FlashPPI: Linear-time prediction of proteome-scale microbial protein interactions

Model Description

FlashPPI is a contrastively trained model for protein-protein interaction (PPI) prediction, grounded in residue-level interactions, that enables full-proteome interaction prediction in minutes.

By reframing PPI prediction as a dense retrieval task, FlashPPI circumvents the $\mathcal{O}(N^2)$ computational bottleneck of traditional all-vs-all structural screening.

• Scalable: Reduces proteome-wide screening from days/months to minutes.
• Interpretable: Predicts fine-grained, residue-level 2D contact maps for retrieved interaction candidates.
• Genomic Priors: Leverages gLM2 initialization to capture cross-protein, multi-gene co-evolutionary signals.

Web Server

FlashPPI is integrated into seqhub.org. You can upload a FASTA and interactively explore whole-proteome networks and contact maps. Explore an example network here.

Installation

pip install -r requirements.txt

Optionally, install Flash Attention for faster inference on GPU:

pip install flash-attn --no-build-isolation

Usage

Fast Proteome-wide PPI Screening (All-vs-All)

Run the prediction script by passing your proteome FASTA file. It will output a predictions file with predicted pairs of interacting proteins and confidence scores. Note: Requires a machine with at least 1 GPU.

python predict_proteome.py --fasta my_proteome.fasta --output predictions.csv

Cross-Proteome PPI Screening (Host–Viral)

Predict interactions between two proteomes, for example a viral genome and its host genome.

python predict_cross_proteome.py \
    --host_fasta host.fasta \
    --viral_fasta virus.fasta \
    --output predictions.csv

Visualizing contact predictions

import torch
import matplotlib.pyplot as plt
from transformers import AutoModel, AutoTokenizer

seq1 = "MKTAYIAKQRQISFVKSHFSRQL"
seq2 = "MSTAGKVIKCKAAVLW"

device = "cuda" if torch.cuda.is_available() else "cpu"

tokenizer = AutoTokenizer.from_pretrained("tattabio/flashppi", trust_remote_code=True)
model = AutoModel.from_pretrained("tattabio/flashppi", trust_remote_code=True).to(device).eval()

inputs1 = tokenizer(seq1, return_tensors="pt").to(device)
inputs2 = tokenizer(seq2, return_tensors="pt").to(device)

with torch.no_grad():
    outputs = model(
        input_ids1=inputs1["input_ids"],
        attention_mask1=inputs1["attention_mask"],
        input_ids2=inputs2["input_ids"],
        attention_mask2=inputs2["attention_mask"],
        return_dict=True
    )

# Extract map and trim padding
contact_map = outputs.contact_map[0].cpu().numpy()
len1, len2 = inputs1["attention_mask"].sum().item(), inputs2["attention_mask"].sum().item()
contact_map = contact_map[:len1, :len2]

plt.imshow(contact_map, cmap="Blues", vmin=0, vmax=1)
plt.savefig("contact_map.png")

License

The model code and inference scripts in this repository are licensed under the Apache License 2.0.

The FlashPPI model weights are hosted on Hugging Face and released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. The weights are freely available for academic and research purposes.

Citing

If you use FlashPPI or our datasets in your research, please cite:

@article {Cornman2026FlashPPI,
	author = {Cornman, Andre and Tranzillo, Matt and Zulaybar, Nicolo G and Bouzit, Imane and Hwang, Yunha},
	title = {Linear-time prediction of proteome-scale microbial protein interactions},
	year = {2026},
	doi = {10.64898/2026.03.01.708874},
	publisher = {Cold Spring Harbor Laboratory},
	URL = {https://www.biorxiv.org/content/early/2026/03/02/2026.03.01.708874},
	journal = {bioRxiv}
}