PROST

What PROST does

PROST, the Protein Language Search Tool, finds remote homologs that sequence-alignment tools miss, and annotates them, in a single pass. It reads your sequence with a protein language model rather than aligning letters, so it detects relationships that survive in the biophysics long after the sequence similarity has eroded. On remote-homology benchmarks it outperforms CS-BLAST and PHMMER.

Paste a sequence, or press Example input to load one. Every hit comes back with an e-value and the enriched GO terms its homologs share, so a search doubles as a first functional hypothesis for an uncharacterized protein. A short protein comes back in a few seconds; a 3000-residue one takes around two minutes.

How the search works

The language model turns your sequence into per-residue embeddings, which a discrete cosine transform compresses into a fixed 475-byte fingerprint. PROST compares that fingerprint against all 569,182 SwissProt proteins, fits the resulting distance distribution, and converts each distance into an e-value from the z-score's normal CDF with a Bonferroni correction for the size of the database. Because every protein is one short vector, the whole database is searched exhaustively — nothing is pruned away by a heuristic first pass.

Automatic GO enrichment

The second threshold picks which homologs feed the enrichment step. For each GO term carried by those homologs, PROST builds a 2×2 contingency table against the term's background frequency across SwissProt and applies a chi-square test with Yates' continuity correction. The p-values are Bonferroni-corrected, and terms below 0.05 are reported, most significant first.

Notes

• This webserver is experimental and runs one search at a time. If someone else is already searching, yours queues behind theirs and the waiting page tells you how many are ahead of you — nothing is dropped.
• Result links are durable — bookmark or share the /result URL, and resubmitting an identical query returns the stored answer immediately.
• For batch processing, problem reports, or anything the form cannot express, email mesihkilinc@gmail.com.