Explainable AI for Network Pharmacology: A Critical Framework for Interpreting Multi-Target Drug Discovery on Biological Graphs

Abstract

Network pharmacology examines drug action across interacting targets, pathways, and disease modules. Increasingly, however, the graph neural networks and knowledge-graph models used for this purpose are difficult to inspect. This critical narrative review considers feature attribution, attention, counterfactual and rule-based explanation, explainable graph neural networks, Bayesian and causal methods, and path-based knowledge-graph reasoning. We assess these approaches against three demands imposed by biological graphs: preservation of heterogeneous node and relation semantics, support for multi-hop reasoning, and concordance with evidence excluded from model training. Among the methods reviewed, none has yet been shown to satisfy all three demands consistently. A model-faithful explanation may still emphasize node degree, network hubs, or heavily studied entities instead of a disease-specific mechanism. To address this problem, we propose five evaluation domains: faithfulness, stability, external biological concordance, bias sensitivity, and translational utility, together with a six-item reporting standard. Both are conceptual proposals; they remain unweighted, uncalibrated, and unevaluated empirically. An exploratory, literature-based application to TxGNN, DDI-GCN, and MINERVA shows that transparency-related reporting is useful but incomplete. This exercise demonstrates how the criteria might be used; it is not an empirical evaluation or calibration of the framework. We did not retrain or rerun any model or conduct new quantitative benchmarking. The field still lacks a unified biomedical benchmark that evaluates prediction and explanation together under heterogeneous, cold-start, and provenance-aware graph settings. Progress will depend on task-matched evaluation, explicit treatment of uncertainty, and experiments designed to distinguish what a model uses from what biology actually does.