// Welcome to PATHOS webserver

Predict variant pathogenicity from protein sequence.

PATHOS is a deep-learning tool that scores the pathogenic potential of protein variants directly from sequence. Variants above the 0.63 threshold are flagged as likely pathogenic.

Cite
Radjasandirane R., Cretin G., Diharce J., de Brevern A.G., & Gelly J.C. (2026). PATHOS: Predicting Variant Pathogenicity by Combining Protein Language Models and Biological Features. AI in the Life Sciences, 100165. doi:10.1016/j.ailsci.2026.100165
20,416
Proteins
Full human proteome, precomputed
216M
Mutations
All SwissProt single-residue variants
682M
Embeddings
PLM precomputed embeddings

Run offline: install the standalone package

github.com/DSIMB/PATHOS

PLM embeddings: precomputed for all 20,416 human proteins across 3 PLMs
682M embeddings · 6.5 TB on Hugging Face

huggingface.co/DSIMB/PATHOS-PLM-EMBEDDINGS
Prediction Input
Input UniProt IDs
Limits: Max 50 proteins per job · Max 10 full-landscape proteins · Max 1,000 mutations per protein · Max 50,000 total mutations · File upload max 1 MB.
In Scores only mode, limits do not apply (max 300,000 mutations).
Or, upload a TXT, CSV, or TSV file:
Format: first column is the UniProt ID, remaining columns or whitespace-separated values are mutations.

Job options
Webserver Statistics
Total Jobs
Unique Users
Jobs (Last 7 Days)
Mutations Analyzed
API Calls
Activity (Last 30 Days)
Geographic Distribution