Back

HARVEST: Unlocking the Dark Bioactivity Data of Pharmaceutical Patents via Agentic AI

Shepard, V.; Musin, A.; Chebykina, K.; Zeninskaya, N. A.; Mistryukova, L.; Avchaciov, K.; Fedichev, P. O.

2026-03-18 bioinformatics
10.64898/2026.03.15.711910 bioRxiv
Show abstract

Pharmaceutical patents contain vast Structure-Activity Relationship tables documenting protein- ligand binding data that are technically public yet computationally inaccessible, rendering this wealth of data effectively dark -- trapped in unstructured archives no existing database has systematically captured. We present HARVEST, a multi-agent large language model pipeline that autonomously extracts structured bioactivity records from USPTO patent archives at $0.11 per document. Applied to 164,877 patents, HARVEST produced 3.36 million activity records, recovering 365,713 unique scaffolds and 1,108 protein targets absent from BindingDB -- completing in under a week a task requiring over 55 years of continuous expert labor. Automated extraction achieves 91% agreement with human curators while exhibiting lower unit-conversion error rates. We further introduce H-Bench, a structurally guaranteed held-out benchmark built from this recovered data. Evaluation of the leading open-source model Boltz-2 on H-Bench reveals a two-dimensional generalization gap: performance degrades both on novel chemical scaffolds and on uncharacterized protein targets, exposing fundamental limitations of models trained on existing public repositories.

Matching journals

The top 8 journals account for 50% of the predicted probability mass.

1
Journal of Chemical Information and Modeling
● 238 papers in training set
Top 0.4%
12.1%
Similar papers in this journal
2
Bioinformatics
◐ 1204 papers in training set
Top 3%
8.7%
Similar papers in this journal
3
Nature Communications
○ 5641 papers in training set
Top 28%
5.3%
Similar papers in this journal
4
Patterns
○ 78 papers in training set
Top 0.2%
5.3%
Similar papers in this journal
5
Journal of Cheminformatics
○ 29 papers in training set
Top 0.1%
5.3%
Similar papers in this journal
6
Proceedings of the National Academy of Sciences
● 2444 papers in training set
Top 11%
4.7%
7
Briefings in Bioinformatics
◐ 354 papers in training set
Top 2%
4.7%
8
Nature Methods
○ 385 papers in training set
Top 2%
4.2%
50% of probability mass above
9
Cell Systems
○ 201 papers in training set
Top 1%
3.9%
10
Artificial Intelligence in the Life Sciences
○ 13 papers in training set
Top 0.1%
3.9%
11
Computational and Structural Biotechnology Journal
● 242 papers in training set
Top 1%
3.3%
12
PLOS Computational Biology
● 1863 papers in training set
Top 9%
3.3%
13
Communications Chemistry
○ 48 papers in training set
Top 0.2%
2.6%
14
Bioinformatics Advances
◐ 203 papers in training set
Top 2%
2.4%
15
PLOS ONE
● 5266 papers in training set
Top 42%
2.3%
16
mAbs
○ 32 papers in training set
Top 0.2%
2.3%
17
Scientific Reports
○ 3612 papers in training set
Top 54%
1.7%
18
iScience
○ 1154 papers in training set
Top 20%
1.6%
19
Journal of Molecular Biology
○ 232 papers in training set
Top 2%
1.5%
20
Chemical Science
● 73 papers in training set
Top 1%
1.5%
21
Nucleic Acids Research
◐ 1281 papers in training set
Top 11%
1.1%
22
Advanced Science
○ 286 papers in training set
Top 7%
1.1%
23
GigaScience
◐ 212 papers in training set
Top 4%
1.0%
24
eLife
● 5828 papers in training set
Top 64%
0.9%
25
Cell Genomics
○ 172 papers in training set
Top 5%
0.8%
26
Nature Biotechnology
○ 172 papers in training set
Top 5%
0.6%
27
Database
◐ 61 papers in training set
Top 1%
0.6%
28
Molecules
○ 39 papers in training set
Top 2%
0.6%

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.