Will this sequence bind?
A pre-screen for de novo binder campaigns. Paste a sequence, choose one of 24 targets, and the ensemble returns a calibrated probability with the evidence behind it.
Running the ensemble
-
Sequence
AI Interpretation
Groq · SHAP-groundedFeature attribution
SHAP, per-feature contributionScore a library
One sequence per line. Every candidate is ranked against the same target, with a hit rate and a shortlist at the end.
Every row is scored against the one target selected below. Drop a FASTA, CSV, TSV or plain text file on the box to fill it.
Scoring the library
Ranked candidates
AI Summary
GroqWhich single mutation helps most?
Every position is scanned against all 19 substitutions and ranked by the gain in predicted binding. This one is slow, because it is a full scan.
Scanning every position
Ranked substitutions
AI Rationale
GroqImprove a seed sequence
A genetic algorithm uses the model as a fitness oracle; ESM-2 then refines the survivors into chemically plausible variants.
Parameters
Evolving candidates
-
Optimisation trajectory
Top candidates
AI Assessment
GroqAsk about the model
Questions about features, the applicability domain, or how to read a result.
Try one of these
About
A machine-learning pre-screen for de novo binder campaigns across 24 human and viral targets, trained on experimental Proteinbase data.
Model
LightGBM over 470 features: ESM-2 target-conditioned prototype-similarity, sequence composition, physicochemical properties, design-method priors and Boltz2 structural metrics. The prototype-similarity features, meaning how closely a candidate resembles known binders of a target in ESM-2 embedding space, are the most informative predictors.
Applicability domain
Two different numbers are often quoted, and they measure different things. The published pipeline reaches AUROC 0.946 on the held-out split using a feature matrix that includes measured structural inputs (ESMFold pLDDT, ProteinMPNN, Boltz2). This tool has none of those: it computes features from sequence alone and fills the rest with training medians. Scored on the same 392 held-out pairs, the served configuration reaches AUROC 0.852, AUPRC 0.52, and enriches binders 3.4× in the top decile. Use 0.85 when judging what this tool will do for you.
Away from the training distribution it degrades further: about 0.77 on a new design campaign, about 0.55 zero-shot to a novel target, recovering to roughly 0.70 with as few as two known binders. It is a binder / non-binder classifier and does not rank affinity.
Targets
24Dataset & licence
This work used Proteinbase by Adaptyv Bio under the ODC-BY licence. 5,253 proteins · 2,630 labelled pairs · experimental binding evaluations. Model MIT · Dataset ODC-By. Paper in preparation.