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

HAR VEST: Unlocking the Dark Bioactivity Data of Pharmaceutical Patents via Agentic AI Viktoriia Shepard, Aibulat Musin, Kristina Chebykina, Natalia A. Zeninskaya, Lukia Mistryukova, Konstantin Avchaciov, and Peter O. Fedichev ∗ GERO PTE. LTD., 133 Cecil Street 14-01 Keck Seng Tower, Singapore 069535 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 sys- tematically 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, recov- ering 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. I. INTRODUCTION Pharmaceutical patents represent one of the largest repositories of experimental protein–ligand interaction (PLI) data ever assembled. Thousands of Structure– Activity Relationship (SAR) tables, each documenting binding affinities across hundreds of compounds, are filed annually with patent offices worldwide and often appear years before or independently of peer-reviewed litera- ture [1–3]. Despite billions of dollars in R&D invest- ment, this knowledge remains effectively “dark”: tech- nically public, yet computationally inaccessible, trapped in unstructured archives that no existing database has systematically captured. This data gap matters acutely. Recent breakthroughs in de novo protein design [4–6] and structure predic- tion [7–9] have transformed what AI can do in drug discovery, but these models face a generalization crisis. Even the best architectures struggle to predict activity in new chemical or biological spaces when trained on sparse data [10–12]. Closing this gap requires two things simul- taneously: massive, diverse training sets and genuinely held benchmarks to demonstrate robust model general- ization [13–15]. Pharmaceutical patents could provide both provided their contents were easily accessible. PLI data is the ground truth for both training and benchmarking [16, 17]. The leading public repository, BindingDB, relies on the slow manual curation of the lit- erature [18] and covers only a fraction of the available patent data. Automating patent extraction has histori- cally failed due to the specific challenges of patent lan- guage and their multimodal nature [19, 20]: information is fragmented across unstructured text, complex tables, ∗ Correspondence email address: [email protected] and chemical diagrams, and high-fidelity extraction de- mands reconstructing the complete link between a spe- cific compound, the assay performed, and the resulting activity against a protein target [20]. Existing pipelines like SureChEMBL index chemical structures at scale but lack systematic extraction of quantitative binding values or mapping to biological targets [21, 22]. The result is a self-reinforcing bottleneck: models are evaluated on the same datasets they were trained on, making it impossible to distinguish genuine generalization from memorization. Agentic AI systems break this bottleneck. Decom- posing complex extraction into specialized sequential agents reduces hallucination rates and maintains accu- rate compound–target associations across documents ex- ceeding 500,000 tokens. With the rapid rise in LLM rea- soning and falling inference costs [23, 24], hierarchies of specialized agents can now mimic expert human work- flows at negligible marginal cost – making systematic patent mining economically feasible for the first time. We presentHARVEST(High-throughput Agent Re- trieval of Values for Evaluated Small-molecules and Tar- gets), an automated multi-agent pipeline for the ex- traction of SARs from USPTO bulk data. The sys- tem autonomously parses patent XML, resolves chem- ical aliases to canonical SMILES, and maps biological targets to UniProt identifiers. When applied to 164,877 patent archives, the pipeline produced 3.36 million ac- tivity records from 40,902 patents at a cost of only $0.11 per document – completing in under a week a task that would require over 55 years of continuous manual ex- pert labor. This dataset substantially expands the known chemical-biological landscape, recovering 365,713 unique scaffolds and 1,108 protein targets entirely absent from BindingDB. A central contribution of this work is H-Bench, an open benchmark derived from HARVEST comprising bioactivity data absent from all existing public reposi- .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint 2 tories. H-Bench supports two distinct evaluation scenar- ios: scaffold-generalization on known targets, and target- generalization across proteins with no prior public bioac- tivity data. Our evaluation of the leading open-source structure-based model Boltz-2 [9] on H-Bench reveals a two-dimensional generalization gap: model performance degrades both when chemistry is novel and when pro- tein targets lack prior bioactivity data – demonstrating that current models have not yet learned fully transfer- able binding physics. Together, HARVEST and H-Bench convert billions of dollars of inaccessible R&D knowledge into open scientific infrastructure, directly addressing the data bottleneck that limits AI-driven therapeutic discov- ery. II. RESULTS A. HARVEST Substantially Expands Public Bioactivity Space The HARVEST pipeline (Section IX) was applied to 164,877 USPTO patent archives pre-selected for the pres- ence of chemical structures and bioactivity mentions. Processing 50 documents in parallel, the system pro- duced a final dataset of 3.36 million activity records extracted from 40,902 patents (25% of the input cor- pus), averaging 82 records per document containing ex- tractable data. The remaining patents either lacked ex- tractable bioactivity data or fell outside the pipeline’s current parsing capabilities (see Section IV). The au- tomated pipeline achieves a consistently higher docu- ment throughput than manual curation across the full 25-year publication window examined (Fig. 1a). Al- though BindingDB reports a higher average number of activity records per patent (Fig. 1b), this is a result of human curators prioritizing the most data-rich doc- uments. HARVEST achieves comparable yield on shared patents, confirming equivalent extraction depth. Because the marginal cost of processing is negligible, HARVEST captures data from thousands of documents that would not justify the cost of manual labor. Overall, HARVEST and BindingDB contain a com- parable total volume of protein–ligand interactions. Af- ter aggregating multiple measurements per compound– target pair and applying inclusion filters (Section IX D), HARVEST yields 2.26M unique PLIs, comparable to BindingDB’s 2.21M across all sources (Fig. 1c). However, these totals reflect fundamentally different source com- positions: BindingDB aggregates records from patents, journal articles, and ChEMBL, whereas HARVEST draws exclusively from patent text. Restricting the comparison to patent-derived records, HARVEST con- tributes nearly three times as many PLIs as BindingDB’s patent subset, confirming substantially deeper coverage of the patent corpus. Across the combined chemical-biological landscape of 8,710 protein targets, only 34.1% are shared between the two databases (Fig. 2a). A further 1,108 targets (12.7%) are covered exclusively by HARVEST, while the major- ity of BindingDB-only targets (53.2%) reflect literature and ChEMBL sources outside the patent corpus. Crit- ically, novelty extends well beyond HARVEST-exclusive proteins: for the 2,969 shared targets, 37.0% of HAR- VEST PLIs (851,642 interactions) and 43.4% of Murcko scaffold clusters (424,772 clusters) are absent from Bind- ingDB (Fig. 2b–c). This indicates that substantial chem- ical diversity remains trapped in patent archives even for the most extensively studied drug targets. The target distribution reflects established drug dis- covery priorities: enzymes and kinases predominate due to their well-defined binding pockets [25], while transcrip- tion factors are less represented due to the difficulty of targeting protein-protein interfaces with small molecules (see Table S1). To assess the utility of this expanded chemical space for structure-activity relationship (SAR) analysis, we quantified the density of activity cliffs. These are de- fined as pairs of structurally similar compounds (Tani- moto similarity≥0.7) that exhibit a substantial differ- ence in biological activity (∆pActivity ≥ 1.5), where pActivity = − log10[M] [26]. Activity cliffs represent high-information data points where minor chemical mod- ifications critically determine a molecule’s effect. Across the 2,969 shared targets with at least one activity cliff, HARVEST provides greater cliff density for 42% of pro- teins, while BindingDB leads for 54%; only 3% show equivalent coverage. This asymmetric complementarity– where each resource uniquely enriches SAR information for distinct protein subsets–argues strongly for merg- ing both datasets to maximize the structural discontinu- ities available for lead optimization and machine learning model training. B. HARVEST Extracts Data Comparable in Quality to Manual Curation To validate the fidelity of automated extraction, we benchmarked HARVEST against the manually curated BindingDB (BDB) dataset [18]. Since HARVEST oper- ates exclusively on US patents, all comparisons use the patent-derived subset of BDB unless otherwise noted. Density distributions of binding affinity, molecular weight, and synthetic accessibility show close agreement between HARVEST and BDB across the full range of values (Figs. 3a–3c), consistent with previously reported patent-derived compound profiles [18]. We next assessed record-level accuracy by pairwise comparison of activ- ity values for PLIs present in both databases, matched by UniProt accession and InChIKey connectivity layer (first block) [27]. Across 319,954 matched PLIs from 5,668 shared patents, the distribution of activity residuals (∆pActivity) is highly centered around zero, with 91.0% of PLIs showing near-identical values (Fig. 4a). This cor- responds to a high quantitative correlation (Pearson r = .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint 3 FIG. 1:(a) Number of patents included per year from USPTO for HARVEST (green) and BindingDB (orange), illustrating the consistently higher document coverage achieved by automated extraction across the full 25-year window. (b) Mean number of activity records per document for HARVEST (green), BindingDB (orange), and their intersection (blue). The convergence of HARVEST and BindingDB on shared patents confirms equivalent extraction depth; the lower overall HARVEST average reflects broader document selection criteria that include patents with fewer activity records. (c) Comparison of unique protein–ligand interactions (PLIs) between HARVEST and BindingDB. The BindingDB bar is decomposed by data source: ChEMBL-derived entries (orange), US Patent extractions (blue), and remaining sources (pink). HARVEST yields 2.26M unique PLIs from patent text alone, exceeding the 2.21M across all BindingDB sources combined. FIG. 2:Comparison of protein and chemical space coverage between HARVEST and BindingDB. (a) Protein target composition across both databases (n = 8,710 total): 34.1% of targets are shared, while 12.7% are covered exclusively by HARVEST and 53.2% exclusively by BindingDB. (b) Protein–ligand interaction (PLI) diversity within HARVEST for the 2,969 shared protein targets: 37.0% of PLIs are novel relative to BindingDB, with an additional 2.4% associated with proteins unique to HARVEST. (c) Murcko scaffold cluster diversity within HARVEST for shared targets: 43.4% of scaffold clusters are absent from BindingDB, with a further 2.6% belonging to HARVEST-exclusive proteins. Orange indicates novel content unique to HARVEST; blue indicates overlap with BindingDB; green indicates targets absent from BindingDB entirely. 0.925, Spearmanρ= 0.937). Validation against 68,209 independent article-derived records confirms this consis- tency (r= 0.851, ρ = 0.875), with the broader residual distribution reflecting inherent experimental variability between separate data sources (Fig. 4b). The residual analysis reveals isolated spikes at ∆pActivity = ±3, the signature of 1,000-fold unit conver- sion errors (nM/µM confusion), the most common cura- tion artifact reported by the BindingDB team [18]. These affect ∼1.4% of patent PLIs and ∼1.2% of article PLIs. To determine which database held the correct value, we manually verified the original patent text for the 20 patents contributing the most discrepant records, collec- tively covering 3,706 of the 5,499 affected PLIs (67%). .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint 4 (a) (b) (c) FIG. 3:Physicochemical property distributions for HARVEST and BindingDB (US patents subset). (a) Binding affinity distributions (combined IC 50,K i,K d, and EC 50). Only exact numeric measurements (relation “=”) are included. (b) Molecular weight distributions, with both peaking near 450 Da. (c) Synthetic accessibility (SA) scores, both peaking at SA≈3. The alignment across three descriptors confirms that HARVEST extracts a representative chemical space without systematic physicochemical bias relative to manual curation. FIG. 4:Distribution of activity residuals (∆pActivity = BDB−HARVEST) for matched compound–target pairs. The y-axis is broken to show both the dominant central peak and the tail structure. (a) US patent-derived BindingDB records (n= 319,954): 91.0% of pairs fall in the central bin (∆pActivity ≈ 0), corresponding to the high correlation (r = 0.925) observed between manual and automated curation. (b) Article-derived BindingDB records (n = 68,209): the broader distribution (59.3% central bin, r = 0.851) reflects genuine measurement variability across independent data sources rather than extraction error. In both panels, the distinct spikes at ∆pActivity = ±3 identify 1,000-fold unit conversion errors (nM/µM confusion), the most frequent artifact in both curation workflows. The sharp central distribution confirms that HARVEST achieves human-level extraction fidelity across hundreds of thousands of records. At the record level, HARVEST held the correct value in 92% of verified PLIs, BindingDB in 5%, with 3% remain- ing ambiguous. This indicates that automated agents are substantially less prone to unit-conversion errors than manual curators. The preceding comparison is restricted to records present in both databases. To assess record-level over- lap, we cross-referenced each record against the other database within the same patent (Table I). Of 606,456 HARVEST records, 80.3% find an exact match in Bind- ingDB; conversely, HARVEST recovers 70.5% of 690,869 BindingDB records. To understand the sources of dis- crepancy, we manually verified the original patent text for the patents contributing the most records in each mismatch category (Tables S3–S5), collectively covering 12.1% of all mismatched records. Target mismatch. When both databases find the same compound but assign it to different proteins, HAR- VEST more often identified the correct target. In the remaining cases, HARVEST misclassified assay readout biomarkers as direct binding targets. Additional dis- agreements arose from differences in protein name reso- lution and from unclear species assignment in the patent text. Compound mismatch. Differences in compound sets mostly reflect extraction coverage. On large patents .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint 5 TABLE I: Cross-validation on 5,668 shared patents. Each row indicates how a record from one database was matched in the other on the same patent. Full match : both compound and target agree. Target mismatch: same compound found but assigned to a different protein. Compound mismatch : same target found but linked to a different compound. No overlap : neither compound nor target found on that patent. HARVEST (n=606,456) in BDB BDB (n=690,869) in HARVEST Full match 486,950 (80.3%) 486,950 (70.5%) Target mismatch 61,823 (10.2%) 53,784 (7.8%) Compound mismatch 47,685 (7.9%) 124,061 (18.0%) No overlap 9,998 (1.6%) 26,074 (3.8%) (500+ examples), HARVEST sometimes extracts fewer compounds than BindingDB. On small and medium patents, the opposite holds: HARVEST captures data formats that manual curators skip, including inline activ- ity values in synthesis text, semi-quantitative statements (“IC50 <5µM for all 590 examples”), and non-numeric activity codes (letter grades, symbolic ratings, +/++ scales) used in place of exact measurements. Additional differences arose from HARVEST extraction or structure- normalization errors. Overall, HARVEST yields some- what fewer PLIs on shared patents (606K vs. 691K), pri- marily due to incomplete extraction from the largest doc- uments. No overlap.The rare cases (1.6–3.8%) where nei- ther compounds nor targets overlap are often caused by BindingDB patent mapping or identifier errors, or by the same target mismatch and compound extraction factors described above. C. H-Bench: A Public Benchmark Dataset A central deliverable of this work isH-Bench, an open benchmark of bioactivity data extracted by HARVEST that is not present in BindingDB. BindingDB serves as the primary training source for most modern bioactivity models, either as a direct source [9] or through its inte- gration into the PDB [28]. H-Bench therefore provides a structurally novel held-out resource for rigorous model evaluation on records likely omitted from training sets. To establish the benchmark’s integrity, we used a graph-based integer linear program (ILP) to maximize structural separation between HARVEST compounds and existing records. This process identified two distinct subpopulations (Fig. 5): the Valid subset (n = 98, 105, median similarity ≈ 0.47), occupying novel chemical space, and the Common subset (n = 245, 836, median similarity ≈ 0.70), containing compounds structurally closer to BindingDB entries. We provide a Python script that implements this algorithm, allowing researchers to compare H-Bench against their own training data. This FIG. 5:Chemical distance of H-Bench compounds from BindingDB, measured as maximum Tanimoto similarity to the nearest BindingDB compound. (a) Valid subset (n = 98,105; median = 0.47): compounds with low structural overlap with BindingDB, released as the H-Bench benchmark. (b) Common subset (n = 245,836; median = 0.70): compounds structurally proximal to BindingDB entries. tool automatically identifies and moves any similar struc- tures into the Common subset, effectively censoring po- tential data leakage and preventing over-optimistic re- sults during model evaluation. H-Bench covers 53 protein targets spanning diverse classes (Supplementary Table S6), including enzymes, membrane receptors, and transcription factors. Of these, 44 targets overlap with BindingDB and 9 are entirely novel. To ensure reliable evaluation, targets were filtered to maintain an activity balance between 20% and 55% (mean ≈ 33% active, Supplementary Table S6), avoiding the extreme class imbalance common in public screen- ing datasets. These two components, the set of entirely new proteins and the novel structural clusters for overlap- ping targets, support two primary evaluation scenarios: generalization to novel chemical structures on known tar- gets and generalization to entirely uncharacterized pro- tein targets. To investigate whether current high-performance mod- els learn transferable binding physics or simply rely on structural similarity to their training data, we evaluated the leading open-source structure-based model Boltz-2 [9] on the H-Bench benchmark. This evaluation uses up to 120 ligands per protein, balanced between active and in- active classes. For the Common subset evaluation, we excluded molecules with Tanimoto similarity above 0.75 to BindingDB to mitigate direct memorization effects. Of the three Boltz-2 output scores evaluated, affinity pred valueperformed best overall. Perfor- mance varied systematically across the three bench- mark subsets: AUC-ROC increased monotonically from Valid/new targets (novel proteins, novel chemistry; AUC≈ 0.52) to Valid/known targets (known proteins, novel chemistry; AUC ≈ 0.63) to Common/known tar- gets (known proteins, structurally familiar chemistry; AUC ≈ 0.70), directly mirroring proximity to train- ing data (Fig. 6a). The score intended for hit discov- .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint 6 FIG. 6:Evaluation of Boltz-2 predicted binding scores against experimental activity on H-Bench. (a) Per-protein AUC-ROC for three Boltz-2 output scores (mean±SEM across proteins), evaluated on three benchmark subsets that differ in both chemical and target novelty: Valid/new targets (9 proteins exclusive to HARVEST, novel chemistry); Valid/known targets (44 proteins shared with BindingDB, novel chemistry); and Common/known targets (44 proteins shared with BindingDB, chemistry structurally proximal to BindingDB). The dashed line marks random performance (AUC = 0.5). Performance increases monotonically with proximity to training data across all three scores, revealing that Boltz-2 predictions reflect structural familiarity more than transferable binding physics. (b) Experimental pActivity (-log 10[M]) vs. predicted affinity pred valuewith linear fits and Pearson r, for Valid/known (orange) and Common/known (blue) subsets. (c) pActivity vs. affinity prob binary. Points in (b, c) are colored by Tanimoto similarity to the nearest BindingDB compound. ery,affinity prob binary, performed near-randomly across all subsets (AUC≈0.52–0.58), indicating a signif- icant challenge for virtual screening applications where the model must distinguish binders from decoys. The confidence scoremetric, which reflects structural plau- sibility of predicted complexes, showed a similar gradient but remained near random on genuinely novel chemistry. A key finding is the systematic performance gap be- tween the two subsets: points colored by Tanimoto sim- ilarity cluster toward higher predicted scores as similar- ity increases (Fig. 6b,c). This reveals that model pre- dictions are often biased by proximity to training data rather than underlying binding physics. Scatter analysis confirms only a weak but statistically significant correla- tion between predicted affinity and experimental activity (Pearsonr≈ −0.31 for the Valid set). Overall, Boltz-2 performance on H-Bench is modest but above random on known targets, and degrades on proteins entirely absent from public bioactivity reposi- tories, which is consistent with results reported on the model’s own proprietary benchmarks [9]. This two– dimensional generalization gap — across both chemical and target space — confirms that H-Bench provides a stringent and unbiased evaluation resource. The results suggest that current structure-based models have not yet learned fully transferable binding physics, and that both novel chemical scaffolds and uncharacterized targets re- main a fundamental challenge for AI-driven drug discov- ery. III. DISCUSSION The pharmaceutical industry has invested hundreds of billions of dollars in protein–ligand interaction experi- ments over the past three decades. Patent law was de- signed to make this knowledge public, yet the practi- cal inaccessibility of unstructured patent archives has meant that this investment remained effectively private– confined to commercial databases behind expensive sub- scriptions or simply uncurated. HARVEST transforms this situation: by processing the full USPTO pharma- ceutical corpus in under a week at $0.11 per document, it demonstrates that the era of dark bioactivity data is ending. The 3.36 million activity records recovered, in- cluding 365,713 structural clusters and 1,108 protein tar- gets absent from BindingDB, represent a qualitative ex- pansion of the computable chemical-biological landscape available to the global research community. This capability distinguishes HARVEST from all ex- isting approaches along two dimensions simultaneously: scale and semantic depth. SureChEMBL provides a high- volume index of chemical structures [21]; however, with- out quantitative binding context or protein identity reso- lution it answers “what molecules appear in patents” but not “what do they do and against which target.” Bind- ingDB provides exactly that semantic depth but is funda- mentally constrained by the throughput of human expert curation [18]. HARVEST achieves BindingDB-level ex- traction fidelity – 91% agreement on matched records, with lower unit conversion error rates than manual cura- tors – at 3,500 times the throughput. Recent LLM-based systems like BioMiner [29] and BioChemInsight [30] rep- .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint 7 HAR VEST BioMiner [29] BioChemInsight [30] Source Patents Articles Patents Documents processed 164,877 11,683 181 Documents with bioactivity 40,902 11,683 N/A Records extracted 3.36M 68K N/A Throughput∗ 2 sec/doc 14 sec/doc N/A SMILES Yes Yes Yes Protein ID UniProt AlphaFold path N/A Unit normalization Yes No No Public dataset Partially Yes No TABLE II: Comparison of automated bioactivity extraction systems. ∗Throughput measured under different conditions: HARVEST uses cloud-based LLM API with 50-document parallelism; BioMiner reports 14s/paper on 8×V100 GPUs. Only HARVEST performs full resolution to UniProt identifiers and standardized units. resent important progress but remain restricted to scien- tific articles or small-scale proofs-of-concept. HARVEST is the first system to demonstrate this combination of fi- delity and scale across a full national patent corpus (Ta- ble II). Beyond raw extraction, a significant advantage of HARVEST is its automated data normalization. The system resolves varied and often ambiguous protein de- scriptors to canonical UniProt identifiers [31] and stan- dardizes diverse activity units into a consistent numeric format. By producing a dataset structured similarly to BindingDB, HARVEST provides a machine-actionable resource that is immediately ready for training machine learning models. This eliminates the massive manual post-processing and data-cleaning efforts typically re- quired when working with “dirty” automated extractions from patent literature. The cost profile of HARVEST fundamentally changes the economics of medicinal chemistry data. Traditional curation projects like BindingDB process approximately 1, 500 patents over 2 years [18]; at that rate, processing our full corpus would require over 55 years of contin- uous expert labor. HARVEST completed this task in under a week (Table II). This efficiency removes the fi- nancial barriers that have long confined large-scale bioac- tivity data to expensive commercial platforms such as Reaxys [32, 33] or GOSTAR [34]. For approximately the cost of six months’ worth of a commercial sub- scription, any research organization can now generate a proprietary-scale dataset in a matter of weeks. This democratization of data allows academic groups to com- pete in a landscape previously dominated by well-funded commercial providers. Critically, this cost structure also changes the update cycle for bioactivity data. Bind- ingDB’s manual curation introduces a lag of years be- tween patent publication and database inclusion [18]. HARVEST can in principle be rerun on new USPTO weekly releases continuously, maintaining a near-real- time mirror of the public patent bioactivity landscape FIG. 7:Cost breakdown for processing the USPTO patent corpus with HARVEST, including development and testing expenses. The production cost of$0.11 per document excludes these one-time development costs. – something no manual system can achieve regardless of funding. A striking finding is that despite recovering 1,108 pro- tein targets entirely absent from BindingDB, these novel targets account for only 2.4% of total extracted interac- tions. The vast majority of new data deepens coverage of established targets rather than revealing entirely new bio- logical associations. This pattern likely reflects two con- verging forces: the pharmaceutical industry’s strategic focus on validated targets where biological risk is lower, and the temporal lag between patent filing and scientific publication. The release of H-Bench addresses a fundamental eval- uation problem in AI-driven drug discovery. Because modern models are often trained on the same core pub- lic datasets, it is difficult to distinguish genuine gener- alization from memorization [10, 11]. H-Bench provides a structurally guaranteed held-out resource because its records are derived from patent literature currently ab- sent from BindingDB. Our three-way evaluation of Boltz- 2 [9] – across novel targets, known targets with novel chemistry, and known targets with familiar chemistry – reveals that the generalization gap is two-dimensional: models degrade both when chemistry is novel and when targets lack prior bioactivity data. This is a more pre- cise characterization of model limitations than binary train/test splits on BindingDB alone can provide, and .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint 8 it points directly toward what the field needs to improve: training data that covers undercharacterized targets, and evaluation frameworks that separately stress-test chemi- cal and biological generalization. H-Bench provides both. The success of HARVEST relies heavily on the high quality of structured data provided by the USPTO. Be- cause the USPTO includes chemical structures as em- bedded ChemDraw files within its XML archives, they can be reliably converted into canonical SMILES. In con- trast, processing raw PDF documents would require com- plex OCR and significantly more intensive processing, which often introduces errors. This highlights the criti- cal importance of maintaining publicly available data in well-organized, machine-readable formats to enable au- tomated discovery. The 2025 announcement by the CNIPA in China to promote XML for electronic patent submissions [35] opens a direct pathway to extend HARVEST to Chinese pharmaceutical literature, which would add a massive and currently inaccessible reservoir of medicinal chem- istry data. More broadly, the value of open, structured data formats for enabling downstream scientific compu- tation cannot be overstated: the difference between a PDF and an XML archive is the difference between dark data and actionable knowledge. The principles demonstrated by HARVEST extend far beyond bioactivity extraction. Much of humanity’s ex- pert knowledge remains practically inaccessible: techni- cally public in patents, regulatory filings, and clinical records, yet effectively “dark” due to the prohibitive cost of human curation. The multi-agent architecture intro- duced here, which decomposes complex document un- derstanding into specialized sequential agents grounded by structured resolution pipelines, is directly applicable to any domain facing this barrier. Recent examples in- clude agentic systems for multilingual pharmaceutical as- set scouting [36, 37] and automated extraction from clin- ical and regulatory documents [38, 39]. As LLM capabili- ties improve and inference costs decline [24], the marginal cost of converting massive document corpora into struc- tured knowledge is approaching zero. The critical ques- tion is no longer whether this conversion is feasible, but which datasets are most vital to recover and how to en- sure the resulting knowledge remains a public good rather than a proprietary asset. IV. LIMITATIONS HARVEST’s current scope defines a clear roadmap for future development. Four categories of patent data re- main to be addressed. First, the system cannot yet process Markush structures. These are highly complex diagrams used to represent large families of molecules simultaneously via variable parts. Automatically “un- packing” these structures into a list of specific, individ- ual compounds is a significant technical challenge that is not yet implemented, although these cases account for only 8% of the data points we excluded (Section IX D). Second, graphical data such as dose–response curves re- main inaccessible, preventing the extraction of param- eters from assays reported only in figure form. Third, protein target resolution is limited by the scope of the curated Swiss-Prot database [40] and by the inherent ambiguity of multi-subunit complexes like integrins or IL-23. Patents often reference these targets using incon- sistent subunit, heterodimer, or domain-level descriptors, making canonical mapping difficult. Finally, LLM safety policies occasionally caused the system to refuse patents targeting high-risk pathogens such as Ebola virus. This has resulted in systematic coverage gaps in certain antivi- ral research areas, where the model interprets the data extraction as a violation of safety guidelines. Our cross-validation is grounded in the 5,668 patents shared with BindingDB, representing 14% of the HAR- VEST corpus. For the remaining 86%, no indepen- dent reference currently exists – establishing such a ref- erence, through prospective experimental validation of HARVEST-derived predictions or through community curation efforts, is a priority for future work. Based on the consistency of activity residual distributions between validated and unvalidated subsets, we estimate the error rate for uncharacterized records at level not exceeding 10-15%, comparable to known error rates in manually curated databases [18]. V. CONCLUSION We have presented HARVEST, a multi-agent LLM pipeline that converts the dark bioactivity data of phar- maceutical patents into open, computable scientific in- frastructure. By decomposing complex document un- derstanding into specialized sequential agents, HAR- VEST achieves BindingDB-level extraction fidelity at 3,500 times the throughput of manual curation – pro- cessing 164,877 patent archives in under a week at$0.11 per document, recovering 365,713 structural clusters and 1,108 protein targets entirely absent from public reposi- tories. At this cost, comprehensive patent mining is no longer a luxury for well-funded commercial providers; it is accessible to any research group with a compelling sci- entific question. The accompanying H-Bench benchmark addresses an equally critical problem: the lack of genuinely held- out evaluation data for protein–ligand interaction mod- els. Our three-way evaluation of Boltz-2 [9] reveals that the generalization gap is two-dimensional – model per- formance degrades both when chemistry is novel and when protein targets lack prior public bioactivity data. This finding exposes a fundamental limitation of models trained exclusively on manually curated repositories, and establishes H-Bench as a stringent, leakage-free resource for driving the development of more robust, physics- aware models. Together, HARVEST and H-Bench represent a prac- .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint 9 tical answer to the data bottleneck that currently limits AI-driven drug discovery. The hundreds of billions of dol- lars in R&D investment embedded in the global patent corpus was always legally public; it was never compu- tationally accessible. As LLM inference costs continue to fall [24], the same approach can be extended to new patent jurisdictions, to regulatory filings, and to any do- main where expert knowledge remains trapped in un- structured text. The era of dark data in medicinal chem- istry is ending. VI. ACKNOWLEDGMENTS We thank Mikhail Batin, Alexey Strygin, and Vita Stepanova, the organizers of the Agentic AI Against Ag- ing (AAAA) hackathon, for providing the venue that fa- cilitated the inception of this work. We are also deeply grateful to Vladimir Manujlov, an employee of Gero, for his assistance and guidance during the hackathon and for his valuable edits and review of this manuscript. We ex- tend our gratitude to Daniel Kravtsov for providing his technical expertise in modern agentic systems, which of- fered helpful insights during the design of the HARVEST pipeline. Finally, we acknowledge the use of the Gemini Large Language Model for assistance in the drafting and linguistic refinement of this manuscript. VII. CONFLICTS OF INTEREST K.A., L.M. and P.F. are employees and equity hold- ers of Gero PTE. LTD., a company developing AI-driven drug discovery tools. Gero proposed the patent mining challenge at the AAAA hackathon and may benefit com- mercially from the methods and datasets described in this work. V.S., A.M., K.C., and N.A.Z. were selected as the winning implementation team by the independent AAAA organizing committee and subsequently contributed to this work under a contract with Gero. All authors de- clare no other competing financial interests. VIII. DATA AVAILABILITY The H-Bench benchmark and the full HARVEST dataset are available at https://github.com/gero-s cience/HARVEST under the Creative Commons Attri- bution 4.0 International (CC BY 4.0) license. The ChemDraw binary file reader is available at https: //github.com/gero-science/cdx_reader. This work also utilizes data from the BindingDB open- source database (September 2025 release), which can be accessed at https://www.bindingdb.org. The raw patent application data used for extraction was obtained from the USPTO Patent Application Full Text Data with Embedded TIFF Images (APPDT), available at https: //data.uspto.gov/bulkdata/datasets/appdt. Detailed dataset statistics, protein family distribu- tions, and activity label balance metrics are provided in the Supplementary Information. IX. MATERIALS AND METHODS A. Patent Data Sources We evaluated several repositories for large-scale bioac- tivity extraction, prioritizing structured data formats, API stability, and cost. Alternative sources such as SureChEMBL [21], Google Patents [41], Google Big- Query Patents Public Data [42] and Lens.org [43] were excluded due to limitations in their data formats (see Supplementary Section S2 for details). USPTO Bulk Data [44] was selected as the primary data source, which provides mostly unlimited downloads supported by an API without any additional costs. The structured XML format of the provided documents pre- serves table hierarchies and chemistry tags, which is es- sential for accurately linking compounds to their biolog- ical targets and activity values. This high-density dis- closure is a major advantage of the patent corpus: US patents contain an average of 160 measurements per doc- ument compared to only 40 per scientific article [18]. We retrieved weekly archives from the USPTO Ap- plication Data (APPDT) using the Bulk Datasets API. To ensure the corpus was rich in extractable PLI data, we applied an initial filter to retain only documents containing either embedded chemical structure attach- ments (CDX/MOL) or specific bioactivity keywords (e.g., ”IC50”, ”Ki”, ”Kd”, ”EC50”) identified via regular ex- pression matching. The keyword filter was optimized for high recall, minimizing false negatives at the cost of increased false positives that are subsequently fil- tered by downstream agents. This pre-filtering is sup- ported by our empirical observation that patents lacking CDX/MOL attachments rarely contain structured quan- titative bioactivity data. B. Patent Family Deduplication A single invention can generate multiple legally dis- tinct application records, such as pre-grant publications, granted patents, continuation or divisional applications, and reissue or reexamination documents. These filings frequently contain identical bioactivity tables and chem- ical examples, which can lead to redundant data extrac- tion and inflated record counts. To ensure each unique experimental observation is represented only once, we ag- gregated related filings into continuity clusters. We constructed these clusters by building a directed graph of parent–child relationships retrieved via the USPTO API. To specifically identify content duplication, we restricted the graph edges to priority claims and conti- nuity links: “is a continuation of,” “is a divisional of,” “is .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint 10 FIG. 8:The HARVEST pipeline architecture. USPTO patent documents are parsed by an XML extractor and passed to a three-stage LLM extraction module: Agent 1 identifies biological targets and assay conditions; Agent 2 extracts quantitative bioactivity measurements; and Agent 3 resolves compound aliases to IUPAC names or embedded chemical identifiers. The extracted records are then processed by two resolution agents operating in parallel: Agent 4 converts chemical identities to canonical SMILES via embedded structure files and py2opsin name-to-structure conversion; Agent 5 maps protein names to UniProt identifiers using UniProt FASTA. The final output is the normalized HARVEST dataset of Document–Assay–Result– Compound–Protein records. a national stage entry of,” and “is a reissue of.” We ex- plicitly excluded “continuation-in-part” (CIP) relation- ships from this automated deduplication, as CIP filings often introduce new substantive data not present in the parent application. For each resulting cluster, we selected the most re- cent document by publication date as the representative record. This strategy ensures the capture of the most complete version of the disclosure, as later filings often include corrected tables or refined IUPAC nomenclature. This deduplication process yielded a 17.7% reduction in the total document volume, resulting in a final corpus of unique inventions for bioactivity extraction. C. Multi-Agent Architecture To extract PLIs from linguistically complex patent documents, we developed a multi-agent architecture con- sisting of five specialized agents operating in sequence (Fig. 8). We adopted this multi-stage decomposition be- cause LLMs often exhibit performance loss when required to identify heterogeneous data types simultaneously [23]. Our preliminary experiments on 500 patents confirmed this limitation, as a monolithic single-prompt strategy resulted in three systematic failure modes. First,attention dilution caused the model to lose con- sistent compound–target associations in documents ex- ceeding 200,000 tokens, leading to misattributed activity values. Second, task interference substantially increased SMILES hallucination rates when chemical names, bio- logical targets, and numeric values were requested simul- taneously. Third, output truncation caused premature response termination in patents containing over 500 ac- tivity records, losing a substantial fraction of data. Decomposing the extraction into sequential agents ad- dresses these three failure modes by narrowing the se- mantic scope of each LLM call. This approach is con- sistent with recent findings that task decomposition and agent specialization improve performance in fron- tier models [45–48]. The sequential architecture also im- proves system traceability and enables targeted prompt optimization for each subtask. Furthermore, this design provides an early termination mechanism: if Agent 1 identifies no biological targets, subsequent steps are skipped to avoid unnecessary computation on irrelevant patents. The pipeline utilizes three LLM-based agents (Agents 1–3) for semantic extraction, followed by two resolution agents (Agents 4–5) for chemical and protein standardization. All agents operate within an asyn- chronous processing framework with configurable worker pools, enabling high-throughput parallel processing of the patent corpus. Agents 1–3: Semantic Extraction Agent 1 (Target Extraction) identifies biological targets such as proteins, enzymes, and receptors, along with test organisms and assay conditions. This extracted context is injected into the subsequent Agent 2 prompt to reduce misattribution errors in documents describing multiple targets. Agent 2 (Activity Extraction) focuses on quanti- tative measurements. It extracts compound aliases (e.g., “Example 1”, “Compound 42”), binding metrics (IC 50, Ki, K d, EC50), numeric values, and measurement units. By embedding the output of Agent 1, Agent 2 performs target-aware extraction, ensuring that each numeric re- .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint 11 sult is correctly linked to its respective assay and protein. Agent 3 (Compound Mapping) resolves the com- mon patent practice of referencing molecules by internal aliases. It maps these aliases to IUPAC nomenclature or embedded chemical identifiers provided in the text. Iso- lating this task into a dedicated agent allows handling the complexity: compound identities can be scattered across hundreds of pages, requiring full-document con- text awareness. The pipeline utilizes three LLM-based agents to han- dle the initial semantic extraction from patent text. These agents are built on google/gemini-2.5-flash, selected for its 1-million-token context window and effi- cient prompt caching. Agents 4–5: Resolution Agent 4 (Chemical Structure Resolution) con- verts resolved chemical identities into standardized SMILES strings. While USPTO patent XML embeds both MOL and CDX (ChemDraw binary) files for all chemical compounds; we use CDX exclusively due to sys- tematic corruption discovered in the distributed MOL representations. This corruption typically occurs dur- ing default CDX-to-MOL conversion, where atom type aliases or common substitutions (e.g., “Me”, “Et”, “HN”, or “CN”) are erroneously interpreted as carbon atoms. By parsing the ChemDraw binary files directly, we avoid these substitution errors. The impact of CDX-based res- olution on extraction fidelity is shown in Supplementary Table S2. The parser is implemented in Python and is released as an open-source tool (see Data Availabil- ity, Section VIII). For patents containing only IUPAC nomenclature without embedded structural files, we use py2opsin [49] as a fallback. Agent 5 (Protein Target Resolution) maps ex- tracted protein names to UniProt identifiers [31] and retrieves their associated amino acid sequences. Due to the prevalence of non-standard nomenclature, unoffi- cial aliases, and context-dependent target specification in patent documents, this task requires LLM assistance. We employ openai/gpt-5-2025-08-07 for this stage, as its superior reasoning capabilities relative to smaller Gemini Flash enable more accurate interpretation of ambiguous biological context. When species is unspecified in the patent text, the agent defaults to Homo sapiens, consis- tent with the therapeutic focus of pharmaceutical patents and the composition of BindingDB, where 83% of binding data derive from human proteins [18]. D. Dataset Inclusion Criteria and Normalization HARVEST targets quantitative protein–ligand bind- ing data. The following criteria, enforced through agent prompts and post-processing, define the records included in the final dataset. Binding metrics. Only IC50, Ki, Kd, and EC50 mea- surements are retained. These four metrics account for over 80% of records extracted by the agents. Metrics such as percent inhibition are excluded, as extracting meaningful binding constants from single-point or curve- dependent measurements would require additional pro- cessing logic not justified by their low prevalence in the corpus. All retained values are normalized to nanomolar (nM). Defined protein targets. Only targets mappable to one or more UniProt identifiers are retained. This in- cludes multi-subunit complexes, which are represented as semicolon-separated accessions. Approximately 87% of extracted records were successfully mapped. Records where the extracted target refers to a cell line or pheno- typic endpoint (e.g., “HeLa cytotoxicity” or “A549 cell viability”) rather than a specific protein are excluded. Direct binding assays. Records derived from phe- notypic or cell-based readouts, such as cytotoxicity, cell viability, proliferation, are excluded. These reported ac- tivity values reflect aggregate cellular responses rather than direct protein–ligand interactions. This filter re- moved 230,306 records (6.3%). Markush structure exclusion. Patents containing only Markush structures or generic R-group representa- tions are excluded. Enumerating individual compounds from these combinatorial representations requires spe- cialized chemical reasoning not currently implemented in the pipeline. Because most patents additionally report concrete compound examples, this criterion removed only 8.0% of records. E. Architectural Design Considerations Several design decisions shaped the final HARVEST architecture and are described here as they may inform similar extraction systems. Context window and the chunking problem. Our initial architecture chunked patent text to fit within 256K-token context windows. Tables were split into row groups with preserved headers and several surrounding paragraphs for local context. This approach introduced systematic errors: chemical names spanning multiple ta- ble cells were truncated at chunk boundaries. More criti- cally, patents frequently define compound aliases (e.g., “Compound 1”) in one section and report bioactivity measurements in distant tables, creating unresolvable cross-references in isolated chunks. The availability of models with 1M+ token context windows resolved these issues entirely, with the additional cost offset by prompt caching, resulting in a 3–5× reduction in per-patent in- ference cost. Explicit grounding constraints. A persistent fail- ure mode was the generation of plausible but fabricated IUPAC names for compound aliases. Agent 3 was there- fore instructed to return chemical identifiers only when found verbatim in the patent text, defaulting to TIFF .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint 12 filename references from chemistry tags when no system- atic name was available. Combined with downstream validation via py2opsin and CDX file cross-referencing, this constraint reduced false positive chemical identifica- tions in the final dataset. Temperature tuning.Setting the model tempera- ture to 0 was adopted to minimize hallucinated content. Even modest values (e.g., temperature = 0.1) produced noticeably higher rates of fabricated compound names and activity values during our internal validation. Few-shot example leakage.When a patent con- tained no extractable bioactivity data, the model occa- sionally output values copied from few-shot prompt ex- amples. We addressed this by engineering the prompts to make “no data found” an explicit valid response and post-processing validation to filter records that matched our prompt examples. F. H-Bench Construction Protein selection.We selected 53 proteins span- ning diverse target families–kinases (e.g. EGFR, JAK1, JAK2, KDR), GPCRs (CNR2, GRM5, DRD2, HTR2A, S1PR1), ion channels (P2RX3, P2RX7, GABRA1), lipid transfer proteins (CETP), nuclear receptors (ESR1, NR3C1, RORC), and epigenetic regulators (HDAC6, BRD4, EZH2). Of these, 44 are “overlap” targets shared with BindingDB (each with>30 structural clusters and >1,000 BindingDB compounds) and 9 are “novel” tar- gets present only in HARVEST (>100 compounds). Di- versity across families was ensured by requiring 2–3 rep- resentatives per ChEMBL L2 target class and removing redundant proteins via sequence clustering (MMseqs2, 40% identity threshold). An activity-balance filter re- tained only proteins for which 20–55% of Valid-subset compounds are active. Cluster-based splitting.For each protein, com- pound Morgan fingerprints (radius 2, 2048 bits) were computed and clustered via complete-linkage hierarchical clustering at a Tanimoto distance threshold of 0.2. Clus- ters were initially labeled by data source: A (HARVEST- only) or B (BindingDB-only). A cluster connectivity graph was then constructed using centroid Tanimoto sim- ilarity (threshold 0.225, ≤2 hops), and an integer linear program (ILP) identified the minimum set of A clus- ters to relabel as buffer C, maximizing structural sepa- ration between the Valid (A) and BindingDB-only (B) subsets. The final partition yields three non-overlapping sets: Valid (structurally distant from BindingDB), Com- mon (buffer zone, included in H-Bench), andBindingDB- only (excluded from H-Bench). Boltz evaluation. Boltz-2 predictions were gener- ated in affinity mode with 5 recycling steps, MSA server queries, and molecular-weight correction enabled. Up to 120 ligands per protein were sampled, balanced between active and inactive classes and drawn from diverse clus- ters. For compounds in the Common set, we selected only molecules with Tanimoto similarity below 0.75. Three output scores were evaluated: affinity pred value, affinity prob binary, andconfidence score. Per- protein AUC-ROC was computed for proteins with ≥5 compounds and both activity classes present, yielding 6,469 predictions across 53 proteins (Valid) and 5,399 predictions across 44 proteins (Common). [1] M. Bregonje, Patents: a unique source for scientific tech- nical information in chemistry related industry?, World Patent Inf 27, 309 (2005). [2] M. E. Valentinuzzi, Patents and scientific papers: Quite different concepts: The reward is found in giving, not in keeping [retrospectroscope], IEEE Pulse 8, 49 (2017). [3] C. Southan, P. Varkonyi, K. Boppana, S. A. R. P. Jagar- lapudi, and S. Muresan, Tracking 20 years of compound- to-target output from literature and patents, PLOS ONE 8, e77142 (2013). [4] J. L. Watson, D. Juergens, N. R. Bennett, B. L. Trippe, J. Yim, H. E. Eisenach, W. Ahern, A. J. Borst, R. J. Ragotte, L. F. Milles, B. I. M. Wicky, N. Hanikel, S. J. Pellock, A. Courbet, W. Sheffler, J. Wang, P. Venkatesh, I. Sappington, S. V´ azquez Torres, A. Lauko, V. De Bor- toli, E. Mathieu, S. Ovchinnikov, R. Barzilay, T. S. Jaakkola, F. DiMaio, M. Baek, and D. Baker, De novo design of protein structure and function with RFdiffu- sion, Nature 620, 1089 (2023). [5] N. R. Bennett, J. L. Watson, R. J. Ragotte, A. J. Borst, D. L. See, C. Weidle, R. Biswas, Y. Yu, E. L. Shrock, R. Ault, P. J. Y. Leung, B. Huang, I. Goreshnik, J. Tam, K. D. Carr, B. Singer, C. Criswell, B. I. M. Wicky, D. Vafeados, M. Garcia Sanchez, H. M. Kim, S. V´ azquez Torres, S. Chan, S. M. Sun, T. T. Spear, Y. Sun, K. O’Reilly, J. M. Maris, N. G. Sgourakis, R. A. Melnyk, C. C. Liu, and D. Baker, Atomically accurate de novo design of antibodies with RFdiffusion, Nature 649, 183 (2026). [6] F. J. Ferreira and A. S. Carneiro, Ai-driven drug dis- covery: a comprehensive review, ACS omega 10, 23889 (2025). [7] J. Jumper, R. Evans, A. Pritzel, T. Green, M. Fig- urnov, O. Ronneberger, K. Tunyasuvunakool, R. Bates, A. ˇZ´ ıdek, A. Potapenko, A. Bridgland, C. Meyer, S. A. A. Kohl, A. J. Ballard, A. Cowie, B. Romera-Paredes, S. Nikolov, R. Jain, J. Adler, T. Back, S. Petersen, D. Reiman, E. Clancy, M. Zielinski, M. Steinegger, M. Pacholska, T. Berghammer, S. Bodenstein, D. Silver, O. Vinyals, A. W. Senior, K. Kavukcuoglu, P. Kohli, and D. Hassabis, Highly accurate protein structure prediction with AlphaFold, Nature 596, 583 (2021). [8] J. Abramson, J. Adler, J. Dunger, R. Evans, T. Green, A. Pritzel, O. Ronneberger, L. Willmore, A. J. Bal- lard, J. Bambrick, S. W. Bodenstein, D. A. Evans, C.-C. Hung, M. O’Neill, D. Reiman, K. Tunyasuvu- .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint 13 nakool, Z. Wu, A. ˇZemgulyt˙ e, E. Arvaniti, C. Beat- tie, O. Bertolli, A. Bridgland, A. Cherepanov, M. Con- greve, A. I. Cowen-Rivers, A. Cowie, M. Figurnov, F. B. Fuchs, H. Gladman, R. Jain, Y. A. Khan, C. M. R. Low, K. Perlin, A. Potapenko, P. Savy, S. Singh, A. Stecula, A. Thillaisundaram, C. Tong, S. Yakneen, E. D. Zhong, M. Zielinski, A. ˇZ´ ıdek, V. Bapst, P. Kohli, M. Jaderberg, D. Hassabis, and J. M. Jumper, Accurate structure pre- diction of biomolecular interactions with AlphaFold 3, Nature630, 493 (2024). [9] S. Passaro, G. Corso, J. Wohlwend, M. Reveiz, S. Thaler, V. R. Somnath, N. Getz, T. Portnoi, J. Roy, H. Stark, D. Kwabi-Addo, D. Beaini, T. Jaakkola, and R. Barzilay, Boltz-2: Towards accurate and efficient binding affinity prediction, bioRxiv 10.1101/2025.06.14.659707 (2025), preprint. [10] P. Tossou, C. Wognum, M. Craig, H. Mary, and E. Noutahi, Real-world molecular out-of-distribution: Specification and investigation, Journal of Chemical In- formation and Modeling 64, 697 (2024). [11] H. Fooladi, T. N. L. Vu, M. Mathea, and J. Kirchmair, Evaluating machine learning models for molecular prop- erty prediction: Performance and robustness on out-of- distribution data, Journal of Chemical Information and Modeling 65, 9871 (2025), pMID: 40947919. [12] D. van Tilborg, L. Rossen, and F. Grisoni, Molecular deep learning at the edge of chemical space, ChemRxiv 10.26434/chemrxiv-2025-qj4k3-v3 (2025), preprint. [13] N. Pallikkavaliyaveetil and S. Chandrasekaran, Small data, big challenges: Machine- and deep-learning strate- gies for data-limited drug discovery, Advanced Drug De- livery Reviews 10.1016/j.addr.2025.115762 (2025). [14] K. Debnath, P. Rana, and P. Ghosh, A survey on deep learning for drug-target binding prediction: mod- els, benchmarks, evaluation, and case studies, Briefings in Bioinformatics 26, bbaf491 (2025). [15] A. Gangwal, A. Ansari, I. Ahmad, A. K. Azad, and W. M. A. W. Sulaiman, Current strategies to address data scarcity in artificial intelligence-based drug discov- ery: A comprehensive review, Computers in Biology and Medicine 179, 108734 (2024). [16] J. Dagdelen, A. Dunn, S. Shen, et al., Structured infor- mation extraction from scientific text with large language models, Nat Commun 15, 1418 (2024). [17] J. Michels et al., Natural language processing methods for the study of protein–ligand interactions, J Chem Inf Model 65, 2191 (2025). [18] T. Liu, L. Hwang, S. K. Burley, C. I. Nitsche, C. Southan, W. P. Walters, and M. K. Gilson, Bindingdb in 2024: a fair knowledgebase of protein-small molecule binding data, Nucleic acids research 53, D1633 (2025). [19] D. M. Jessop, S. E. Adams, and P. Murray-Rust, Mining chemical information from open patents, J Cheminform 3, 40 (2011). [20] M. Krallinger, F. Leitner, O. Rabal, M. Vazquez, J. Oyarzabal, and A. Valencia, Chemdner: the drugs and chemical names extraction challenge, J Cheminform7, S1 (2015). [21] G. Papadatos, M. Davies, N. Dedman, et al., Surechembl: a large-scale, chemically annotated patent document database, Nucleic Acids Res 44, D1220 (2016). [22] Y. Gadiya, S. Shetty, M. Hofmann-Apitius, P. Gribbon, and A. Zaliani, Exploring surechembl from a drug dis- covery perspective, Scientific data 11, 507 (2024). [23] S. Qiao, R. Fang, Z. Qiu, X. Wang, N. Zhang, Y. Jiang, P. Xie, F. Huang, and H. Chen, Benchmarking agen- tic workflow generation, arXiv preprint arXiv:2410.07869 (2024). [24] G. Appenzeller, Llmflation: Llm inference cost going down fast, https://a16z.com/llmflation-llm-infer ence-cost/ (2024), discusses Moore’s-law-like declines in LLM inference cost at fixed quality. [25] A. L. Hopkins and C. R. Groom, The druggable genome, Nature reviews Drug discovery 1, 727 (2002). [26] D. Stumpfe, H. Hu, and J. Bajorath, Evolving concept of activity cliffs, ACS Omega 4, 14360 (2019). [27] S. R. Heller, A. McNaught, I. Pletnev, S. Stein, and D. Tchekhovskoi, Inchi, the iupac international chemical identifier, J Cheminform 7, 23 (2015). [28] H. M. Berman, J. Westbrook, Z. Feng, G. Gilliland, T. N. Bhat, H. Weissig, I. N. Shindyalov, and P. E. Bourne, The protein data bank, Nucleic Acids Research 28, 235 (2000). [29] J. Yan, J. Zhu, Y. Yang, Q. Liu, K. Zhang, Z. Zhang, X. Liu, B. Zhang, K. Gao, J. Xiao, et al., Biominer: A multi-modal system for automated mining of protein- ligand bioactivity data from literature, bioRxiv , 2025 (2025). [30] Z. Wang, F. Fu, W. Zhang, L. Yan, Y. Meng, J. Wu, H. Wu, G. Xu, and S. Chen, Biocheminsight: An open- source toolkit for automated identification and recogni- tion of optical chemical structures and activity data in scientific publications, arXiv preprint arXiv:2504.10525 (2025). [31] The UniProt Consortium, Uniprot: the universal pro- tein knowledgebase in 2023, Nucleic Acids Res 51, D523 (2023). [32] Elsevier, Reaxys, https://www.elsevier.com/product s/reaxys (2026), accessed: 22 February 2026. [33] Elsevier, Overview of reaxys data, https://www.elsevi er.support/dataasaservice/answer/overview-of-r eaxys-data (2026), accessed: 22 February 2026. [34] Excelra, Gostar™ small molecules, https://www.excelr a.com/databases/gostar (2026), accessed: 22 February 2026. [35] China National Intellectual Property Administration, Promoting the use of extensible markup language (xml) format for submitting electronic patent application doc- uments, https://www.cnipa.gov.cn/art/2025/5/26/a rt_75_199841.html (2025), accessed: 2026-01-25. [36] A. Vinogradova, V. Vinogradov, L. Greenwood, I. Yasny, D. Kobyzev, S. Kasbekar, K. Nguyen, D. Radkevich, R. Doronin, and A. Doronichev, Hunt globally: Wide search AI agents for drug asset scouting in invest- ing, business development, and competitive intelligence, arXiv preprint arXiv:2602.15019 (2026), preprint. [37] A. Vinogradova, V. Vinogradov, D. Radkevich, I. Yasny, D. Kobyzev, I. Izmailov, K. Yanchanka, R. Doronin, and A. Doronichev, LLM-based agents for competitive land- scape mapping in drug asset due diligence, arXiv preprint arXiv:2508.16571 (2025), preprint. [38] L. Liu, V. Blake, M. Barman, B. Gallego, T. Churches, G. Kennedy, S.-Y. Ooi, G. P. Delaney, and L. Jorm, Us- ing natural language processing to extract information from clinical text in electronic medical records for popu- lating clinical registries: a systematic review, Journal of the American Medical Informatics Association 33, 484 (2026). .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint 14 [39] Z. Yang, H. Yuan, R. Sayeed, A. L. M. Tan, E. Cai, M. Moro, X. Li, H. Ying, N. Brown, G. Weber, et al. , CLINES: Clinical LLM-based information extraction and structuring agent, medRxiv , 2025 (2025), preprint. [40] E. Boutet, D. Lieberherr, M. Tognolli, M. Schneider, and A. Bairoch, Uniprotkb/swiss-prot, Methods in molecular biology (Clifton, N.J.) 406, 89 (2007). [41] Google, Google patents, https://patents.google.com/ (2026), accessed: 22 February 2026. [42] Google, Google patents public datasets: connecting pub- lic, paid, and private patent data, https://cloud.goog le.com/blog/topics/public-datasets/google-paten ts-public-datasets-connecting-public-paid-and -private-patent-data/ (2017), accessed: 22 February 2026. [43] The Lens, Lens.org, https://www.lens.org/lens/user /subscriptions (2026), accessed: 22 February 2026. [44] USPTO, Uspto bulk data directory, https://data.usp to.gov/bulkdata/datasets (2026), accessed: 22 Febru- ary 2026. [45] T. Zhou, J. Xu, G. Liu, J. Liu, H. Wang, and E. Wu, An approach for systematic decomposition of complex LLM tasks, arXiv preprint arXiv:2510.07772 (2026), intro- duces ACONIC framework showing decomposition im- proves agent performance on complex tasks. [46] S. Montazeri, Y. Feng, and K. Sha, PublicAgent: Multi- agent design principles from an LLM-based open data analysis framework, arXiv preprint arXiv:2511.03023 (2025), shows specialization provides value independent of model strength with 97.5% agent win rates. [47] Y. Yang, Q. Peng, J. Wang, Y. Wen, and W. Zhang, LLM-based multi-agent systems: Techniques and busi- ness perspectives, arXiv preprint arXiv:2411.14033 (2024), discusses dynamic task decomposition and or- ganic specialization advantages over single-agent sys- tems. [48] U.-e. Nisa et al., Agentic ai: The age of reasoning—a review, Journal of Artificial Intelligence (2025), a survey of agentic AI systems and their reasoning capabilities. [49] D. M. Lowe, P. T. Corbett, P. Murray-Rust, and R. C. Glen, Chemical name to structure: Opsin, an open source solution, J Chem Inf Model 51, 739 (2011). .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint S1 Supplementary Information HAR VEST: Unlocking the Dark Bioactivity Data of Pharmaceutical Patents via Agentic AI Appendix S1: Supplementary tables Target Class Proteins Novel Compounds Novel Clusters Enzyme Kinase 433 235,302 117,642 Transferase 256 56,736 28,143 Protease 221 44,680 23,166 Oxidoreductase 206 35,676 18,692 Other 154 35,461 18,335 Hydrolase 244 30,281 15,282 Unspecified 115 15,267 7,960 Phosphatase 61 7,345 3,884 Lyase 45 1,233 714 Membrane receptor Family A G protein-coupled receptor 387 122,178 59,877 Other 64 28,526 13,787 Unspecified 68 13,054 6,321 Ion channel Voltage-gated ion channel 98 23,129 11,354 Ligand-gated ion channel 53 18,263 8,982 Other 18 12,322 5,732 Unclassified protein 313 46,928 22,499 Transcription factor Nuclear receptor 64 24,212 11,638 Unspecified 70 18,099 8,796 Epigenetic regulator 108 39,917 20,460 Unclassified 755 33,066 15,806 Secreted protein 80 26,798 12,500 Transporter Electrochemical transporter 95 13,335 6,372 Other 33 5,134 2,429 Other 82 10,391 4,897 Other cytosolic protein 54 10,359 5,357 Total 4,077 907,692 450,625 TABLE S1: Distribution of novel protein–ligand interactions (PLIs) and structural clusters contributed by HARVEST, grouped by ChEMBL target classification. For each Level 1 (L1) target class (bold font) containing multiple Level 2 (L2) families, L2 subcategories are listed with indentation. Novel compounds/clusters include all ligands associated with HARVEST-only proteins or ligands not present in BindingDB for shared proteins. Structural clusters were defined using hierarchical complete-linkage clustering of Morgan fingerprints (radius = 2, 2048 bits) with a Tanimoto distance threshold of 0.2, followed by aggregation of adjacent clusters Appendix S2: Alternative Patent Data Sources Several alternative patent data sources were evaluated but found unsuitable for corpus-scale bioactivity extraction: SureChEMBL (https://www.surechembl.org) provides free access to patent-extracted chemical structure data through an open API [21]. However, tabular data in the processed output are concatenated without clear delimiters, making it difficult to distinguish decimal separators from column boundaries. The API also exhibits instability under sustained high-throughput querying, returning HTTP 500 errors at the request rates required for corpus-scale extraction. .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint S2 Google Patents(https://patents.google.com) andGoogle BigQuery Patents Public Data were both evaluated but found to be unsuitable: Google Patents implements aggressive rate limiting and IP-based blocking during programmatic access, while BigQuery, though well-structured, proved cost-prohibitive for extracting the complete pharmaceutical patent corpus with full-text descriptions in academic research budgets. Lens.org provides a more cost-effective alternative with institutional subscription options, but patent description fields are provided in plain text rather than structured markup, complicating the disambiguation of tabular data and chemical structure references essential for Document–Assay–Result–Compound–Protein relationship extraction. Appendix S3: Impact of CDX-based structure resolution TABLE S2: Cross-validation on shared patents without CDX-based chemical structure resolution (using only MOL files and py2opsin). Without the CDX parser, the full-match rate drops from 80.3% to 72.7% for HARVEST records and from 70.5% to 63.6% for BDB records (cf. Table I), while compound mismatches approximately double (from 7.9% to 15.5% and from 18.0% to 25.0%, respectively). These results demonstrate the improvement in extraction fidelity achieved by parsing chemical structures directly from the original ChemDraw binary files. HARVEST (n=603,929) in BDB BDB (n=690,907) in HARVEST Full match 439,113 (72.7%) 439,113 (63.6%) Target mismatch 57,079 (9.5%) 49,151 (7.1%) Compound mismatch 93,471 (15.5%) 172,515 (25.0%) No overlap 14,266 (2.4%) 30,128 (4.4%) Appendix S4: Cross-validation case studies TABLE S3: Manual review of the largest cross-validation mismatches (a): Target mismatch — the same compound was found in both databases but assigned to different proteins. Patent T arget UniProt ID (HAR VEST) UniProt ID (BindingDB) V erdict US20140275087A1 GlyT1SC6A9 HUMAN MGAT1 HUMANBindingDB incorrectly recognized a protein US20150376212A1 FLAPAL5AP HUMAN FEN1 HUMANBindingDB incorrectly recognized a protein US20240246937A1 IL4I1OXLA HUMAN IL1RA HUMANBindingDB incorrectly recognized a protein US20250051309A1 MK2, p38αMAPK2 HUMAN, MK14 HUMAN MK01 HUMAN, 4EBP1 HUMAN BindingDB incorrectly recognized a protein US20170226089A1 NR2BNMDZ1 HUMAN; NMDE1 HUMAN; NMDE2 HUMAN; NMDE3 HUMAN; NMDE4 HUMAN; NMD3A HUMAN; NMD3B HUMAN RXRA HUMANHARVEST and BindingDB incorrectly recognized a protein US20150191464A1 IRAK4IRAK4 HUMAN, TLR2 HUMAN IRAK4 HUMANHARVEST incorrectly recognized assay US20140057889A1 Bcl-XL, Bcl-2B2CL1 HUMAN, BCL2 HUMAN B2CL1 MOUSE, BCL2 MOUSE, BCL2 HUMAN No information about the organism US20120035149A1 PKR1, PKR2PKR1 HUMAN, PKR2 HUMAN G3HIE5 CRIGR, G3H407 CRIGR No information about the organism US20240059703A1 KRAS G12C,C118A (aa 1-169), His-tagged RASK HUMAN SOS1 HUMANHARVEST and BindingDB do not process protein modifications US20170334917A1 PG-K1 (aa101-181), His-tagged PLMN HUMAN TPA HUMANHARVEST and BindingDB do not process protein modifications .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint S3 TABLE S4:Manual review of the largest cross-validation mismatches (b): Compound mismatch — the same protein was found in both databases but associated with different compound sets. Patent H B Ovlp V erdict US20230255941A1 44 953 33 HARVEST incomplete extraction US20200095247A1 57 804 56 HARVEST incomplete extraction US20130172317A1 68 2136 52 HARVEST incomplete extraction US20220175775A1 79 1080 79 HARVEST incomplete extraction US20140094448A1 205 1116 179 HARVEST incomplete extraction US20240208941A1 89 332 89 HARVEST incomplete extraction US20190322661A1 32 462 30 HARVEST incomplete extraction US20240294506A1 2 499 2 HARVEST incomplete extraction US20250129020A1 201 454 197 HARVEST incomplete extraction US20240294524A1 84 783 68 HARVEST incomplete extraction (activity table not present in the XML package) US20240254137A1 37 892 4 Mixed case: HARVEST incomplete extraction and non-matching source documents US20190241588A1 754 766 59 HARVEST chemical structure error US20250017938A1 781 1 1 BindingDB incomplete curation US20160304519A1 515 48 48 BindingDB incomplete curation US20220041601A1 284 8 4 BindingDB incomplete curation US20120058988A1 562 94 92 BindingDB incomplete curation US20180118724A1 332 259 146 HARVEST extraction error US20130079303A1 322 318 115 HARVEST chemical structure error US20180273529A1 443 298 40 HARVEST chemical structure error US20230286970A1 393 405 392 Different inclusion threshold for inactive measurements TABLE S5: Manual review of the largest cross-validation mismatches (c): No overlap — neither compounds nor targets matched between databases for the same patent. Patent H B V erdict US20180118720A1 105 42 BindingDB patent mapping error US20210009607A1 138 25 BindingDB patent mapping error US20240343719A1 117 20 BindingDB patent mapping error US20200157106A1 5 800 BindingDB patent mapping error US20160115131A1 97 516 BindingDB patent mapping error US20210053946A1 8 612 BindingDB accession or identifier error US20190343826A1 37 608 BindingDB accession or identifier error US20210188777A1 163 7 BindingDB incomplete curation US20240209001A1 103 10 BindingDB incomplete curation US20220315606A1 429 3 BindingDB incomplete curation US20240400523A1 1 619 HARVEST incomplete extraction US20250064789A1 16 540 Mixed case: identifier mismatch and HARVEST incomplete extraction US20160060224A1 63 505 Technical identifier limitation US20230365594A1 214 8 Different assay focus across sources .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint