Paper List

Journal: ArXiv Preprint
Published: Unknown
BioinformaticsAI/ML

DeepFRI Demystified: Interpretability vs. Accuracy in AI Protein Function Prediction

Yale University | Microsoft

Ananya Krishna, Valentina Simon, Arjan Kohli
Figure
Figure
Figure
Figure

The 30-Second View

IN SHORT: This study addresses the critical gap between high predictive accuracy and biological interpretability in DeepFRI, revealing that the model often prioritizes structural motifs over functional residues, complicating reliable identification of drug targets.

Innovation (TL;DR)

  • Methodology Comprehensive benchmarking of three post-hoc explainability methods (GradCAM, Excitation Backpropagation, PGExplainer) on DeepFRI with quantitative sparsity analysis.
  • Methodology Development of a modified DeepFool adversarial testing framework for protein sequences, measuring mutation thresholds required for misclassification.
  • Biology Revealed that DeepFRI prioritizes amino acids controlling protein structure over function in >50% of tested proteins, highlighting a fundamental accuracy-interpretability trade-off.

Key conclusions

  • DeepFRI required 206 mutations (62.4% of 330 residues) in the lac repressor for misclassification, demonstrating extreme robustness but potentially missing subtle functional alterations.
  • Explainability methods showed significant granularity differences: PGExplainer was 3× sparser than GradCAM and 17× sparser than Excitation Backpropagation across 124 binding proteins.
  • All three methods converged on biochemically critical P-loop residues (0-20) in ARF6 GTPase, validating DeepFRI's focus on conserved functional motifs in straightforward domains.
Background and Gap： Current protein function prediction models lack comprehensive interpretability analysis and robustness testing, creating barriers to adoption in drug discovery where regulatory compliance requires transparent, verifiable AI decisions.

Abstract: Machine learning technologies for protein function prediction are black box models. Despite their potential to identify key drug targets with high accuracy and accelerate therapy development, the adoption of these methods depends on verifying their findings. This study evaluates DeepFRI, a leading Graph Convolutional Network (GCN)-based tool, using advanced explainability techniques—GradCAM, Excitation Backpropagation, and PGExplainer—and adversarial robustness tests. Our findings reveal that the model’s predictions often prioritize conserved motifs over truly deterministic residues, complicating the identification of functional sites. Quantitative analyses show that explainability methods differ significantly in granularity, with GradCAM providing broad relevance and PGExplainer pinpointing specific active sites. These results highlight trade-offs between accuracy and interpretability, suggesting areas for improvement in DeepFRI’s architecture to enhance its trustworthiness in drug discovery and regulatory settings.