Step-by-step workflow for integrating a new generative (transition) model into the proteingen library. Covers choosing between GenerativeModel and GenerativeModelWithEmbedding, implementing abstract methods, writing tests, and avoiding common gotchas. This skill is for generative models only โ not PredictiveModel subclasses.
Step-by-step workflow for integrating a new predictive model into the proteingen library. Covers decomposing a pretrained predictor into four layers (raw model, binary logit function, template model class, PredictiveModel subclass), identifying what already exists vs what's missing, and only building the missing pieces. This skill is for predictive models only โ not GenerativeModel subclasses.
Guide the user through planning and implementing a protein design pipeline by following ProteinGen workflows. Use when the user says they want to follow a workflow (e.g. ProteinGuide, Continued Pretraining) or wants help planning their library design pipeline. Walks through documentation in-order and recursively, helping the user make decisions at each step before writing code.
Evaluate and plot log-likelihood trajectories for generative models under progressive unmasking. Use when comparing how well one or more generative models predict masked amino acids as context is revealed, or when evaluating fine-tuned vs base models on a protein dataset.