Use Protein Model Fine-Tuning to learn measured properties or class labels from protein data, then Protein Model Inference to score new candidates with the saved training model. This is a supervised prediction workflow; it does not automatically design or experimentally validate proteins.
Define the dataset and target
Specify whether each endpoint is numerical regression or classification. Keep units and assay conditions consistent. Avoid treating missing measurements as zero or interpreting differences in assay protocol as sequence effects.
The service accepts a dataset CSV with ID and Sequence. Add measured target columns for training and list them in Targets. Every training row needs at least one known target; blank target cells represent missing measurements.
ID,Sequence,Activity,Group
protein_1,MQIFVKTLTGKTITLEVEPSDTIENVKAKIQDKEGIPPDQQRLIFAGKQLEDGRTLSDYNIQKESTLHLVLRLRGG,1.2,family_a
This row illustrates formatting only. It is not a measured ubiquitin dataset or enough data to train a model. Review the current training form for accepted models, target settings, limits, and split options.
Prevent leakage
Separate training, validation, and test examples according to the generalization question. Closely related sequences, shared constructs, or repeated measurements can inflate random-split performance.
Use Group or explicit Split columns according to the service’s documented options when appropriate. Confirm realized splits in the output. A held-out test set should represent future candidates and should not influence architecture, hyperparameter, or checkpoint selection.
Supply structural context when required
Sequence-only modes do not require structures. Structure or mixed modes for compatible models require matching uploaded structures and CSV Structure and Chain references.
The chosen chain sequence must match the CSV sequence, including unresolved residues. Use unique filenames and one structural model per file. Repairing or truncating a structure without reconciling its sequence can silently change the training problem.
Train and evaluate
Review compute requirements before submitting. The current service requires at least 300 available credits to submit training; this is not a guaranteed total charge.
Inspect training curves, summary.csv, per-target metrics.csv, predictions, and splits.csv. The selected checkpoint uses validation performance; test data does not select it. Interpret regression errors in the target’s units and classification metrics with class balance in mind.
The saved best_model.pt includes learned weights and reuse settings. Preserve the training job, dataset provenance, selected targets, split decisions, and package/model context with your report.
Run inference on new candidates
Open Protein Model Inference and select the compatible training job. New CSV inputs need ID and Sequence, plus matching structures when the trained model requires them. You do not need measured target columns for prediction.
Review predictions.csv in the training target’s units and label definitions. Probabilities should not be assumed experimentally calibrated. Evaluate new candidates experimentally, especially when their sequences or assay conditions differ substantially from the training data.