🧬 DrugCMF: Confidence-aware Multimodal Fusion for Drug-Target Interaction Prediction

🚀 Official implementation of DrugCMF, accepted at AAAI 2026 🎉

---

📰 News

• [2026-01] 🎉 DrugCMF accepted at AAAI 2026!
• [2026-05] 🚀 Code and datasets released.

---

🔬 Overview

💊 Drug-target interaction (DTI) prediction is a fundamental task in drug discovery. 1D sequences 🧬 and 3D structures 🏗️ capture complementary information but vary in reliability across different drug–protein pairs 🧑‍🔬. DrugCMF addresses this modality reliability gap through a confidence-aware multimodal fusion 🔗 framework.

---

🏆 Key Contributions

1. 🧩 Quadruple-Modal Representations — Each DTI pair is represented by four views: MolFormer (1D drug) + ProtT5 (1D protein) as the textual modality, and UniMol (3D drug) + SaProt (3D protein) as the structural modality.
2. 🔄 Bidirectional Cross-Attention — Within each modality, drug and protein representations attend to each other before fusion, combined with residual connections and average/max pooling.
3. 🎯 Trusted-Confidence Prediction (TCP) — A learned confidence score per modality that adaptively weights each modality's contribution. The TCP loss aligns confidence scores with softmax probabilities via MSE, mitigating the impact of unreliable representations.
4. 🧠 Cross-Modal Transformer & Attentional Pooling — Modality features are stacked and refined through a Transformer encoder, then aggregated via a learnable multi-head attentional pooler before final classification.

---

📊 Experimental Results

🏠 In-domain Performance

🌍 Cross-domain Performance

---

⚙️ Implementations

📦 Datasets

Datasets can be downloaded from:

• ☁️ Google Drive
• 💿 Baidu Netdisk

✂️ Splits:

• 🎲 Random — Random train/val/test split
• ❄️ Cold — Cold-start setting (novel drugs/proteins in test, Human only)
• 🔷 Cluster — Cluster-based split for domain generalization (BindingDB / BioSNAP only)

Each split CSV contains 6 columns (SMILES, Protein, Y, Structure, SMILES_id, Protein_id). Cluster splits additionally include drug_cluster and target_cluster columns.

---

🤖 Pre-trained Models

Model weights are downloaded automatically from HuggingFace 🤗:

---

⏳ Preprocessing

Generate all four embeddings offline before training 🏗️:

python preprocess/get_embedding.py \
    --dataset bindingdb \
    --data_path datasets/bindingdb/processed_full.csv \
    --save_path embeddings/bindingdb \
    --device cuda:0 \
    --seed 42

Or generate for all three datasets at once ⚡:

bash preprocess/run_embedding.sh

This produces .pt files in embeddings/{dataset}/{type}_embeddings/ containing seq_emb and mask tensors 📁.

---

📋 Requirements

All dependencies are listed in requirements.txt 📄. Key packages:

🚂 Training

Train on the Human dataset with a cold-start split:

python main.py --dataset human --split cold --device cuda:0

Run with multiple random seeds 🌱:

bash train.sh

---

📁 Repository Structure

├── main.py                  # Training & evaluation entry point
├── model.py                 # DrugCMF model architecture
├── modules.py               # Building blocks (ModalityInteractionLearner, Classifier, MultiModalAttentionalPooler)
├── attention.py             # BidirectionalCrossAttention, AttentionalPooler
├── dataloader.py            # DTIDataset & DataLoader (loads pre-computed embeddings)
├── train_util.py            # Training & evaluation loops (AdamW, ReduceLROnPlateau, early stopping)
├── utils.py                 # Reproducibility seed, collate function, checkpoint I/O, logger
├── requirements.txt         # Full Python dependencies
├── train.sh                 # Multi-seed training script (11 seeds)
│
├── preprocess/              # Offline embedding generation
│   ├── get_embedding.py     # Generate all 4 embeddings for a dataset
│   └── run_embedding.sh     # SLURM batch script for all datasets
│
├── unimol/                  # UniMol implementation (based on DP Technology)
│   ├── unimol.py
│   ├── unimol_utils.py
│   └── transformers.py
│
└── datasets/                # Preprocessed CSV splits
    ├── README.md
    ├── human/
    │   ├── processed_full.csv
    │   ├── cold/   {train,val,test}.csv
    │   └── random/ {train,val,test}.csv
    ├── bindingdb/
    │   ├── processed_full.csv
    │   ├── random/ ...csv
    │   └── cluster/ {source_train,target_test}.csv
    └── biosnap/
        ├── processed_full.csv
        ├── random/ ...csv
        └── cluster/ {source_train,target_test}.csv

---

📖 Citation

If you find DrugCMF useful, please cite our paper 🙏:

@inproceedings{yang2026bridging,
  title={Bridging the Modality Reliability Gap in Drug-Target Interaction Prediction via a Confidence-aware Multimodal Fusion Framework},
  author={Yang, Jie and Zhang, Junxiong and Qian, Kun and Yang, Qingyu and Li, Weikai and Cheng, Zhen},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={40},
  number={32},
  pages={27529--27537},
  year={2026}
}

---

⚖️ License

• 🧑‍💻 All code except unimol/: CC BY-NC 4.0 — Creative Commons Attribution-NonCommercial 4.0 International. Sharing and adaptation for non-commercial purposes only, with attribution.
• 🧪 unimol/ (from DP Technology / Uni-Core): MIT License, as originally licensed.