Seminar, Hosted by Foresight Institute — Computation Group
One Nucleotide Is Enough: Adversarial Robustness of Genomic Foundation Models
New Mexico Institute of Mining and Technology
Abstract
Protein language models fine-tuned for clinical variant interpretation perform well at baseline, reaching an area under the ROC curve of 0.761 on a cardiomyopathy evaluation set, competitive with established predictors at clinically relevant operating points. Yet a single nucleotide edit elsewhere in the gene, one that does not change the label, drops performance to 0.394: rather than degrading towards chance, the model's ranking systematically inverts. Huixin Zhan examines the technical mechanisms behind this instability and its clinical implications, starting from the fundamentals of sequence-based language models, contrasting natural-language corpora with protein sequence modelling, and deriving core ideas of adversarial machine learning from the fast gradient sign method to advanced iterative attacks. She introduces SafeGenes, a framework that evaluates robustness under two clinically motivated threat scenarios through continuous embedding perturbations, biologically plausible single-nucleotide edits and representation-level stress tests of model reuse. Across five pretrained backbones and two cardiac variant datasets, iterative attacks push the false-positive rate to nearly 1.0 at 95% sensitivity and displace a third of true pathogenic variants from a fixed-size clinical review queue. The talk closes by distinguishing theoretical robustness stress tests from practical deployment exploits and analysing the security perimeter that emerges as these models enter agentic pipelines involving evidence retrieval, knowledge-graph traversal and shared database writes.
Topics
- Genomic Foundation Models, broader topic: foundation models
- adversarial robustness
- variant interpretation
- protein language models
- clinical variant interpretation, broader topic: variant interpretation
- single-nucleotide edits, search World Wide
- SafeGenes, search World Wide
- cardiomyopathy, search World Wide
- fast gradient sign method, search World Wide
- iterative attacks, search World Wide
- pathogenic variants