GRAM-DTI Data

Pre-training and downstream evaluation data for GRAM-DTI: Adaptive Multimodal Representation Learning for Drug-Target Interaction Prediction (ICLR 2026).

Paper: https://arxiv.org/abs/2509.21971

Code: https://github.com/uta-smile/GRAM-DTI

Contents

File	Size	Description
embeddings/embeddings.h5	256 MB	Pre-computed embeddings for the four modalities, 50,968 rows each
embeddings/metadata.json	168 MB	The raw text and labels for the same 50,968 pairs, in the same row order
cleaned_pretrain_after_rdk.json	146 MB	The source records the embeddings were computed from, keyed by PubChem CID
downstream/	65 MB	The four downstream evaluation datasets, and make_splits.py to generate their cross-validation folds

The pre-training data covers 50,968 molecule–protein pairs, built from 6,545 molecules and 4,418 protein sequences.

embeddings/embeddings.h5

Dataset	Shape	Encoder
smiles_embeddings	(50968, 768)	ibm/MoLFormer-XL-both-10pct
text_embeddings	(50968, 768)	laituan245/molt5-base (encoder), on the molecule description
hta_embeddings	(50968, 768)	laituan245/molt5-base (encoder), on the hierarchical taxonomic annotation
protein_embeddings	(50968, 1280)	facebook/esm2_t33_650M_UR50D

All arrays are float32. Each embedding is the first-token hidden state of the encoder's last layer. Inputs were truncated to 512 tokens (1,024 for proteins).

embeddings/metadata.json

A list of 50,968 objects. Row i matches row i of every array in embeddings.h5.

Field	Description
CID	PubChem compound ID
smiles	SMILES string
description	Text description of the molecule
hta	Hierarchical taxonomic annotation text
protein_sequence	Target protein amino-acid sequence
protein_name, organism, function_description	Target protein annotations
ic50_value	Mean IC50 in nM, or null
ic50_label	0 if IC50 < 10 nM, 1 if < 1000 nM, 2 otherwise; null when no IC50 is available
has_ic50	Whether the pair has a measured IC50

16,035 pairs have a measured IC50 (1,529 / 4,396 / 10,110 in classes 0 / 1 / 2). The remaining 34,933 pairs have no IC50 label.

cleaned_pretrain_after_rdk.json

A dictionary keyed by CID with 6,546 entries. Each entry holds:

• smiles_info: SMILES, Description, HTA
• bindingdb_info: targets, a list of protein targets, each with target_name, organism, sequence, uniprot_ids, function_description and binding_data
• smiles_validation: the result of matching the SMILES against BindingDB

downstream/

Four evaluation datasets, one directory each:

Directory	Task	Interactions	Drugs	Proteins
downstream/yamanishi_08/	DTI prediction	5,127	791	989
downstream/Hetionet/	DTI prediction	49,942	1,440	18,935
downstream/activation/	Mechanism of action	1,913	1,426	281
downstream/inhibition/	Mechanism of action	21,055	14,049	1,088

Each directory holds four files:

File	Description
dti.csv	Positive drug–target pairs, by ID (tab-separated)
drug_smiles.csv (drug_smi.csv for the MoA sets)	Drug ID to SMILES (tab-separated)
protein_seq.csv (tar_seq.csv for the MoA sets)	Protein ID to amino-acid sequence (tab-separated)
retrieval.csv	The positive pairs with SMILES and sequence filled in, used by the retrieval task (comma-separated)

The ID, SMILES and sequence files are unchanged copies of the data released with DTIAM. retrieval.csv was derived from them for this work.

Cross-validation folds are not included. downstream/make_splits.py generates them from the three ID, SMILES and sequence files, for three settings:

Setting	How the pairs are split
split_warm	At random
split_drug_cold	By drug, so test drugs are unseen in training
split_protein_cold	By protein, so test proteins are unseen in training

Negative pairs are sampled at a 10:1 ratio to the positives, from the unobserved pairs between the drugs and proteins listed in the files. The seed is fixed (42), so the script writes the same folds every time. The paper uses 10 folds for the DTI datasets and 5 for the MoA datasets.

Usage

Download the pre-training data into the directory the training script expects:

hf download FengJiang02/GRAM-DTI --repo-type dataset \
  --include "embeddings/*" "cleaned_pretrain_after_rdk.json" --local-dir pretrain_data

Then run pre-training from the directory that contains pretrain_data/:

python pretraining/run.py --embedding_dir pretrain_data/embeddings

To read the embeddings directly:

import h5py, json

with h5py.File("pretrain_data/embeddings/embeddings.h5", "r") as f:
    smiles = f["smiles_embeddings"][:]
    protein = f["protein_embeddings"][:]

with open("pretrain_data/embeddings/metadata.json") as f:
    metadata = json.load(f)

The downstream scripts read from data/<dataset>/:

hf download FengJiang02/GRAM-DTI --repo-type dataset --include "downstream/*" --local-dir .
mv downstream data

Then generate the folds, which are written to data/<dataset>/data_folds/<setting>/:

python data/make_splits.py data/yamanishi_08 data/Hetionet
python data/make_splits.py data/activation data/inhibition --n_splits 5

The script needs numpy, pandas and scikit-learn. The folds store the full SMILES and sequence on every row, so they are large: about 11 GB for Hetionet and 15 GB for all four datasets.

Sources

The molecules, with their descriptions and hierarchical taxonomic annotations, come from the TRIDENT dataset of 47,269 PubChem molecules. Protein targets and IC50 measurements come from BindingDB (the May 2025 release, BindingDB_All_202505).

To prevent leakage, every (SMILES, protein) pair that appears as a positive in a downstream evaluation set (Yamanishi_08, Hetionet, and the activation and inhibition MoA sets) was removed before pre-training. Individual molecules and proteins from those sets can still appear in the pre-training data, in other pairs.

The downstream datasets come from DTIAM; please cite it as well if you use them.

Citation

If you use this data, please cite:

@article{jiang2025gramdti,
  title={GRAM-DTI: adaptive multimodal representation learning for drug target interaction prediction},
  author={Jiang, Feng and Mollaysa, Amina and Ma, Hehuan and Mansi, Tommaso and Huang, Junzhou and Prakash, Mangal and Liao, Rui},
  journal={arXiv preprint arXiv:2509.21971},
  year={2025}
}