How to Use Protein Model Inference

Inference means applying a trained model to new examples. Use this service after a successful Protein Model Fine-Tuning run to predict the same measured properties or categories for new proteins. The saved model determines the targets, class labels, and sequence/structure requirements automatically; you do not need to understand or repeat its training settings.

1. Select your trained model

Open the results page of a successfully completed Protein Model Fine-Tuning version 1.0 job and copy its 24-character job ID into Training Job ID. Your account must have access to that job. Choose a model trained on the properties and types of proteins relevant to your new candidates.

2. Prepare the proteins to predict

Upload a CSV with one protein per row and these required columns:

ID,Sequence
candidate_1,ACDEFGHIK
candidate_2,ACDEFGHIKA

IDs must be unique and Sequence must contain amino-acid letters without FASTA headers. You do not need target measurement columns, Group, or Split for prediction. Supports up to 100000 proteins and 100 MB per CSV, with a maximum sequence length of 2048 amino acids; some models may require shorter sequences. Sequences are not silently shortened to fit a model.

If the training job used Sequence mode, provide sequences only. If it used Structure mode, every row needs a Structure filename and Chain ID plus the corresponding uploaded structure. If it used Mixed mode, structures are optional per row; leave Structure and Chain blank for proteins without them. For example, candidate_1.pdb and A select chain A from that uploaded file. Every selected structure chain must exactly match its CSV Sequence, including resolved residues. Upload PDB or mmCIF files with unique names and one structural model per file.

3. Read the predictions

summary.csv identifies the model, input mode, protein count, and target count. predictions.csv preserves each protein ID and target name. Regression gives one numerical prediction per protein and target in the training target's original units. Classification gives a predicted category and a separate row for each possible class probability; the predicted category is repeated on those rows. For example, a binary target produces two probability rows for each protein. Observed values remain blank because these are new predictions.

The saved model supplies its class names and numerical scaling automatically. For binary targets, the positive class is the second class in the text-sorted label order used during training; labels 0 and 1 make 1 positive. A high class probability indicates model confidence, not a guarantee of correctness. Probabilities have not been calibrated against a separate calibration dataset.

Use the training job's independent test metrics to judge whether its predictions are useful for your application. Predictions may be less reliable for protein families or measurement conditions very different from the training data. This service applies the trained model; it does not retrain it or establish accuracy on the new proteins. Validate important predictions experimentally.