Log Odds Mutation Scoring

log-odds-mutation-scoring

Pro-PRIME: A general Temperature-Guided Language model to engineer enhanced Stability and Activity in Proteins — Jiang et al. (2023) (arXiv:2307.12682, 2023)

What this evaluates

Evaluates a protein language model's zero-shot capability to score single-site amino acid mutations by comparing contextual likelihoods of wild-type versus mutant residues. It probes how well masked language modeling objectives capture evolutionary and structural constraints for mutant effect prediction.

Datasets

• Unspecified protein sequences — total ?; splits: test (-1)

Metrics

• log-odds ratio (primary) — range: other
  • Computed as the difference between the log-likelihood of the mutated amino acid and the wild-type amino acid at a specific position, conditioned on the surrounding sequence context.

Input / output format

Input: A protein sequence containing a single-site mutation (wild-type and mutant amino acids at the target position).

Output: A scalar log-odds score representing the relative likelihood of the mutation.

Scoring recipe

def compute_log_odds(wt_seq, mut_seq, model):
    wt_log_prob = model.log_prob(wt_seq)
    mut_log_prob = model.log_prob(mut_seq)
    return mut_log_prob - wt_log_prob

Common pitfalls

• Zero-shot log-odds scores may not directly correlate with experimental stability or activity without calibration.
• Fine-tuning on homologous sequences is unsupervised and can introduce phylogenetic bias if not controlled.

Evidence (verbatim from paper)

The mutations are then scored using the log-odds ratio at the mutated position. (See Fig. 1C; the details can be found in the Methods).

Citation

@misc{jiang2023proprime,
  title={Pro-PRIME: A general Temperature-Guided Language model to engineer enhanced Stability and Activity in Proteins},
  author={Jiang et al. (2023)},
  year={2023},
  note={arXiv:2307.12682}
}

• arXiv: 2307.12682