Hybrid Gated Fusion: A Multimodal Deep Learning Framework for Protein Functional Annotation

A PyTorch framework for multimodal protein function prediction that integrates four modalities: sequence, text, structure, and protein-protein interaction (PPI) embeddings using a hybrid gated bilinear fusion architecture.

Overview

MMFP predicts Gene Ontology (GO) term annotations for proteins by combining information from multiple modalities using advanced fusion techniques. The framework supports:

• Three GO aspects: Biological Process (BPO), Cellular Component (CCO), Molecular Function (MFO)
• Multiple fusion methods: Concatenation, Bilinear Gated Fusion, Hybrid (bilinear + auxiliary heads)
• Flexible modality handling: Gracefully handles missing modalities via learned masking
• CAFA-compliant evaluation: Standard evaluation metrics including F-max, weighted F-max, and S-min
• Temporal decontamination: Historical UniProt text (pre-CAFA3 cutoff) via UniSave to prevent data leakage

Temporal Decontamination

To prevent data leakage, test-set text embeddings use historical UniProt records retrieved from UniSave with a cutoff date of 2016-02-17 (before the CAFA3 assessment period). Train/validation text embeddings use current UniProt descriptions. This ensures the model cannot exploit post-assessment functional annotations at test time.

Installation

git clone https://github.com/psipred/PFP.git
cd PFP/MMFP
pip install -r requirements.txt

Data Preparation

Option A: Download Precomputed Data (Recommended)

Download precomputed CAFA3 embeddings and data splits:

# Download precomputed data from Zenodo: https://zenodo.org/records/19498341
wget https://zenodo.org/records/19498341/files/mmfp_cafa3_data.tar.gz
# Extract to data directory
tar -xzf mmfp_cafa3_data.tar.gz -C ./data

Required data:

Directory	Description	Size
embedding_cache/prott5/	ProtT5-XL sequence embeddings (1024-D)	~550 MB
embedding_cache/exp_text_embeddings_temporal/	Temporal PubMedBERT text embeddings (768-D)	~280 MB
embedding_cache/IF1/	ESM-IF1 structure embeddings (512-D)	~270 MB
embedding_cache/ppi/	STRING PPI network embeddings (512-D)	~230 MB
{BPO,CCO,MFO}_*.npy/.npz/.json	Train/valid/test splits and GO term labels	-
{BPO,CCO,MFO}_ia.txt	Information Accretion weights	-
go.obo	Gene Ontology structure (place in parent dir)	-

Option B: Extract from Scratch

See scripts/README.md for detailed extraction instructions.

Reproducing Paper Results

Paper Table	Description	Command
Table 1	CAFA3 Comparison (Full Model)	python scripts/reproduce_full_model.py
Table 2	Modality Contribution	python scripts/reproduce_modality_contribution.py
Table 3	Ablation Study	python scripts/run_ablation.py
Table 4	Masking Evaluation	python scripts/reproduce_masking_eval.py

Training from Scratch

# Full model (Table 1)
python train.py \
--seq-model prott5 \
--fusion-types gated_bilinear \
--aspects BPO CCO MFO \
--use-late-fusion \
--text-embedding-dir data/embedding_cache/exp_text_embeddings_temporal \
--output-base results/full_model \
--seed 42
# Evaluation with CAFA metrics
python scripts/reproduce_full_model.py

Expected Results (Table 1)

Aspect	Fmax	wFmax
BPO	0.601	0.515
MFO	0.702	0.605
CCO	0.706	0.566

Project Structure

MMFP/
├── mmfp/
│ ├── models.py # Fusion models and architectures
│ ├── dataset.py # Dataset and data loading
│ └── evaluation.py # CAFA evaluation utilities
├── scripts/
│ ├── reproduce_full_model.py # Table 1 reproduction
│ ├── reproduce_modality_contribution.py # Table 2 reproduction
│ ├── run_ablation.py # Table 3 reproduction
│ ├── reproduce_masking_eval.py # Table 4 reproduction
│ ├── extract_uniprot_text.py # Text extraction + temporal bundle
│ ├── embed_uniprot_descriptions.py # PubMedBERT embedding
│ ├── extract_prott5_embeddings.py # ProtT5 sequence embeddings
│ ├── extract_esm_if1_embeddings.py # ESM-IF1 structure embeddings
│ ├── extract_ppi_embeddings.py # STRING PPI embeddings
│ ├── extract_historical_uniprot_text.py # Historical UniSave extraction
│ ├── check_alphafold_coverage.py # AlphaFold PDB download
│ └── prepare_cafa3_data.py # CAFA3 benchmark splits
├── train.py # Main training script
├── requirements.txt
└── README.md

Citation

If you use this code in your research, please cite:

@article{mmfp2025,
title={Hybrid Gated Fusion: A Multimodal Deep Learning Framework for Protein Functional Annotation},
author={Zijian Zhou and Daniel W. A. Buchan},
journal={TODO},
year={2025}
}

License

MIT License