New training-free method generates novel protein sequences from small alignments

Researchers have developed a novel training-free method called stochastic attention (SA) for generating protein sequences from small alignment families. Unlike traditional deep learning models that require extensive data, SA utilizes a Hopfield energy model and Langevin dynamics to draw samples, eliminating the need for GPUs or pre-training. This approach has demonstrated success across various protein families, producing sequences with low composition divergence, novelty, and structural plausibility, outperforming existing methods like profile HMMs and the MSA Transformer. SA's ability to operate automatically from seed alignments and its favorable scoring by independent language models make it a promising tool for the vast number of small protein families. AI

IMPACT Enables novel protein design for small families previously inaccessible to deep learning methods.

RANK_REASON The cluster contains a research paper detailing a new method for protein sequence generation. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• AlphaFold2
• ESM2-650M
• ESMFold
• EvoDiff
• Hopfield energy
• Langevin dynamics
• MSA Transformer
• Pfam
• profile hidden Markov models
• stochastic attention

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →