Protein language models (PLMs) are good at proposing biologically plausible protein sequences. However, steering them toward a design goal (e.g., designing novel protein sequences with increased immunogenicity) is difficult, as the PLMs’ generation process is largely uncontrolled. Activation steering of pretrained PLMs can help, as one can nudge PLMs’ internal activations along a vector (referred to as the “steering vector”) pointing from sequences that lack the desired properties toward those that possess the target properties. Such an activation steering approach is fast and training-free, but it faces several limitations in a real-world setting: when labeled data is scarce, labeling is costly, and the available labels might be noisy.
In our recent paper entitled “Optimizing protein design through uncertainty-weighted steering of protein language models,” we propose PROSOUNDS that addresses these limitations. PROSOUNDS makes steering uncertainty-aware by utilizing a probabilistic surrogate model to predict the property of interest for unlabeled sequences and then assessing the reliability of the predictions via uncertainty quantification (UQ). More specifically, UQ is used to assess the epistemic uncertainty of the surrogate, which is then used to reweight each sequence’s contribution to the steering vector such that confident predictions count more and unreliable guesses count less. PROSOUNDS also projects out domain-specific bias from the steering vector, improving its out-of-distribution performance and generalizability.
PROSOUNDS has been applied to three protein design tasks: toxicity reduction, solubility enhancement, and immunogenicity improvement. Across all three tasks, PROSOUNDS consistently and significantly beats the state-of-the-art baseline, showing its potential in real-world protein design tasks.
This work has been presented at the 25th European Conference on Computational Biology (ECCB 2026) in Geneva, Switzerland. For details, please refer to the ECCB conference proceedings published in the journal Bioinformatics:
Alif Bin Abdul Qayyum, Yingtong Zhou, Xiaoning Qian, Byung-Jun Yoon, Optimizing protein design through uncertainty-weighted steering of protein language models, Bioinformatics, Volume 42, Issue Supplement_2, August 2026, btag436, https://doi.org/10.1093/bioinformatics/btag436