Likelihood Curves (TrpB)

Architecture Breakdown

Data: TrpB sequences from SaProtHub (used as evaluation targets, not training data).

Models: ESMC (pretrained, evaluated — not trained) → generative_modeling.

Sampling: None (evaluation only).

Evaluation: compute_log_prob_trajectory + plot_log_prob_trajectories — measures log p(true token) at masked positions across noise levels → Likelihood Curves, evaluation.

Evaluate how well a model predicts masked amino acids under progressive unmasking, using real TrpB sequences from SaProtHub.

Quick Start

uv run python examples/trpb_likelihood_curves.py --device cuda --n-sequences 50 --n-time-points 20

What It Produces

A plot showing average log p(true token) at masked positions vs. fraction unmasked. This characterizes how a model's predictive confidence changes as it sees more context — a key diagnostic for masked language models.

See the Likelihood Curves workflow for interpretation and usage.

Source: examples/trpb_likelihood_curves.py