[Paper Review] Counterfactual Explanation with Multi-Agent Reinforcement Learning for Drug Target Prediction

Tri Minh Nguyen, Thomas P. Quinn+2 more|arXiv (Cornell University)|Mar 24, 2021

Computational Drug Discovery Methods4 citations

TL;DR

This paper proposes MACDA, a multi-agent reinforcement learning framework that generates counterfactual explanations for drug-target binding affinity models by jointly optimizing drug and protein modifications. It achieves more parsimonious and valid counterfactual explanations with improved interpretability of substructure interactions, validated on the Davis dataset and a case study with ABL1 and Nilotinib.

ABSTRACT

Motivation: Many high-performance DTA models have been proposed, but they are mostly black-box and thus lack human interpretability. Explainable AI (XAI) can make DTA models more trustworthy, and can also enable scientists to distill biological knowledge from the models. Counterfactual explanation is one popular approach to explaining the behaviour of a deep neural network, which works by systematically answering the question "How would the model output change if the inputs were changed in this way?". Most counterfactual explanation methods only operate on single input data. It remains an open problem how to extend counterfactual-based XAI methods to DTA models, which have two inputs, one for drug and one for target, that also happen to be discrete in nature. Methods: We propose a multi-agent reinforcement learning framework, Multi-Agent Counterfactual Drug target binding Affinity (MACDA), to generate counterfactual explanations for the drug-protein complex. Our proposed framework provides human-interpretable counterfactual instances while optimizing both the input drug and target for counterfactual generation at the same time. Results: We benchmark the proposed MACDA framework using the Davis dataset and find that our framework produces more parsimonious explanations with no loss in explanation validity, as measured by encoding similarity and QED. We then present a case study involving ABL1 and Nilotinib to demonstrate how MACDA can explain the behaviour of a DTA model in the underlying substructure interaction between inputs in its prediction, revealing mechanisms that align with prior domain knowledge.

Motivation & Objective

• To address the lack of interpretability in high-performance black-box drug-target affinity (DTA) models.
• To extend counterfactual explanation methods to DTA models with two discrete inputs: drug molecules and protein sequences.
• To enable joint optimization of drug and protein modifications for more accurate and human-interpretable counterfactual instances.
• To isolate and reveal substructure-level interactions between drug functional groups and protein residues that influence binding affinity predictions.
• To provide a stable, scalable framework for generating counterfactuals that reflect biologically plausible changes in drug-protein complexes.

Proposed method

• MACDA employs a multi-agent reinforcement learning (MARL) framework with two agents: one for drug molecules and one for protein sequences.
• Each agent takes actions such as adding or removing atoms, bonds, residues, or motifs to generate counterfactual instances.
• The agents collaborate under a shared multi-objective reward function that maximizes binding affinity change while minimizing input modification (joint change).
• The framework uses discrete action spaces suited for molecular and sequence-level modifications, avoiding gradient-based methods unsuitable for discrete data.
• Counterfactual generation is guided by the model’s prediction shift, ensuring changes lead to significant output variation.
• The method enables joint optimization of both drug and protein inputs, avoiding suboptimal single-input counterfactuals.

Experimental results

Research questions

• RQ1 How can counterfactual explanations be effectively generated for drug-target binding affinity models with two discrete inputs?
• RQ2 Can multi-agent reinforcement learning generate more parsimonious and valid counterfactual explanations than single-input or gradient-based methods?
• RQ3 To what extent do the generated counterfactuals reflect known biological mechanisms of drug-protein interaction?
• RQ4 How do joint modifications of drug and protein affect the stability and interpretability of counterfactual explanations?
• RQ5 Can the framework isolate and highlight key substructure interactions responsible for DTA model predictions?

Key findings

• MACDA generates counterfactual explanations that are more parsimonious—requiring fewer modifications—without sacrificing explanation validity, as measured by encoding similarity and QED scores.
• The framework successfully identifies key residue positions in ABL1, such as LYS.247, GLU.286, and VAL.448, which are known binding sites for drugs like Nilotinib and Asciminib.
• In the ABL1-Nilotinib case study, MACDA highlighted the critical role of the (trifluoromethyl)benzene group interacting with GLU.286, which aligns with the crystal structure showing a hydrogen bond between them.
• Counterfactuals revealed that the interaction between the pyrimidine group of Nilotinib and MET.318 also contributes significantly to the model’s prediction.
• The method demonstrated a stable learning process and effective joint optimization of drug and protein modifications, avoiding the instability seen in single-input counterfactual generation.
• The framework successfully disentangles and explains high-order interactions between drug substructures and protein residues, supporting biological plausibility and domain knowledge validation.