Multi-Modal Fusion-based Efficient Model for Enhanced Generalization in Drug-Target Interaction Prediction

In the pursuit of accelerated and cost-effective drug discovery, accurately predicting Drug-Target Interactions (DTIs) remains a crucial challenge. Advanced computational approaches can significantly reduce the time and financial burden of traditional drug discovery methods. This study proposes MMF-DTI, a Multi- Modal Fusion-based model designed to enhance the generalization capabilities in DTI predictions. By integrating multi-modal data, including drug and target sequences, molecular properties, and target function descriptions, MMF-DTI demonstrates marked improvements in predicting interactions for both new drugs and targets. Our model incorporates a novel fusion strategy that combines features extracted from various data modalities through multi-modality encoders, including sequence encoders for SMILES and FASTA formats and additional encoders for molecular properties and target function descriptions. The efficacy of MMF-DTI is validated against benchmark datasets, showcasing its superior performance in generalization, particularly in cold-drug and cold-target scenarios. Furthermore, the model maintains efficient resource usage, enabling its operation even in constrained computational environments. MMF-DTI’s approach not only advances the field of DTI prediction but also underscores the potential of multimodal data fusion in enhancing model generalization, offering promising avenues for future drug discovery efforts.

• Binary classification: DAVIS, Binding DB, and BIOSNAP

• Downloaded from MolTrans

• Preprocessed UniProt is avaiable at [HuggingFace space](https://huggingface.co/datasets/jonghyunlee/UniProt_functional text descriptions)

• You can install the required libraries by running pip install -r requirements.txt
• If you encounter any installation errors, please don't hesitate to reach out to us for assistance.

• You can run experiments using config files (under config folder). For example you can launch an experiment by running python run.py -c config/BIOSNAP_Labmda-True_Text-False_Prop-False_Missing-0_Unseen-No.yaml -s 42 or python run.py --config config/BIOSNAP_Labmda-True_Text-False_Prop-False_Missing-0_Unseen-No.yaml --seed 42
• Hyperparameters are recorded in config file, therefore, you can modify experiments easily.

dataset: 
    name: "BIOSNAP" # Options: DAVIS, BindingDB, BIOSNAP
    missing: 0 # Options: 70, 80, 90, 95 (Only for BIOSNAP)
    unseen: "drug" # Options: No, drug, target
    use_sampler: False # Options: True, False
    use_enumeration: False # Options: True, False
    
prot_length: 
    teacher: 545
    student: 545

lambda:
    learnable: True
    fixed_value: -1

prot_encoder:
    hidden_size: 1024
    num_hidden_layers: 2
    num_attention_heads: 16
    intermediate_size: 4096
    hidden_act: "gelu"

multimodality:
    use_text_feat: False # Options: True, False
    use_property: False # Options: True, False

training_config:
    batch_size: 32
    num_workers: 16
    epochs: 30
    hidden_dim: 1024
    learning_rate: 0.0001
    device: 0
    seed: -1