Help
How to submit sequences to NABpred and how to read what it returns.
What NABpred does
NABpred predicts whether a nucleic-acid-binding protein (NAB) — an RNA-binding protein, transcription factor, or chromatin regulator — is likely to undergo liquid–liquid phase separation (LLPS). Submit a protein sequence (single, or a batch of up to 16 sequences) and NABpred returns an LLPS-propensity score along with supporting context to help you judge the prediction.
Try it yourself
On the homepage, click TDP-43 next to “load:” above the sequence box to fill in an example — human TDP-43 (UniProt Q13148, full length) — then press Run prediction. Or open the live worked example, which shows real output for this sequence generated by the same pipeline your own submissions use.
Input
- Formats: a raw sequence, or FASTA (one or more records) pasted into the sequence box; or a CSV, TSV or Excel file (column 1 = identifier, column 2 = sequence) via Batch upload.
- Length: 10–2,000 residues. Sequences longer than 1,022 residues are scored on their first 1,022 residues (the ESM2 input limit); the result says so.
- Alphabet: the 20 standard amino acids. Whitespace, line breaks, position numbers and a trailing stop
*are ignored; ambiguous codes (X, B, Z, U, O, J), alignment gaps and other characters are reported as errors rather than silently removed. - Batch size: up to 16 sequences per submission.
- Organism category (optional): choose it above the input box to see the reliability for that group. One category applies to the whole submission.
- Response time: usually a few seconds per sequence; the first request after a quiet period can take up to about a minute while the protein language model starts up.
How to read your results
1. LLPS score (0–1)
The core prediction: the average of the probabilities from a five-model ensemble. Higher values indicate greater predicted propensity to undergo phase separation. 0.5 is the default decision threshold (≥ 0.5 → “LLPS”). The Score margin label (high / moderate / borderline) only describes how far the score is from 0.5; it is not a reliability estimate.
2. Percentile context
Your score is placed against the distribution of scores on NABpred’s validated 470-protein NAB holdout set, e.g. “This score is higher than 82% of proteins in our validated reference set (n=470).” The percentile is approximate: live predictions use Hugging Face–hosted ESM2 embeddings, whereas the reference distribution was computed with the local embedding pipeline.
3. Taxonomy-aware reliability
If you specify the organism group of your query, NABpred displays an empirically measured reliability tier based on its benchmarked performance on that group (Hou et al. cross-species benchmark, 6 taxa: Animals-Mammal, Animals-Non-mammal, Fungi, Plants, Prokaryotes, Viruses; n = 2,420 in total). This reflects historical accuracy for that group, not a per-prediction confidence score.
| Organism category | Tier | Held-out AUROC |
|---|---|---|
| Mammal | high | 0.79 |
| Non-mammalian animal | high | 0.75 |
| Prokaryote | moderate | 0.68 |
| Fungi | moderate | 0.67 |
| Plant | moderate-low | 0.61 |
| Viral | low | 0.38 |
Viral proteins: in held-out testing NABpred performed below chance on viral sequences (AUROC 0.38). Treat viral predictions as unreliable; tools such as catGRANULE2 or DeePhase are better validated for this category.
4. Disorder profile
A per-residue intrinsic-disorder propensity plot (FoldIndex-based: Kyte–Doolittle hydropathy and net charge over a 51-residue window, sign-inverted so that values above zero are disorder-leaning), drawn along your sequence from residue 1 to the last residue scored, as a visual complement to the single aggregate score. It is a sequence-composition estimate, not a dedicated disorder predictor.
Batch submission
Paste several FASTA records into the sequence box, or upload a CSV/TSV/Excel file under Batch upload (up to 16 sequences per batch). Results include:
- a summary (counts, mean/median and top score), a histogram of the batch’s scores and a top-5 leaderboard;
- a results table with score, call, percentile and reliability tier per row, filterable to all rows, LLPS calls only, or errors — invalid entries are listed with the reason rather than dropped;
- downloads: CSV (one row per sequence: score, call, percentile, reliability, mean disorder, errors) and JSON (the complete output, including per-residue disorder profiles).
Click any row to see that sequence’s full detailed result (gauge, reliability banner, percentile note and disorder profile).
Programmatic access (API)
Send a JSON request to POST https://nabpred.com/nabpred/predict:
curl -X POST https://nabpred.com/nabpred/predict \
-H "Content-Type: application/json" \
-d '{"sequence": "MSEYIRVTEDENDEPIEIPSEDDGTVLLSTVTAQFPGACGLRYRNPVSQCMRGV", "taxon": "animal_mammal"}'
Use "sequences": [{"id": "...", "sequence": "..."}, …] for up to 16 sequences per request. Errors are returned as JSON with a readable error message. See the API section on the homepage for all fields.
Limitations (please read)
- NABpred is trained and validated specifically on nucleic-acid-binding proteins. Predictions on non-NAB proteins are out of scope and not validated.
- Performance varies by taxonomic group; viral protein predictions are the least reliable in our validation (below chance).
- Percentiles are approximate (see above), and the disorder profile is a composition-based estimate.
- Predictions are computational estimates, not experimental measurements, and should be followed up with experimental validation.
Data & privacy
NABpred does not require registration, does not set cookies and does not store submitted sequences. To compute ESM2 embeddings, each sequence is sent to Hugging Face’s hosted inference service.
Licence
NABpred is freely available under the Creative Commons Attribution 4.0 International licence (CC BY 4.0).
Questions: Prof. Rizwan Hasan Khan, Interdisciplinary Biotechnology Unit, Aligarh Muslim University, Aligarh 202002, India.