DeepPurpose is a deep learning library for drug-target interaction (DTI) prediction, published in Bioinformatics. This guide covers data preparation for training DTI models.

Input format

Create a CSV with SMILES, FASTA targets, and labels:

smiles,protein,affinity
CC(=O)Oc1ccccc1C(=O)O,>sp|P00533|EGFR_HUMAN...SEQUENCE...,8.1
CC1=CC=CC=C1,>sp|P00533|EGFR_HUMAN...SEQUENCE...,6.4

Train/val/test splits

import pandas as pd
from sklearn.model_selection import train_test_split

df = pd.read_csv("inputs/data.csv")
train, temp = train_test_split(df, test_size=0.3, random_state=42)
val, test = train_test_split(temp, test_size=0.5, random_state=42)

train.to_csv("inputs/train.csv", index=False)
val.to_csv("inputs/val.csv", index=False)
test.to_csv("inputs/test.csv", index=False)

Choosing encoders

DeepPurpose provides 15+ compound encoders and 7+ protein encoders that can be combined into 50+ neural architectures.

Compound (drug) encoders

Encoder	Type	Description
`Morgan`	Fingerprint	Morgan/ECFP circular fingerprints
`Pubchem`	Fingerprint	PubChem substructure fingerprints
`Daylight`	Fingerprint	Daylight-type fingerprints
`rdkit_2d_normalized`	Descriptor	RDKit 2D molecular descriptors
`CNN`	Deep Learning	1D CNN on SMILES strings
`CNN_RNN`	Deep Learning	CNN + RNN hybrid on SMILES
`RNN`	Deep Learning	RNN/LSTM on SMILES sequences
`Transformer`	Deep Learning	Transformer encoder on SMILES
`MPNN`	GNN	Message Passing Neural Network on molecular graphs
`DGL_GCN`	GNN	Graph Convolutional Network (DGL)
`DGL_NeuralFP`	GNN	Neural fingerprints via DGL
`DGL_AttentiveFP`	GNN	Attentive fingerprints via DGL
`DGL_GIN_AttrMasking`	GNN	GIN with attribute masking (pretrained)
`DGL_GIN_ContextPred`	GNN	GIN with context prediction (pretrained)

Protein (target) encoders

Encoder	Type	Description
`AAC`	Descriptor	Amino Acid Composition
`PseudoAAC`	Descriptor	Pseudo Amino Acid Composition
`Conjoint_triad`	Descriptor	Conjoint triad features
`Quasi-seq`	Descriptor	Quasi-sequence order descriptors
`CNN`	Deep Learning	1D CNN on amino acid sequences
`CNN_RNN`	Deep Learning	CNN + RNN hybrid
`RNN`	Deep Learning	RNN/LSTM on protein sequences
`Transformer`	Deep Learning	Transformer encoder on sequences

Encoder selection tips

Small datasets (<1000 samples): Use fingerprint encoders (Morgan, Pubchem) with descriptor-based protein encoders (AAC, PseudoAAC)
Medium datasets: CNN-based encoders offer a good balance of performance and speed
Large datasets: Transformer and GNN encoders can capture more complex patterns
Pretrained options: DGL_GIN_AttrMasking and DGL_GIN_ContextPred leverage pretrained molecular representations

Featurize with DeepPurpose utils

DeepPurpose provides built-in dataset utilities and benchmark datasets:

from DeepPurpose import dataset

# Load a benchmark dataset (e.g., DAVIS, KIBA, BindingDB)
X_drugs, X_targets, y = dataset.load_process_DAVIS(
    path='./data',
    binary=False,  # Regression task
    convert_to_log=True,
    threshold=30
)

# Or process your own data
X_drug, X_target, y = dataset.data_process(
    X_drug=drug_smiles_list,
    X_target=protein_sequence_list,
    y=affinity_labels,
    drug_encoding="Morgan",
    target_encoding="CNN",
    split_method="random",
    frac=[0.7, 0.1, 0.2],
)

Available benchmark datasets

DeepPurpose includes loaders for common DTI benchmark datasets:

DAVIS: Kinase inhibitor binding affinities (Kd values)
KIBA: Kinase Inhibitor BioActivity dataset
BindingDB: Large-scale binding affinity database
DrugBank: Drug-target associations
Tox21: Toxicity prediction dataset
HIV: HIV replication inhibition dataset

Data quality tips

Deduplicate SMILES/targets before splitting.
Validate FASTA headers and ensure sequences contain only valid amino acids.
Standardize SMILES (e.g., RDKit canonicalization) before training to reduce variance.
For binding affinity data, consider log transformation (pKd, pKi) for better model convergence.
Remove invalid SMILES using RDKit's MolFromSmiles() validation.

Split strategies

DeepPurpose supports multiple splitting strategies for robust evaluation:

# Random split (default)
train, val, test = dataset.data_process(..., split_method='random')

# Cold drug split (test on unseen drugs)
train, val, test = dataset.data_process(..., split_method='cold_drug')

# Cold target split (test on unseen proteins)
train, val, test = dataset.data_process(..., split_method='cold_protein')

Cold splits are more challenging but better reflect real-world drug discovery scenarios where you predict interactions for novel compounds or targets.

Data prep & featurization