esm2-sequence-scoring

Pseudo-log-likelihood (PLL) scores for protein sequence plausibility can be computed using the ESM2 model. This workflow processes FASTA files to generate PLL scores and per-sequence embeddings for screening design variants.

Yes, generating protein sequence embeddings for clustering and similarity search without full structure prediction is supported. The ESM2 model provides rich sequence representations for downstream modeling directly from FASTA inputs.

ESM2 sequence scoring requires Python 3.8+ and PyTorch 1.10+, preferably PyTorch 2.0+. You also need access to the esm2 sequence scoring resources or Modal runtime as specified in the documentation.

Zero-shot variant effect screening is performed by running the ESM2 sequence-scoring workflow on a FASTA file. This yields pseudo-log-likelihood scores to quantify sequence plausibility and support design screening in protein-design pipelines.

Integrating sequence plausibility scores into protein design pipelines is done by computing ESM2 embeddings and PLL scores. These outputs support screening design variants, clustering sequence space, and feeding sequence scores into downstream QC steps.

Using ESM2 embeddings enables rapid plausibility checks and representation learning without the computational overhead of full structure prediction. This approach accelerates screening design variants and clustering sequence space efficiently.