A curated multi-endpoint drug-protein bioactivity dataset with construct-aware protein targets

Lina Abou-Abbas; Nour Mhanna; Hasan Bazzi; Sleiman El Hajj; Abdulkader Helwan; Rabah Omrani; Khadidja Henni

cs.LGOct 23, 2026

A curated multi-endpoint drug-protein bioactivity dataset with construct-aware protein targets

Authors: Lina Abou-Abbas, Nour Mhanna, Hasan Bazzi, Sleiman El Hajj, Abdulkader Helwan, Rabah Omrani, Khadidja Henni

Organizations: Department of Electrical and Computer Engineering, Lebanese American University, Byblos, Lebanon · Institute of Applied Artificial Intelligence, T´ELUQ University, Montreal, Canada

Abstract

Public drug-protein bioactivity resources commonly identify targets by UniProt accession, although one accession may represent distinct experimental sequences. We present Curate-DTI, a construct-aware dataset derived from the 1 July 2026 BindingDB export, with IC50, Ki, Kd and EC50 processed independently. The release contains 2,919,717 measurements, 1,291,872 standardized ligands, 9,668 sequence-defined protein entities and 2,275,683 endpoint-specific interactions spanning 8,336 reviewed UniProt accessions. Protein entities combine the accession with the reported experimental sequence, preserving constructs that accession-only processing would merge. Across the four endpoints, 4,204 endpoint-specific ligand-accession groups showed different classification labels across sequence-defined protein entities. Exact and inequality-qualified observations provide relation-aware classification evidence, whereas regression targets use exact values only. Connected Parquet tables preserve individual measurements, standardized structures, protein sequences, aggregated interactions and available assay and source provenance. The resource supports endpoint-specific classification and regression while allowing users to select molecular representations, protein representations and evaluation splits. Preprocessing code, a column dictionary, exclusion reports and validation summaries accompany the release.

Figures & tables

Figure 1: BindingDB preprocessing workflow for the four activity endpoints: IC 50 , $K_{i}$ , $K_{d}$ , and EC 50 . Measurements for each endpoint were processed independently. Single-chain records underwent activity-value validation, reviewed Swiss-Prot accession checking, experimental sequence validation, and ligand structure standardization. The resulting construct-aware measurements were then used to build shared ligand and protein-entity tables together with endpoint-specific interaction datasets. Multi-chain measurements were retained separately, along with the available chain-level information.

Processing stage	IC 50	$K_{i}$	$K_{d}$	EC 50
Endpoint reported	2,204,608	626,880	130,072	269,888
Exactly one target chain	2,061,581	597,553	126,972	259,974
SMILES and primary accession reported	2,008,206	584,594	125,658	251,885
Positive finite value or boundary	2,008,054	584,565	125,625	251,850
Exact $pE$ restricted to $[3,12]$	2,002,416	580,881	124,729	251,246
Accession present in reviewed Swiss-Prot	1,999,681	580,431	124,699	250,562

Table 1: Endpoint-specific attrition during construction of the curated single-chain measurement tables. Values indicate the number of measurements retained after each sequential processing stage.

Endpoint	Measurements	Exact ( $=$ )	Right-censored ( $>$ )	Left-censored ( $<$ )
IC 50	1,982,041	1,561,575	254,428	166,038
$K_{i}$	565,541	489,745	57,603	18,193
$K_{d}$	124,325	77,589	39,357	7,379
EC 50	247,810	196,715	40,772	10,323
Total	2,919,717	2,325,624	392,160	201,933

Table 2: Relations among the retained endpoint-specific measurements. Right-censored measurements report a lower boundary, whereas left-censored measurements report an upper boundary.

Relation	Boundary	Classification evidence
$=$	$\leq 1{,}000$ nM	Class 1
$=$	$>1{,}000$ nM	Class 0
$<$	$\leq 1{,}000$ nM	Class 1
$<$	$>1{,}000$ nM	Ambiguous
$>$	$\geq 1{,}000$ nM	Class 0
$>$	$<1{,}000$ nM	Ambiguous

Table 3: Measurement-level classification rules at the operational threshold of 1,000 nM.

Quantity	Count
Retained measurements	2,919,717
Endpoint-specific interactions	2,275,683
Distinct drug–protein pairs across endpoints	2,193,251
Standardized ligands	1,291,872
Sequence-defined protein entities	9,668
Reviewed UniProt accessions	8,336

Table 4: Overview of the multi-endpoint construct-aware dataset.

Organism	Protein entities	Percentage (%)
Homo sapiens	6,050	62.58
Mus musculus	869	8.99
Rattus norvegicus	841	8.70
Bos taurus	201	2.08
Escherichia coli strain K12	137	1.42
Mycobacterium tuberculosis H37Rv	100	1.03

Table 5: Most represented organism labels among the sequence-defined protein entities. Counts describe protein entities rather than measurements or interactions.

Figure 2: Organism coverage of the endpoint-specific interaction records. The twelve most represented organism labels are shown. Counts combine the IC 50 ,

K_{i}

K_{d}

and EC 50 interaction tables; therefore, a drug–protein pair reported for more than one endpoint contributes one interaction record to each corresponding endpoint.

Source-group type	Measurements	Unique groups	Measurements (%)
Article DOI	1,420,613	38,263	48.66
Patent	1,228,477	8,218	42.08
PMID	169,782	4,787	5.82
BindingDB entry DOI	100,072	1,072	3.43
PMID associated with multiple DOI values	68	1	$<0.01$
Release-specific fallback	705	705	0.02

Table 6: Source provenance of the retained measurement records. Percentages are rounded independently and may not sum exactly to 100%.

Figure 3: Relationships among the four principal record types in the model-ready release. Measurement and interaction records reference the shared ligand and protein-entity tables through Drug_ID and Protein_Entity_ID . Repeated measurements are aggregated separately within each endpoint to produce endpoint-specific interaction records.

Data table	Rows	Primary key	Purpose
Measurements	2,919,717	Measurement_ID	Individual retained single-chain IC 50 , $K_{i}$ , $K_{d}$ and EC 50 observations, including reported relations, task evidence and available assay and source provenance.
Standardized ligands	1,291,872	Drug_ID	One record for each standardized ligand structure, including canonical chemical representations, calculated properties and available external identifiers.
Protein entities	9,668	Protein_Entity_ID	One record for each sequence-defined protein entity, including experimental and canonical sequences, reviewed Swiss-Prot annotations and construct characterization.
Endpoint-specific interactions	2,275,683	Interaction_ID	One aggregated record for each endpoint, ligand and protein-entity combination, including classification and regression task fields and quality indicators.

Table 7: Primary model-ready tables included in the data release.

Endpoint	Interaction file	Classification target	Regression target
IC 50	BindingDB_ic50_ interactions.parquet	y_class	pIC50
$K_{i}$	BindingDB_ki_ interactions.parquet	y_class	pKi
$K_{d}$	BindingDB_kd_ interactions.parquet	y_class	pKd
EC 50	BindingDB_ec50_ interactions.parquet	y_class	pEC50

Table 8: Endpoint-specific interaction files and prediction targets.

Endpoint	Valid classification	Class 1	Class 0	Conflicts	Valid regression	Dispersed regression
IC 50	1,500,552	915,116	585,436	27,532	1,220,359	12,104
$K_{i}$	433,755	302,979	130,776	5,363	382,935	2,636
$K_{d}$	83,604	39,912	43,692	2,176	56,580	2,112
EC 50	187,772	104,110	83,662	3,616	154,209	1,396
Total	2,205,683	1,362,117	843,566	38,687	1,814,083	18,248

Table 9: Validation summary for the endpoint-specific classification and regression task views.

Protein-entity category	Entities
Exact canonical sequence	8,191
Exact canonical fragment	548
Substitution variant	417
Fragment with variants	328
Insertion/deletion variant	139
Canonical sequence with terminal extension	24

Table 10: Characterization of the retained sequence-defined protein entities.

IC 50	1,548,382	1,492,055	27,532	30,692	3,681
	Interactions		Classification conflicts		Construct-dependent groups
Endpoint	Construct	Accession	Construct	Accession
$K_{i}$	445,298	441,518	5,363	5,598	244
$K_{d}$	86,858	84,379	2,176	2,387	258
EC 50	195,145	194,390	3,616	3,635	21
Total	2,275,683	2,212,342	38,687	42,312	4,204

Table 11: Comparison of construct-aware and UniProt-accession-level aggregation. Dependent groups are endpoint-specific drug–accession groups whose classification outcomes vary among experimental protein sequences.

Drug Discovery

Drug discovery identifies, designs, and tests molecules that can safely affect disease biology.

See all 164 papers

BERT4DTI : BERT-based Model for Predicting Drug-Protein Interactions
Sep 27, 2026Thanina Hamitouch, Khadidja Henni, Abdelkrim Arie +3
Ecole Nationale Supérieure d’Informatique (ESI) Algiers, Algeria · Institut d’Intelligence Artificielle Appliquée, TELUQ University Montreal, Canada · Department of Electrical and Computer Engineering Lebanese American University Byblos, Lebanon
Beyond Manual Curation: Augmenting Targeted Protein Degradation Databases via Agentic Literature Extraction Workflows
May 11, 2026Yaochen Rao, Farzaneh Jalalypour, N. M. Anoop Krishnan +1Training Data Curation Document Information Extraction
Department of Computer Science and Engineering, Chalmers University of Technology and University of Gothenburg, Gothenburg, SE · Yardi School of Artificial Intelligence, Indian Institute of Technology Delhi, New Delhi, IN
BioMiner: A Multi-modal System for Automated Mining of Protein-Ligand Bioactivity Data from Literature
Apr 23, 2026Jiaxian Yan, Jintao Zhu, Yuhang Yang +8Document Information Extraction Multimodal Large Language Models
State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China · Center for Quantitative Biology, Academy for Advanced Interdisciplinary Studies, Peking University · Princeton University +2

Sep 27, 2026J/K move · Enter open · S save

BERT4DTI : BERT-based Model for Predicting Drug-Protein Interactions

Thanina Hamitouch, Khadidja Henni, Abdelkrim Arie, Amina Selma Haichour+2

Ecole Nationale Supérieure d’Informatique (ESI) Algiers, Algeria · Institut d’Intelligence Artificielle Appliquée, TELUQ University Montreal, Canada · Department of Electrical and Computer Engineering Lebanese American University Byblos, Lebanon

Open paper AskPDF arXiv

Understanding how drugs interact with protein targets is fundamental to drug discovery, drug repurposing and the early identification of promising therapeutic candidates before costly experimental testing. Sequence-based DTI models face three practical limitations: labelled interactions are scarce and unevenly distributed, large pretrained chemical and protein encoders are expensive to fine-tune end-to-end, and independently encoded sequences do not capture pair-specific dependencies. We present BERT4DTI, which encodes SMILES strings with ChemBERTa and amino-acid sequences with ProtBERT, applies bidirectional mutual attention between token-level representations, and classifies the resulting interaction features using convolutional layers and a multilayer perceptron. To reduce trainable size, ProtBERT is truncated to 18 retained layers and only the last two layers of each encoder are fine-tuned. On BIOSNAP, DAVIS and BindingDB, BERT4DTI is competitive, achieving the best ROC-AUC and PR-AUC on BIOSNAP and the highest sensitivity on all three benchmarks. An ablation on DAVIS shows that mutual attention improves PR-AUC and specificity. With 125M trainable parameters compared with 353M for full BERT fine-tuning, BERT4DTI provides a favourable performance-parameter trade-off for sequence-based DTI screening, while leaving runtime profiling, calibration and leakage-audited validation for future work.

Abstract

Drug Discovery

Explore similar work

BERT4DTI : BERT-based Model for Predicting Drug-Protein Interactions

Beyond Manual Curation: Augmenting Targeted Protein Degradation Databases via Agentic Literature Extraction Workflows

BioMiner: A Multi-modal System for Automated Mining of Protein-Ligand Bioactivity Data from Literature