Paste a sequence and get a calibrated suitability score in seconds. A first-pass prioritizer for ranking candidates before you commit instrument time, not a verdict.
No login. Real model, real score. Try it below ↓
Real outcomes are the bottleneck. Tell us what actually happened. Even if you haven't run it yet, the planned outcome and a 2-week follow-up still help.
Then tell us the outcome. Picking one submits everything above.
Three steps, no setup. Built for bench scientists who want a sanity check before spending a week on sample prep.
Drop in a single protein in plain text or FASTA. We clean and validate it client-side before sending anything.
The v0.4 model combines 47 hand-engineered features (physicochemical, N-glycosylation sequon counts, transmembrane-helix prediction) with mean-pooled ESM-2 protein-language embeddings, trained on curated native-MS data from PDB and EuropePMC.
You see a suitability score from 0 to 100, the specific risk factors that hurt it, and concrete buffer and instrument recommendations.
Every prediction returns a calibrated score, the risk factors behind it, and whether the sample can even be resolved.
An isotonically calibrated probability with a 95% interval and an out-of-distribution flag, not a raw guess.
Membrane topology, glycosylation, aggregation and charge, each with a severity and a targeted buffer or prep recommendation.
For glycoproteins, ADCs and other heterogeneous proteins it simulates the mass envelope and computes the resolving power needed to separate the proteoforms. A spectrum fine for an intact mass may not resolve glycoforms, so suitability depends on what you need to measure. Arithmetic from published glycan masses.
Two signals, one call. A calibrated ML score (47 biophysical features plus ESM-2 embeddings) and a physics-based heterogeneity and resolvability calculator (published glycan masses, R = m / FWHM).
No raw sequence is kept. It is hashed before storage, with no login, no IP, and no email unless you choose to leave one.
Cluster-aware ROC-AUC 0.835 on an open benchmark. It is most reliable as a positive-suitability triage tool, not a validated failure detector: trust a high-confidence positive, and treat a low score as a flag for manual review.
47 biophysical features plus ESM-2 protein-language embeddings, isotonically calibrated, with an out-of-distribution flag. Glycoproteins and ADCs also get a physics-based heterogeneity and resolvability calculation.
Yes, free during beta: 50 predictions per month per IP, no login required.
Yes. There is a REST API, a Python SDK (pip install nativeready) and a CLI, all batch-capable for screening many candidates before you book instrument time.
Built by Brhanu Znabu, with an open benchmark and a bioRxiv preprint. The code, dataset and trained model are open under MIT on GitHub.
Open benchmark and triage model for native MS suitability. v0.4 trained on public datasets (PDB, UniProt, EuropePMC), with explicit N-glycosylation sequon and transmembrane-helix features. The risk panel surfaces membrane-topology and glycosylation signals with targeted experimental recommendations (nanodiscs / amphipols / SMA for membrane proteins; PNGase F deglycosylation for glycoproteins). Code, dataset, and trained model on GitHub. Methodology and benchmarks: bioRxiv preprint (DOI 10.64898/2026.05.03.722506).