(Be Cautious!) Bio-Foundation Models Are Not Yet Robust to Biological Plausible Perturbations and ML Transformations

Jinhao Duan, Ruichen Zhang, Gengwei Zhang, Huaizhi Qu, Jie Peng, Sijia Liu, Tianlong Chen

Keywords: bio-foundation models, trustworthy foundation models, robustness

Abstract: Biological Foundation Models (Bio-FMs) have demonstrated remarkable success across diverse biomedical domains, enabling advances in drug discovery, protein design, and molecular analysis. However, the robustness of Bio-FMs remains underexplored, particularly in terms of the unique risks and perturbations they may encounter in real-world deployment and how these challenges impact their utility. In this work, we characterize the robustness of Bio-FMs from both biology and machine learning (ML) perspectives, and we observe that Bio-FMs are not yet robust to biological data curation and ML transformations. Specifically, (i) from the biological data curation perspective, we design biologically plausible perturbations that mimic corruptions commonly observed in biological experiments, and assess their impact on Bio-FMs; (ii) from the ML perspective, we probe how data transformations, preprocessing, and embedding affect model performance. We systematically evaluate state-of-the-art Bio-FMs on a spectrum of protein-related downstream tasks, spanning protein design, generation, function prediction, cryo-EM reconstruction, and structure classification, over structure, sequence, and image modalities. Our results reveal that most Bio-FMs are vulnerable to both ML transformations and biological perturbations; however, cryo-EM reconstruction models (e.g., CryoDRGN) exhibit a surprising robustness, which maintains stability even under worst-case adversarial scenarios. Notably, we also find that subtle biological perturbations, which are often imperceptible to current measurement tools, yet induce severe discrepancies in Bio-FM outputs, leading to critical failures. Our work highlights underappreciated vulnerabilities and provides a new perspective for evaluating and improving the trustworthiness of Bio-FMs.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 6535