Inference-Time Toxicity Mitigation in Protein Language Models

Researchers developed Logit Diff Amplification (LDA) as an inference-time safety mechanism for protein language models to prevent toxic protein generation. The method reduces predicted toxicity rates while maintaining biological plausibility and structural viability, addressing dual-use safety concerns in AI-driven protein design.

• →Protein language models can inadvertently generate toxic proteins through domain adaptation to specific taxonomic groups.
• →Logit Diff Amplification (LDA) provides inference-time toxicity control without requiring model retraining.
• →LDA consistently reduces predicted toxicity rates across four taxonomic groups while preserving biological plausibility.
• →The method maintains distributional similarity to natural proteins and structural viability better than activation-based steering methods.
• →This research addresses growing safety concerns around dual-use potential of AI protein design tools.