted in the selection of 174 compounds for a follow-up screening in Plasmodium falciparum growth inhibition and cellular assays. Experimental validation revealed 25 of the selected compounds to be active, yielding a hit rate of 14.2%, with the most potent hit having an EC50 value of 95.6 nM. Subsequently, many studies have reported the application of ML and DL-based QSAR workflows as promising VS tools.

That sound cool and useful. A human expert may be able to detect a pattern on those structures.

• But can we figure out an evaluation mechanism that singles out which fragments are most important?

• At the same time, isolated fragments may be misleading, because the shape of the molecule is also as important. Maybe a "fragments vs shape" could make up a directionally useful plot?

Paper Idea

Feature Attribution Methods, such as SHAP and LIME help answer the question "Why was P obtained?".

On the other hand, contrastive explanations answer the question "Why P rather than Q?".

Here, "Q" is the foil (an alternative event), and it's an important tool to make explanations simpler: the aspects that are similar do not need to be explained, it is the differences that count towards explaining "Why P rather than Q".

The paper "Contrastive explanations for machine learning predictions in chemistry" (2025) develops a framework called MolCE for:

deriving contrastive explanations for machine learning models in chemistry to systematically generate intuitive explanations of predictions in high-dimensional feature spaces.