Semantically Rich Local Dataset Generation for Explainable AI in Genomics

Local explanations for AI predictions in genomics are now more feasible: The paper addresses a key challenge in making black box AI models interpretable by developing a method to generate local explanations for individual genomic predictions, which could help researchers understand why AI models make specific predictions about gene regulation.

A new approach uses evolutionary algorithms to generate meaningful test cases: Rather than random testing, the researchers use Genetic Programming to systematically evolve DNA sequence variations that maintain realistic biological properties while exploring how predictions change, achieving approximately 30% better performance than random approaches.

The method balances two competing requirements in responsible AI: The approach must keep generated sequences biologically plausible (syntactically similar to real DNA) while still creating meaningful variations that change model predictions (semantic diversity)—a tension that mirrors broader responsible AI challenges of maintaining realism while enabling thorough testing.

Scalability to real-world genomic data remains demonstrated but limited: While the proof-of-concept shows promise on short RNA sequences and demonstrates generalizability to larger sequences, the paper focuses on relatively constrained genomic domains (RNA splicing), leaving questions about applicability to more complex genomic prediction tasks.

This work supports more trustworthy biomedical AI applications: By making genomic AI models more interpretable and testable, this approach could help clinicians and researchers better understand and validate AI-assisted predictions before using them in healthcare decisions.