Adhesion Protein Predictor

A bioinformatics tool that scores proteins from FASTA files as secreted cell-surface glycoproteins (FLO/ALS-like). Despite the project name, this is step 1 of a multi-step pipeline and not an adhesin predictor. The removed 0.1.0 model made calls on S. cerevisiae S288C of which about 12% were known adhesins. Mechanism-specific adhesin classes are separate tools; see docs/TOOL-ARCHITECTURE.md section 2.0 for the steps and names.

Requirements

This will run with pytorch on CPUs but will be much faster if on a GPU system with torchvision installed having CUDA bindings.

Usage

Training

No trained model ships with this package. Train one before you predict. The training data in data/ are FLO, ALS1 related proteins from Saccharomyces and Candida. The model that this data produces scores surface glycoproteins, not adhesins. A model trained on a surface-glycoprotein label is planned (see docs/PLAN-2026-09-30-pipeline-and-decisions.md).

surface_glyco_train --positive data/positive --negative data/negative

Training writes the model to ./models by default. It also writes a JSON model card next to the model. surface_glyco_predict and surface_glyco_evaluate do not look in ./models. They look in the packaged model directory, or in the directory in SURFACE_GLYCO_MODELS_DIR. After training, pass the model with --model, as in the examples below, or run export SURFACE_GLYCO_MODELS_DIR=$PWD/models.

Application

You can run on a single file at a time and produce a report for each query file. Note: FungiDB raw-file downloads now require a login (HTTP 401 as of 2026-09); fetch proteomes from NCBI Datasets or UniProt, or download from FungiDB while logged in.

mkdir -p query
pushd query
curl -O https://fungidb.org/a/service/raw-files/release-68/CalbicansSC5314/fasta/data/FungiDB-68_CalbicansSC5314_AnnotatedProteins.fasta
curl -O https://fungidb.org/a/service/raw-files/release-68/CneoformansJEC21/fasta/data/FungiDB-68_CneoformansJEC21_AnnotatedProteins.fasta
curl -O https://fungidb.org/a/service/raw-files/release-68/Spombe972h/fasta/data/FungiDB-68_Spombe972h_AnnotatedProteins.fasta
popd
for qorg in $(ls query/*.fasta)
do
   surface_glyco_predict --model models/surface_glyco_model_esm2_t6_8M_UR50D.pkl --input $qorg --output $(basename $qorg .fasta).surface_glyco.csv
done

You can run on a single folder and all results will be combined in a single file. It will look for all .fasta, .fa, .pep, .aa with or without .gz extensions.

surface_glyco_predict --model models/surface_glyco_model_esm2_t6_8M_UR50D.pkl --input query --output Combinedquery.surface_glyco.csv

Development Setup

1. Install Dependencies:

   pip install -r requirements.txt

2. Setup Pre-commit Hooks:

   pre-commit install

Training writes a JSON model card next to the model (surface_glyco_model_<esm>.json) that records the ESM model, layer, pooling and training data. surface_glyco_predict and surface_glyco_evaluate read the ESM model, layer and pooling from the card. They stop with an error that names the field if the card is missing, the pooling is not supported, or --model-name differs from the card. surface_glyco_predict prints how many sequences it truncates at 1022 residues. The output column is surface_glycoprotein_score and the labels are surface_glycoprotein and other. The default 0.5 threshold is not calibrated. See docs/model-review/ for the current model review and validation plan.

cellsurface_sorting_hat (in development)

cellsurface_sorting_hat sorts the proteins of a proteome into cell surface categories. The current calls are signal_peptide_protein, tandem_repeat_protein, wall_family_domain, cocci_specificity_rank_top15, serodiagnostic_marker_candidate, iuis_allergen_similarity, iuis_allergen_homolog and cell_wall_adhesion_candidate. Design: docs/superpowers/specs/2026-10-04-orchestrator-design.md.

The core engine reads module result tables from a work directory:

cellsurface_sorting_hat --fasta proteome.faa --taxon TAXON_ID --taxdump nodes.dmp \
    --workdir work --out out

TAXON_ID is the NCBI taxon ID of the proteome. Look it up in the NCBI taxonomy.

It writes calls.long.tsv.gz, calls.wide.tsv.gz, evidence.tsv.gz, proteins.tsv.gz, report.md and run.json.

The module wrappers and the calibration commands now exist:

• cellsurface_sorting_hat_module turns the output of one tool (SignalP, Pfam hmmsearch, the repeat detectors, BLASTP against the IUIS allergens, TMHMM) or one lookup table (antigen ranking, Cys-rich tiers, spherule expression) into a module table in the work directory.
• cellsurface_sorting_hat_calibrate writes status sources from a truth set (truth), from the Phase C metrics (phasec) and runs the Pfam specificity, allergen leave-species-out and panel checks.

No calibration has been run on HPCC data yet (Tasks 11 to 15 of the plan are not run). Every module status is therefore unvalidated. The Phase C numbers for rule R0 exist, and calibrate phasec can turn them into status entries (it needs the Phase C cluster files and the SignalP module and mode that Phase C used), but no status file has been written from them in a real work directory. A status counts only for the call named by call= in its notes, and truth accepts only a call that reads one module, and only when the identity of --module in the run.json next to --calls-long matches the module record in the work directory.

Author

Jason Stajich, jason.stajichucr.edu

Code was developed with support from opencode.ai, co-pilot and various code models for setting up classifier framework