Protein folding neural networks can be subject to adversarial attacks

Summary

Protein folding neural networks can be subject to adversarial attacks. Small, calculated changes in AA sequence via adversarial attacks can lead to huge RMSD changes in AlphaFold2 and RosettaFold, and there are strategies to identify these ((1), (2), (3)). This is compared to adversarial attacks in vision NNs. It is unclear how correlated to weaknesses of these NNs are.

See also

• Inversion of protein folding neural networks

1.

Jha SK, Ramanathan A, Ewetz R, Velasquez A, Jha S. Protein Folding Neural Networks Are Not Robust. 2021; Available from: https://arxiv.org/abs/2109.04460

2.

Alkhouri I, Jha S, Beckus A, Atia G, Velasquez A, Ewetz R, et al. On the Robustness of AlphaFold: A COVID-19 Case Study. 2023; Available from: https://arxiv.org/abs/2301.04093

3.

Yuan Z, Shen T, Xu S, Yu L, Pang B, Gan B, et al. AF2-mutation: adversarial sequence mutations against AlphaFold2 in protein tertiary structure prediction. Acta Materia Medica. 2024;3(4). Available from: https://doi.org/10.15212/amm-2024-0047