VFUSE: Virulent Feature Understanding with Sparse autoEncoders

Researchers introduce VFUSE, a mechanistic interpretability tool using sparse autoencoders to audit protein design models for hazardous features. The approach successfully identifies virulent design patterns in popular open-weight models like RoseTTAFold3 and RFDiffusion3, achieving up to 0.84 AUROC detection rates while maintaining model performance.

VFUSE addresses a critical gap in AI safety by introducing the first feature-level virulence audit of protein design models. As generative models become increasingly powerful at protein synthesis, the risk of misuse—whether accidental or intentional—grows substantially. This work demonstrates that sparse autoencoders (SAEs) can extract interpretable, monosemantic features from diffusion-transformer activations that reliably indicate hazardous protein designs, a breakthrough for mechanistic interpretability in this domain.

The methodology builds on emerging trends in AI interpretability, where researchers move beyond black-box model behavior toward understanding internal representations. By training SAEs on protein folding and synthesis models, the authors reveal that certain latent features consistently activate only for dangerous designs. The 0.84 AUROC performance significantly outperforms linear probes in original model spaces, suggesting SAE latent spaces compress safety-relevant information more efficiently.

This research carries implications for responsible AI deployment. As protein design tools proliferate in open-weight formats, auditing capabilities become essential infrastructure. The approach could establish precedent for safety audits across other generative domains—from chemistry to materials science. However, the detection of hazardous features doesn't automatically prevent their generation; enforcement mechanisms remain unclear.

Looking ahead, the field should watch for whether this audit framework becomes standard practice in protein AI releases, how organizations implement detection-based safeguards, and whether similar mechanistic interpretability approaches prove effective in other high-stakes generative domains. The tension between open research and biosecurity remains unresolved.

• →VFUSE introduces the first sparse autoencoder audit of protein design models, detecting hazardous features at 0.84 AUROC
• →Linear probes in SAE latent spaces outperform original model representations for hazard detection without sacrificing performance
• →Monosemantic features from the SAE fire exclusively on dangerous protein designs, enabling interpretable feature-level analysis
• →This represents the first mechanistic interpretability study applied to all-atom diffusion models for proteins
• →The work establishes a potential framework for safety auditing of open-weight generative models in high-stakes domains