TRION Cloud

TSAP (Boltz-2)

Target-specific activity prediction with a fine-tuned Boltz-2

Predict how active your compounds are against one target. Use a model we have already tuned for a target, or train your own on your activity data: Boltz-2's affinity predictor is fine-tuned on your measurements, checked on compounds it never saw, then used to score new compounds.

Boltz-2 fine-tuning (TSAP pipeline v2.2) · Doc updated Oct 14, 2026

Run example (EGFR)Run
Input
Target + activity data, or a pre-trained model
Output
Predicted activity, model quality
Engine
Boltz-2
Cost
From 100 cr, quoted before submit
Standard preset
About 1 hour

When to use

  • You have measured activities (IC50, Ki, Kd, EC50) for at least 30 compounds on one target and want to rank new ideas
  • Your target is in the pre-trained catalog and you want quick predictions without your own data
  • General docking or cofolding scores do not separate your actives well enough

A model is only as good as the data it learned from. Predictions for compounds very unlike the training set are less reliable. Always check the model's quality on held-out compounds before trusting its ranking.

How it works

TargetActivity data01Place compoundskeptdroppedkeptkeptOnly poses in thereference site pass02Split the databy activityTrainingHoldout 20%Holdout is neverused for training03Fine-tuneRest of Boltz-2, not trainedAffinity head, trainedBest epoch keptNever beats pretrained:credits refunded04Check on holdoutAccuracyreportedNew compounds04PredictPredicted activity
  1. 01Place each compound

    Boltz-2 builds structures of the target with each compound and keeps poses that sit in the reference ligand's binding site (residue-contact overlap, or core RMSD in High Accuracy).

  2. 02Split the data

    Compounds that pass are split into training and a 20% holdout set, spread across the activity range. The holdout is never used for training.

  3. 03Fine-tune

    Only Boltz-2's affinity head is trained, starting from its pretrained weights, so it learns your target's activity scale. The best epoch is kept; if training never beats the pretrained model, you get your credits back.

  4. 04Check and predict

    The tuned model scores the holdout compounds to report its accuracy, then predicts the new compounds you submitted.

Input formats

ItemAcceptedNotes
ModelA pre-trained catalog model, or train a new onePre-trained models fix their own settings and reference structure
Target (training)1–10 chains: sequence, 4-character PDB ID, or .pdb/.cif uploadFrom a structure you pick the chain to train on; the default is the longest. MSA, template and cyclic are optional per chain
Binding site reference (training)A structure with a bound ligand: .cif, .pdb, .sdf, .mol or .mol2Required when you also predict new compounds. Water, ions and additives such as GOL are not offered as ligands
Activity data (training).txt, .csv or .tsv with a header row: SMILES, activity value, optional IDAt least 30 compounds. One unit per job (M down to fM, or values already on a p-scale)
Compounds to predict.txt, .smi or .csv, one per line: SMILES, or ID then a tab then SMILESAt least one with a pre-trained model, optional when training

Activity values are converted to the p-scale: pX = −log10 of the molar concentration, so 1 nM becomes 9.

Tutorial: predict EGFR activity with a pre-trained model

EGFR, the lung cancer target from the Virtual Screening and AI-Cofolding docs, has a pre-trained TSAP model in the catalog with PDB 1M17 as its reference structure, so you only need the compounds you want scored. Run example at the top of this page fills all of this in for you.

  1. Pick the model

    In Model, keep Use a pre-trained model and choose the EGFR card. It shows the reference structure, the number of training compounds and its accuracy on held-out compounds.

    Model step with the EGFR pre-trained model card selected, showing its holdout accuracy

  2. Add compounds

    In Compounds, upload a file with one SMILES per line, optionally with an ID and a tab before it. The example uses five compounds from the EGFR sample file linked in the step; each compound adds GPU time.

    Compounds step with a compound file uploaded and its preview

  3. Check the quote and submit

    Advanced Settings are fixed by the model. In Project Settings, name the project, check the quoted credits and time, and submit.

  4. Read the predictions

    Open the project. Sort by predicted activity, check structure confidence and the binding site overlap of each compound, and open a compound to see its pose.

    TSAP result page: the prediction summary with training accuracy and its scatter, and the predicted-compounds table

Reading the results

ValueMeaningHow to read it
Pred pIC50 (or pKi, pKd, pEC50, pActivity)Predicted activity on the p-scale of your dataHigher is more active. 1 unit is a tenfold change. 7.0 and above is counted as a strong binder
Structure confidence (ipTM)Boltz-2's confidence in the predicted complex (0–1)Low values mean the pose, and so the prediction, is shakier
Binding site overlap (Jaccard) or Reference RMSDHow closely the pose matches the reference ligand's siteCompounds whose poses never reach the site are excluded and listed with the reason
Pearson r, R², Spearman ρ, RMSEModel accuracy on the holdout compoundsTraining runs show them for your model; pre-trained runs show the model's accuracy from when it was trained, not on your compounds
vs pretrainedAccuracy gain over Boltz-2 without fine-tuningShows whether fine-tuning helped

With fewer than 8 holdout compounds, the accuracy numbers are marked as not reliable.

Example result: EGFR

A real job, shown read-only. Sign in to run the same setup yourself.

CompoundPred pIC50ipTMJaccard
CHEMBL3282168.970.9480.91
CHEMBL291978.860.9470.91
CHEMBL2850637.730.9400.78
CHEMBL2900967.560.9430.81
CHEMBL2899594.940.9400.76

Cost and time

The price follows the GPU time the job needs: the number of compounds, protein length and sampling steps, plus a training step when you train. With a pre-trained model you only pay for prediction. The wizard shows the server's quote, and you pay exactly that at submit. The minimum is 100 cr (1 cr is $0.01).

PresetPoses, stepsWizard's rough guide
Quick5, 100About 30 min
Standard10, 200About 1 hour
High Accuracy20, 200About 2.5 hours

The EGFR example above (five compounds) took about 38 minutes from submit to results and cost 451 cr. Jobs that would clearly run past the 5-hour limit are blocked before you submit.

Credits come back in full if the job fails, if too few compounds remain after the binding site check, or if fine-tuning never beats the pretrained model. If you stop a job yourself, half comes back.

Citations

If you use these results in a report or paper, please cite the methods below.

  • Passaro S, Corso G, Wohlwend J, et al.. Boltz-2: Towards accurate and efficient binding affinity prediction. bioRxiv (2025). doi:10.1101/2025.06.14.659707Structure and affinity model that TSAP fine-tunes
  • RDKit contributors. RDKit: Open-source cheminformatics. Zenodo (2026). doi:10.5281/zenodo.591637SMILES parsing and checks
  • Stamos J, Sliwkowski MX, Eigenbrot C. Structure of the epidermal growth factor receptor kinase domain alone and in complex with a 4-anilinoquinazoline inhibitor. J Biol Chem 277, 46265-46272 (2002). doi:10.1074/jbc.M207135200Reference structure of the EGFR model (PDB 1M17)

FAQ