I-SAFE: Wasserstein Coherence Metrics for Structural Auditing of Scientific AI Models

2026-05-20

AI Summary

Scientific AI models often rely on data shortcuts rather than genuine mechanistic understanding, leading to predictions lacking scientific interpretability. To address this, this work proposes the I-SAFE framework, which uniquely integrates structured interventions with distributional consistency metrics to audit black-box models under structurally guided perturbations. The framework introduces three complementary dimensions—position, ordinality, and shape—and incorporates external structural priors, such as KLIFS binding pocket annotations, evaluated via Wasserstein distance, quantile-based measures, and a translation-invariant variant of Wasserstein Consistency Metric (WCM). Validated on the Davis kinase drug–target interaction (DTI) task, I-SAFE successfully distinguishes the response characteristics of DeepConvDTI, DeepDTA, and TAPB, demonstrating both its effectiveness and model-agnostic applicability.

Abstract

Deep learning models are increasingly used in scientific prediction tasks where strong benchmark performance is often interpreted as evidence of scientifically meaningful behavior. This interpretation is fragile, as models may exploit shortcut features, dataset-specific regularities, or distributional biases that are predictive on held-out data but not aligned with domain-relevant structure. To address this limitation, we introduce the \textsc{I-SAFE} (Interventional Secure, Accurate, Fair and Explainable) framework, a post-hoc distributional auditing framework for scientific AI models centered on the Wasserstein Coherence Metric (WCM). Given a trained black-box predictor and an external structural prior encoding domain knowledge about task-relevant input structure, \textsc{I-SAFE} evaluates raw model outputs under structurally guided perturbations of the input. The proposed audit measures output-distribution coherence through three complementary metrics: a Quantile-Based Metric (QBM) for location-level coherence, the WCM for ordinal coherence, and a translation-invariant WCM variant for shape coherence. We instantiate \textsc{I-SAFE} on drug--target interaction (DTI) prediction using the Davis kinase benchmark, KLIFS (Kinase--Ligand Interaction Fingerprints and Structures) binding-pocket annotations, and three sequence-based DTI models: DeepConvDTI, DeepDTA, and TAPB. Although the models operate in a comparable predictive regime, \textsc{I-SAFE} reveals substantially different distributional response profiles, a distinction invisible to accuracy-based evaluation. The framework is model-agnostic and applicable to any domain where inputs admit a structured decomposition and an external prior is available.

Problem

• scientific AI models
• structural coherence
• distributional bias
• shortcut features
• model auditing

Innovation

• Wasserstein Coherence Metric
• structural auditing
• I-SAFE
• distributional coherence
• scientific AI