netbiolab/LoGoPPI-Bernett

LoGoPPI Bernett

LoGoPPI is a protein–protein interaction prediction model built on ESM-2. It combines a Global sequence representation with a residue-level Maxsim score. The two calibrated scores are combined with fixed weights of 0.5 and 0.5.

Model overview

This checkpoint was trained independently on the Bernett human PPI benchmark. It does not load or continue training from the LoGoPPI Cross-species checkpoint.

Training and evaluation data

The data originate from the Bernett human PPI benchmark, Figshare version 3, which is based on high-confidence human interactions from HIPPIE v2.3 and uses leakage-controlled dataset construction.

The data/bernett/ directory contains proteins.fasta and query,text,label CSV files for training, checkpoint selection, calibration, and testing.

Split	Pairs
Training	163,192
Validation selection	29,630
Validation calibration	29,630
Test	52,048

Results

Metric	Value
AUPR	0.684319
AUROC	0.701470
NLL	0.632073
F1	0.673916

Usage

Install LoGoPPI v2.0.0 and download the model files:

git clone https://github.com/netbiolab/LoGoPPI.git
cd LoGoPPI
conda env create -f environment.yml
conda activate logoppi

python - <<'PY'
from huggingface_hub import snapshot_download

snapshot_download(
    "netbiolab/LoGoPPI-Bernett",
    revision="v2.0.0",
    local_dir="models/bernett",
    ignore_patterns=["data/*"],
)
PY

Run inference with a FASTA file and a query,text pair CSV:

python inference.py \
  --model_dir models/bernett \
  --fasta_path proteins.fasta \
  --pair_csv pairs.csv \
  --output_path predictions.csv \
  --gpus 0

Training, testing, sequence-only input, and embedding reuse are documented in the GitHub repository.

License and citation

LoGoPPI is released under the Apache License 2.0. Please cite the accompanying LoGoPPI study and the Bernett benchmark when using this model or the distributed data.