Pick one of five DeepPurpose task types, a workflow preset and a drug and protein encoder pair, then run a model-zoo inference and export predictions.csv.

DeepPurpose is a comprehensive deep learning library for drug-target interaction (DTI) prediction, published in Bioinformatics. It implements 15+ compound encoders and 7+ protein encoders with over 50 combined neural architectures.

Webapp analysis tasks

The .bloom* webapp provides five analysis task types:

Task Type	Description	Required Inputs
Drug-Target Interaction	Predict binding affinity between compounds and proteins	Drug (SMILES), Target (protein sequence)
Drug Property Prediction	Predict ADMET and molecular properties	Drug compounds (one per line), Properties to predict
Protein-Protein Interaction	Predict protein binding interactions	Two protein sequences
Drug-Drug Interaction	Predict drug combination effects	Two drug compounds (SMILES)
Protein Function Prediction	Classify protein function	Protein sequence, Function class

Input source selection

For DTI analysis, select input sources for each field:

Input Type	Description
Manual Entry	Enter SMILES or protein sequences directly
From Previous Job	Import compounds/sequences from completed workflows

Compatible upstream sources: AiZynthFinder, REINVENT, ZINC, ESMFold outputs.

Research workflow presets

Click preset buttons to configure optimal settings for specific workflows:

Preset	Task Type	Encodings	Use Case
Virtual Screening	DTI	Morgan + AAC	High-throughput library screening (~30s per 1000 compounds)
Lead Optimization	Drug Property	Transformer/CNN	SAR analysis, multi-parameter optimization (~2-5 min/compound)
Target Discovery	DTI	CNN + CNN	Drug repurposing, target deconvolution (~1-2 min/compound)
ADMET Profiling	Drug Property	Morgan	Safety screening, toxicity prediction (~15-30s/compound)

Encoder options

Drug encodings: CNN, CNN_RNN, Transformer, MPNN, Morgan, Pubchem, Daylight, rdkit_2d, ESPF, DGL_GCN, DGL_NeuralFP, DGL_GIN_AttrMasking, DGL_GIN_ContextPred, DGL_AttentiveFP

Protein encodings: CNN, CNN_RNN, Transformer, AAC, PseudoAAC, Conjoint_triad, Quasi-seq-order, ConjointTriadVec, CTD

DTI prediction modes

Mode	Description
Single Prediction	Analyze one drug-target pair
Drug Repurposing	Find new targets for existing drugs
Virtual Screening	Large-scale compound evaluation

Advanced options (expandable panel)

Parameter	Range	Default	Description
Pretrained Model	Model zoo selection	None	Use pretrained weights or train from scratch
Job Title	Text	—	Optional job description
Batch Size	1 - 512	32	Processing batch size
Output Format	json, csv, tsv	json	Results file format
Use CUDA	Toggle	Enabled	GPU acceleration

Available at https://github.com/kexinhuang12345/DeepPurpose

Install

Use CUDA-matched torch. Example for CUDA 12.1:

pip install "torch==2.3.1" --index-url https://download.pytorch.org/whl/cu121
pip install "deeppurpose>=0.2.9" pandas scikit-learn

Check GPU access:

python - <<'PY'
import torch
print("CUDA available:", torch.cuda.is_available())
print("Device count:", torch.cuda.device_count())
PY

Run a model zoo inference

from deeppurpose import dataset, encoder_model, models

drug, target, y = dataset.load_process_DAVIS('./inputs', binary=False, device='cuda')

model = models.model_zoo("Davis_BindingDB_KD_GNN", device='cuda')
preds = model.predict(drug[:10], target[:10])
print("Predictions shape:", preds.shape)
print("First score:", preds[0])

Minimal job layout

deeppurpose-job/
├─ inputs/
│  └─ davis/                # cached dataset
├─ outputs/
│  └─ predictions.csv
└─ scripts/
   └─ run_inference.py

Export predictions

import pandas as pd

pd.DataFrame({"score": preds}).to_csv("outputs/predictions.csv", index=False)

DeepPurpose quickstart