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

2026-05-23 · Source: cs.LG updates on arXiv.org · Field: Science & Research — Artificial Intelligence & Machine Learning, Health & Medical Research, Research Methodology & Innovation · Depth: Expert, extended

Summary

The I-SAFE (Interventional Secure, Accurate, Fair and Explainable) framework introduces a post-hoc distributional auditing method for scientific AI models, utilizing Wasserstein Coherence Metrics (WCM). This framework evaluates how raw model outputs reorganize under structurally guided input perturbations, addressing the fragility of interpreting strong benchmark performance as evidence of scientifically meaningful behavior. Applied to drug–target interaction (DTI) prediction on the Davis kinase benchmark, I-SAFE audited DeepConvDTI, DeepDTA, and TAPB models using KLIFS binding-pocket annotations as a structural prior. Despite comparable predictive performance (DeepConvDTI 0.876 AUROC, TAPB 0.882 AUROC, DeepDTA 0.907 AUROC), I-SAFE revealed substantially different distributional response profiles. Notably, TAPB exhibited significantly more coherent quantile-level and ordinal responses under KLIFS-aligned pocket perturbations (ΔQBM=+0.093, ΔWCM=+0.063) than non-pocket controls, a distinction invisible to accuracy-based evaluation. The framework is model-agnostic and applicable where inputs admit structured decomposition and an external prior.

Key takeaway

For AI Scientists and Machine Learning Engineers developing scientific prediction models, relying solely on benchmark accuracy is insufficient. You should integrate I-SAFE's distributional auditing to verify if your models align with domain-relevant structural knowledge. This framework helps you understand how model outputs reorganize under targeted perturbations, revealing crucial insights beyond predictive performance. Consider applying QBM and WCM to assess location and ordinal coherence, especially for black-box models where retraining is impractical.

Key insights

Scientific AI model performance benchmarks are insufficient to confirm alignment with domain-relevant structural behavior.

Principles

• Predictive accuracy does not imply structural alignment.
• Interventional evaluation reveals model response to controlled input changes.
• Distributional metrics offer richer insights than scalar summaries.

Method

I-SAFE is a post-hoc framework using a structural prior to define mechanistic and spurious input perturbations. It measures output-distribution coherence via Quantile-Based Metric (QBM), Wasserstein Coherence Metric (WCM), and Translation-Invariant WCM.

In practice

• Apply I-SAFE to black-box scientific predictors.
• Use KLIFS annotations as a structural prior for DTI models.
• Compare QBM and WCM contrasts for location and ordinal coherence.

Topics

• Scientific AI
• Model Auditing
• Wasserstein Coherence Metrics
• Drug-Target Interaction
• Deep Learning Interpretability
• Structural Priors

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• When Probing Accuracy Saturates, Fragility Resolves: A Complementary Metric for LLM Pre-Training Analysis — Fragility, a new metric, reveals LLM pre-training dynamics invisible to standard probing accuracy.
• Conformal Graph Prediction with Z-Gromov Wasserstein Distances — Conformal prediction for graph-valued outputs ensures coverage guarantees using Z-Gromov-Wasserstein distances and adaptive SCQR.
• Benchmarking Instance-Dependent Label Noise with Controlled Corruptions — CILN generates explicit, controllable instance-dependent label noise via input corruptions, revealing new algorithm failure modes.
• Partial Causal Structure Learning for Valid Selective Conformal Inference under Interventions — Selective conformal prediction under interventions can be robustly maintained by learning partial causal structures.
• What properties of reasoning supervision are associated with improved downstream model quality? — Intrinsic data metrics can reliably predict reasoning dataset utility before model training, with scale-dependent predictors.
• The Blind Spot in AI Safety — AI safety research's focus on coherent vs. incoherent failure overlooks "epistemic drift," a critical unmeasured failure mode.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.