{"paper_id":"0d8cb4aa-fa91-4594-9fa8-0dacea37ccd5","body_text":"HAR VEST: Unlocking the Dark Bioactivity Data of Pharmaceutical Patents via\nAgentic AI\nViktoriia Shepard, Aibulat Musin, Kristina Chebykina, Natalia A. Zeninskaya,\nLukia Mistryukova, Konstantin Avchaciov, and Peter O. Fedichev ∗\nGERO PTE. LTD., 133 Cecil Street 14-01 Keck Seng Tower, Singapore 069535\nPharmaceutical patents contain vast Structure–Activity Relationship tables documenting protein–\nligand binding data that are technically public yet computationally inaccessible, rendering this\nwealth of data effectively dark — trapped in unstructured archives no existing database has sys-\ntematically captured. We present HARVEST, a multi-agent large language model pipeline that\nautonomously extracts structured bioactivity records from USPTO patent archives at $0.11 per\ndocument. Applied to 164,877 patents, HARVEST produced 3.36 million activity records, recov-\nering 365,713 unique scaffolds and 1,108 protein targets absent from BindingDB — completing\nin under a week a task requiring over 55 years of continuous expert labor. Automated extraction\nachieves 91% agreement with human curators while exhibiting lower unit-conversion error rates. We\nfurther introduce H-Bench, a structurally guaranteed held-out benchmark built from this recovered\ndata. Evaluation of the leading open-source model Boltz-2 on H-Bench reveals a two-dimensional\ngeneralization gap: performance degrades both on novel chemical scaffolds and on uncharacterized\nprotein targets, exposing fundamental limitations of models trained on existing public repositories.\nI. INTRODUCTION\nPharmaceutical patents represent one of the largest\nrepositories of experimental protein–ligand interaction\n(PLI) data ever assembled. Thousands of Structure–\nActivity Relationship (SAR) tables, each documenting\nbinding affinities across hundreds of compounds, are filed\nannually with patent offices worldwide and often appear\nyears before or independently of peer-reviewed litera-\nture [1–3]. Despite billions of dollars in R&D invest-\nment, this knowledge remains effectively “dark”: tech-\nnically public, yet computationally inaccessible, trapped\nin unstructured archives that no existing database has\nsystematically captured.\nThis data gap matters acutely. Recent breakthroughs\nin de novo protein design [4–6] and structure predic-\ntion [7–9] have transformed what AI can do in drug\ndiscovery, but these models face a generalization crisis.\nEven the best architectures struggle to predict activity in\nnew chemical or biological spaces when trained on sparse\ndata [10–12]. Closing this gap requires two things simul-\ntaneously: massive, diverse training sets and genuinely\nheld benchmarks to demonstrate robust model general-\nization [13–15]. Pharmaceutical patents could provide\nboth provided their contents were easily accessible.\nPLI data is the ground truth for both training and\nbenchmarking [16, 17]. The leading public repository,\nBindingDB, relies on the slow manual curation of the lit-\nerature [18] and covers only a fraction of the available\npatent data. Automating patent extraction has histori-\ncally failed due to the specific challenges of patent lan-\nguage and their multimodal nature [19, 20]: information\nis fragmented across unstructured text, complex tables,\n∗ Correspondence email address: peter.fedichev@gero.ai\nand chemical diagrams, and high-fidelity extraction de-\nmands reconstructing the complete link between a spe-\ncific compound, the assay performed, and the resulting\nactivity against a protein target [20]. Existing pipelines\nlike SureChEMBL index chemical structures at scale but\nlack systematic extraction of quantitative binding values\nor mapping to biological targets [21, 22]. The result is a\nself-reinforcing bottleneck: models are evaluated on the\nsame datasets they were trained on, making it impossible\nto distinguish genuine generalization from memorization.\nAgentic AI systems break this bottleneck. Decom-\nposing complex extraction into specialized sequential\nagents reduces hallucination rates and maintains accu-\nrate compound–target associations across documents ex-\nceeding 500,000 tokens. With the rapid rise in LLM rea-\nsoning and falling inference costs [23, 24], hierarchies of\nspecialized agents can now mimic expert human work-\nflows at negligible marginal cost – making systematic\npatent mining economically feasible for the first time.\nWe presentHARVEST(High-throughput Agent Re-\ntrieval of Values for Evaluated Small-molecules and Tar-\ngets), an automated multi-agent pipeline for the ex-\ntraction of SARs from USPTO bulk data. The sys-\ntem autonomously parses patent XML, resolves chem-\nical aliases to canonical SMILES, and maps biological\ntargets to UniProt identifiers. When applied to 164,877\npatent archives, the pipeline produced 3.36 million ac-\ntivity records from 40,902 patents at a cost of only $0.11\nper document – completing in under a week a task that\nwould require over 55 years of continuous manual ex-\npert labor. This dataset substantially expands the known\nchemical-biological landscape, recovering 365,713 unique\nscaffolds and 1,108 protein targets entirely absent from\nBindingDB.\nA central contribution of this work is H-Bench, an\nopen benchmark derived from HARVEST comprising\nbioactivity data absent from all existing public reposi-\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint \n\n2\ntories. H-Bench supports two distinct evaluation scenar-\nios: scaffold-generalization on known targets, and target-\ngeneralization across proteins with no prior public bioac-\ntivity data. Our evaluation of the leading open-source\nstructure-based model Boltz-2 [9] on H-Bench reveals a\ntwo-dimensional generalization gap: model performance\ndegrades both when chemistry is novel and when pro-\ntein targets lack prior bioactivity data – demonstrating\nthat current models have not yet learned fully transfer-\nable binding physics. Together, HARVEST and H-Bench\nconvert billions of dollars of inaccessible R&D knowledge\ninto open scientific infrastructure, directly addressing the\ndata bottleneck that limits AI-driven therapeutic discov-\nery.\nII. RESULTS\nA. HARVEST Substantially Expands Public\nBioactivity Space\nThe HARVEST pipeline (Section IX) was applied to\n164,877 USPTO patent archives pre-selected for the pres-\nence of chemical structures and bioactivity mentions.\nProcessing 50 documents in parallel, the system pro-\nduced a final dataset of 3.36 million activity records\nextracted from 40,902 patents (25% of the input cor-\npus), averaging 82 records per document containing ex-\ntractable data. The remaining patents either lacked ex-\ntractable bioactivity data or fell outside the pipeline’s\ncurrent parsing capabilities (see Section IV). The au-\ntomated pipeline achieves a consistently higher docu-\nment throughput than manual curation across the full\n25-year publication window examined (Fig. 1a). Al-\nthough BindingDB reports a higher average number of\nactivity records per patent (Fig. 1b), this is a result\nof human curators prioritizing the most data-rich doc-\numents. HARVEST achieves comparable yield on shared\npatents, confirming equivalent extraction depth. Because\nthe marginal cost of processing is negligible, HARVEST\ncaptures data from thousands of documents that would\nnot justify the cost of manual labor.\nOverall, HARVEST and BindingDB contain a com-\nparable total volume of protein–ligand interactions. Af-\nter aggregating multiple measurements per compound–\ntarget pair and applying inclusion filters (Section IX D),\nHARVEST yields 2.26M unique PLIs, comparable to\nBindingDB’s 2.21M across all sources (Fig. 1c). However,\nthese totals reflect fundamentally different source com-\npositions: BindingDB aggregates records from patents,\njournal articles, and ChEMBL, whereas HARVEST\ndraws exclusively from patent text. Restricting the\ncomparison to patent-derived records, HARVEST con-\ntributes nearly three times as many PLIs as BindingDB’s\npatent subset, confirming substantially deeper coverage\nof the patent corpus.\nAcross the combined chemical-biological landscape of\n8,710 protein targets, only 34.1% are shared between the\ntwo databases (Fig. 2a). A further 1,108 targets (12.7%)\nare covered exclusively by HARVEST, while the major-\nity of BindingDB-only targets (53.2%) reflect literature\nand ChEMBL sources outside the patent corpus. Crit-\nically, novelty extends well beyond HARVEST-exclusive\nproteins: for the 2,969 shared targets, 37.0% of HAR-\nVEST PLIs (851,642 interactions) and 43.4% of Murcko\nscaffold clusters (424,772 clusters) are absent from Bind-\ningDB (Fig. 2b–c). This indicates that substantial chem-\nical diversity remains trapped in patent archives even for\nthe most extensively studied drug targets.\nThe target distribution reflects established drug dis-\ncovery priorities: enzymes and kinases predominate due\nto their well-defined binding pockets [25], while transcrip-\ntion factors are less represented due to the difficulty of\ntargeting protein-protein interfaces with small molecules\n(see Table S1).\nTo assess the utility of this expanded chemical space\nfor structure-activity relationship (SAR) analysis, we\nquantified the density of activity cliffs. These are de-\nfined as pairs of structurally similar compounds (Tani-\nmoto similarity≥0.7) that exhibit a substantial differ-\nence in biological activity (∆pActivity ≥ 1.5), where\npActivity = − log10[M] [26]. Activity cliffs represent\nhigh-information data points where minor chemical mod-\nifications critically determine a molecule’s effect. Across\nthe 2,969 shared targets with at least one activity cliff,\nHARVEST provides greater cliff density for 42% of pro-\nteins, while BindingDB leads for 54%; only 3% show\nequivalent coverage. This asymmetric complementarity–\nwhere each resource uniquely enriches SAR information\nfor distinct protein subsets–argues strongly for merg-\ning both datasets to maximize the structural discontinu-\nities available for lead optimization and machine learning\nmodel training.\nB. HARVEST Extracts Data Comparable in\nQuality to Manual Curation\nTo validate the fidelity of automated extraction, we\nbenchmarked HARVEST against the manually curated\nBindingDB (BDB) dataset [18]. Since HARVEST oper-\nates exclusively on US patents, all comparisons use the\npatent-derived subset of BDB unless otherwise noted.\nDensity distributions of binding affinity, molecular\nweight, and synthetic accessibility show close agreement\nbetween HARVEST and BDB across the full range of\nvalues (Figs. 3a–3c), consistent with previously reported\npatent-derived compound profiles [18]. We next assessed\nrecord-level accuracy by pairwise comparison of activ-\nity values for PLIs present in both databases, matched\nby UniProt accession and InChIKey connectivity layer\n(first block) [27]. Across 319,954 matched PLIs from\n5,668 shared patents, the distribution of activity residuals\n(∆pActivity) is highly centered around zero, with 91.0%\nof PLIs showing near-identical values (Fig. 4a). This cor-\nresponds to a high quantitative correlation (Pearson r =\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint \n\n3\nFIG. 1:(a) Number of patents included per year from USPTO for HARVEST (green) and BindingDB (orange), illustrating\nthe consistently higher document coverage achieved by automated extraction across the full 25-year window. (b) Mean\nnumber of activity records per document for HARVEST (green), BindingDB (orange), and their intersection (blue). The\nconvergence of HARVEST and BindingDB on shared patents confirms equivalent extraction depth; the lower overall\nHARVEST average reflects broader document selection criteria that include patents with fewer activity records. (c)\nComparison of unique protein–ligand interactions (PLIs) between HARVEST and BindingDB. The BindingDB bar is\ndecomposed by data source: ChEMBL-derived entries (orange), US Patent extractions (blue), and remaining sources (pink).\nHARVEST yields 2.26M unique PLIs from patent text alone, exceeding the 2.21M across all BindingDB sources combined.\nFIG. 2:Comparison of protein and chemical space coverage between HARVEST and BindingDB. (a) Protein target\ncomposition across both databases (n = 8,710 total): 34.1% of targets are shared, while 12.7% are covered exclusively by\nHARVEST and 53.2% exclusively by BindingDB. (b) Protein–ligand interaction (PLI) diversity within HARVEST for the\n2,969 shared protein targets: 37.0% of PLIs are novel relative to BindingDB, with an additional 2.4% associated with proteins\nunique to HARVEST. (c) Murcko scaffold cluster diversity within HARVEST for shared targets: 43.4% of scaffold clusters are\nabsent from BindingDB, with a further 2.6% belonging to HARVEST-exclusive proteins. Orange indicates novel content\nunique to HARVEST; blue indicates overlap with BindingDB; green indicates targets absent from BindingDB entirely.\n0.925, Spearmanρ= 0.937). Validation against 68,209\nindependent article-derived records confirms this consis-\ntency (r= 0.851, ρ = 0.875), with the broader residual\ndistribution reflecting inherent experimental variability\nbetween separate data sources (Fig. 4b).\nThe residual analysis reveals isolated spikes at\n∆pActivity = ±3, the signature of 1,000-fold unit conver-\nsion errors (nM/µM confusion), the most common cura-\ntion artifact reported by the BindingDB team [18]. These\naffect ∼1.4% of patent PLIs and ∼1.2% of article PLIs.\nTo determine which database held the correct value,\nwe manually verified the original patent text for the 20\npatents contributing the most discrepant records, collec-\ntively covering 3,706 of the 5,499 affected PLIs (67%).\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint \n\n4\n(a)\n (b)\n (c)\nFIG. 3:Physicochemical property distributions for HARVEST and BindingDB (US patents subset). (a) Binding affinity\ndistributions (combined IC 50,K i,K d, and EC 50). Only exact numeric measurements (relation “=”) are included. (b)\nMolecular weight distributions, with both peaking near 450 Da. (c) Synthetic accessibility (SA) scores, both peaking at\nSA≈3. The alignment across three descriptors confirms that HARVEST extracts a representative chemical space without\nsystematic physicochemical bias relative to manual curation.\nFIG. 4:Distribution of activity residuals (∆pActivity = BDB−HARVEST) for matched compound–target pairs. The y-axis\nis broken to show both the dominant central peak and the tail structure. (a) US patent-derived BindingDB records\n(n= 319,954): 91.0% of pairs fall in the central bin (∆pActivity ≈ 0), corresponding to the high correlation (r = 0.925)\nobserved between manual and automated curation. (b) Article-derived BindingDB records (n = 68,209): the broader\ndistribution (59.3% central bin, r = 0.851) reflects genuine measurement variability across independent data sources rather\nthan extraction error. In both panels, the distinct spikes at ∆pActivity = ±3 identify 1,000-fold unit conversion errors\n(nM/µM confusion), the most frequent artifact in both curation workflows. The sharp central distribution confirms that\nHARVEST achieves human-level extraction fidelity across hundreds of thousands of records.\nAt the record level, HARVEST held the correct value in\n92% of verified PLIs, BindingDB in 5%, with 3% remain-\ning ambiguous. This indicates that automated agents\nare substantially less prone to unit-conversion errors than\nmanual curators.\nThe preceding comparison is restricted to records\npresent in both databases. To assess record-level over-\nlap, we cross-referenced each record against the other\ndatabase within the same patent (Table I). Of 606,456\nHARVEST records, 80.3% find an exact match in Bind-\ningDB; conversely, HARVEST recovers 70.5% of 690,869\nBindingDB records. To understand the sources of dis-\ncrepancy, we manually verified the original patent text\nfor the patents contributing the most records in each\nmismatch category (Tables S3–S5), collectively covering\n12.1% of all mismatched records.\nTarget mismatch. When both databases find the\nsame compound but assign it to different proteins, HAR-\nVEST more often identified the correct target. In the\nremaining cases, HARVEST misclassified assay readout\nbiomarkers as direct binding targets. Additional dis-\nagreements arose from differences in protein name reso-\nlution and from unclear species assignment in the patent\ntext.\nCompound mismatch. Differences in compound\nsets mostly reflect extraction coverage. On large patents\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint \n\n5\nTABLE I: Cross-validation on 5,668 shared patents. Each\nrow indicates how a record from one database was matched\nin the other on the same patent. Full match : both\ncompound and target agree. Target mismatch: same\ncompound found but assigned to a different protein.\nCompound mismatch : same target found but linked to a\ndifferent compound. No overlap : neither compound nor\ntarget found on that patent.\nHARVEST\n(n=606,456)\nin BDB\nBDB\n(n=690,869)\nin HARVEST\nFull match 486,950 (80.3%) 486,950 (70.5%)\nTarget mismatch 61,823 (10.2%) 53,784 (7.8%)\nCompound mismatch 47,685 (7.9%) 124,061 (18.0%)\nNo overlap 9,998 (1.6%) 26,074 (3.8%)\n(500+ examples), HARVEST sometimes extracts fewer\ncompounds than BindingDB. On small and medium\npatents, the opposite holds: HARVEST captures data\nformats that manual curators skip, including inline activ-\nity values in synthesis text, semi-quantitative statements\n(“IC50 <5µM for all 590 examples”), and non-numeric\nactivity codes (letter grades, symbolic ratings, +/++\nscales) used in place of exact measurements. Additional\ndifferences arose from HARVEST extraction or structure-\nnormalization errors. Overall, HARVEST yields some-\nwhat fewer PLIs on shared patents (606K vs. 691K), pri-\nmarily due to incomplete extraction from the largest doc-\numents.\nNo overlap.The rare cases (1.6–3.8%) where nei-\nther compounds nor targets overlap are often caused by\nBindingDB patent mapping or identifier errors, or by the\nsame target mismatch and compound extraction factors\ndescribed above.\nC. H-Bench: A Public Benchmark Dataset\nA central deliverable of this work isH-Bench, an open\nbenchmark of bioactivity data extracted by HARVEST\nthat is not present in BindingDB. BindingDB serves as\nthe primary training source for most modern bioactivity\nmodels, either as a direct source [9] or through its inte-\ngration into the PDB [28]. H-Bench therefore provides\na structurally novel held-out resource for rigorous model\nevaluation on records likely omitted from training sets.\nTo establish the benchmark’s integrity, we used a\ngraph-based integer linear program (ILP) to maximize\nstructural separation between HARVEST compounds\nand existing records. This process identified two distinct\nsubpopulations (Fig. 5): the Valid subset (n = 98, 105,\nmedian similarity ≈ 0.47), occupying novel chemical\nspace, and the Common subset (n = 245, 836, median\nsimilarity ≈ 0.70), containing compounds structurally\ncloser to BindingDB entries. We provide a Python script\nthat implements this algorithm, allowing researchers to\ncompare H-Bench against their own training data. This\nFIG. 5:Chemical distance of H-Bench compounds from\nBindingDB, measured as maximum Tanimoto similarity to\nthe nearest BindingDB compound. (a) Valid subset\n(n = 98,105; median = 0.47): compounds with low\nstructural overlap with BindingDB, released as the H-Bench\nbenchmark. (b) Common subset (n = 245,836; median\n= 0.70): compounds structurally proximal to BindingDB\nentries.\ntool automatically identifies and moves any similar struc-\ntures into the Common subset, effectively censoring po-\ntential data leakage and preventing over-optimistic re-\nsults during model evaluation.\nH-Bench covers 53 protein targets spanning diverse\nclasses (Supplementary Table S6), including enzymes,\nmembrane receptors, and transcription factors. Of these,\n44 targets overlap with BindingDB and 9 are entirely\nnovel. To ensure reliable evaluation, targets were filtered\nto maintain an activity balance between 20% and 55%\n(mean ≈ 33% active, Supplementary Table S6), avoiding\nthe extreme class imbalance common in public screen-\ning datasets. These two components, the set of entirely\nnew proteins and the novel structural clusters for overlap-\nping targets, support two primary evaluation scenarios:\ngeneralization to novel chemical structures on known tar-\ngets and generalization to entirely uncharacterized pro-\ntein targets.\nTo investigate whether current high-performance mod-\nels learn transferable binding physics or simply rely on\nstructural similarity to their training data, we evaluated\nthe leading open-source structure-based model Boltz-2 [9]\non the H-Bench benchmark. This evaluation uses up to\n120 ligands per protein, balanced between active and in-\nactive classes. For the Common subset evaluation, we\nexcluded molecules with Tanimoto similarity above 0.75\nto BindingDB to mitigate direct memorization effects.\nOf the three Boltz-2 output scores evaluated,\naffinity\npred valueperformed best overall. Perfor-\nmance varied systematically across the three bench-\nmark subsets: AUC-ROC increased monotonically from\nValid/new targets (novel proteins, novel chemistry;\nAUC≈ 0.52) to Valid/known targets (known proteins,\nnovel chemistry; AUC ≈ 0.63) to Common/known tar-\ngets (known proteins, structurally familiar chemistry;\nAUC ≈ 0.70), directly mirroring proximity to train-\ning data (Fig. 6a). The score intended for hit discov-\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint \n\n6\nFIG. 6:Evaluation of Boltz-2 predicted binding scores against experimental activity on H-Bench. (a) Per-protein AUC-ROC\nfor three Boltz-2 output scores (mean±SEM across proteins), evaluated on three benchmark subsets that differ in both\nchemical and target novelty: Valid/new targets (9 proteins exclusive to HARVEST, novel chemistry); Valid/known targets\n(44 proteins shared with BindingDB, novel chemistry); and Common/known targets (44 proteins shared with BindingDB,\nchemistry structurally proximal to BindingDB). The dashed line marks random performance (AUC = 0.5). Performance\nincreases monotonically with proximity to training data across all three scores, revealing that Boltz-2 predictions reflect\nstructural familiarity more than transferable binding physics. (b) Experimental pActivity (-log 10[M]) vs. predicted\naffinity pred valuewith linear fits and Pearson r, for Valid/known (orange) and Common/known (blue) subsets. (c)\npActivity vs. affinity prob binary. Points in (b, c) are colored by Tanimoto similarity to the nearest BindingDB compound.\nery,affinity prob binary, performed near-randomly\nacross all subsets (AUC≈0.52–0.58), indicating a signif-\nicant challenge for virtual screening applications where\nthe model must distinguish binders from decoys. The\nconfidence\nscoremetric, which reflects structural plau-\nsibility of predicted complexes, showed a similar gradient\nbut remained near random on genuinely novel chemistry.\nA key finding is the systematic performance gap be-\ntween the two subsets: points colored by Tanimoto sim-\nilarity cluster toward higher predicted scores as similar-\nity increases (Fig. 6b,c). This reveals that model pre-\ndictions are often biased by proximity to training data\nrather than underlying binding physics. Scatter analysis\nconfirms only a weak but statistically significant correla-\ntion between predicted affinity and experimental activity\n(Pearsonr≈ −0.31 for the Valid set).\nOverall, Boltz-2 performance on H-Bench is modest\nbut above random on known targets, and degrades on\nproteins entirely absent from public bioactivity reposi-\ntories, which is consistent with results reported on the\nmodel’s own proprietary benchmarks [9]. This two–\ndimensional generalization gap — across both chemical\nand target space — confirms that H-Bench provides a\nstringent and unbiased evaluation resource. The results\nsuggest that current structure-based models have not yet\nlearned fully transferable binding physics, and that both\nnovel chemical scaffolds and uncharacterized targets re-\nmain a fundamental challenge for AI-driven drug discov-\nery.\nIII. DISCUSSION\nThe pharmaceutical industry has invested hundreds of\nbillions of dollars in protein–ligand interaction experi-\nments over the past three decades. Patent law was de-\nsigned to make this knowledge public, yet the practi-\ncal inaccessibility of unstructured patent archives has\nmeant that this investment remained effectively private–\nconfined to commercial databases behind expensive sub-\nscriptions or simply uncurated. HARVEST transforms\nthis situation: by processing the full USPTO pharma-\nceutical corpus in under a week at $0.11 per document,\nit demonstrates that the era of dark bioactivity data is\nending. The 3.36 million activity records recovered, in-\ncluding 365,713 structural clusters and 1,108 protein tar-\ngets absent from BindingDB, represent a qualitative ex-\npansion of the computable chemical-biological landscape\navailable to the global research community.\nThis capability distinguishes HARVEST from all ex-\nisting approaches along two dimensions simultaneously:\nscale and semantic depth. SureChEMBL provides a high-\nvolume index of chemical structures [21]; however, with-\nout quantitative binding context or protein identity reso-\nlution it answers “what molecules appear in patents” but\nnot “what do they do and against which target.” Bind-\ningDB provides exactly that semantic depth but is funda-\nmentally constrained by the throughput of human expert\ncuration [18]. HARVEST achieves BindingDB-level ex-\ntraction fidelity – 91% agreement on matched records,\nwith lower unit conversion error rates than manual cura-\ntors – at 3,500 times the throughput. Recent LLM-based\nsystems like BioMiner [29] and BioChemInsight [30] rep-\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint \n\n7\nHAR\nVEST BioMiner\n[29]\nBioChemInsight\n[30]\nSource\nPatents Articles Patents\nDocuments\nprocessed 164,877 11,683 181\nDocuments with\nbioactivity 40,902 11,683 N/A\nRecords extracted 3.36M 68K N/A\nThroughput∗ 2 sec/doc 14 sec/doc N/A\nSMILES Yes Yes Yes\nProtein ID UniProt AlphaFold path N/A\nUnit normalization Yes No No\nPublic dataset Partially Yes No\nTABLE II: Comparison of automated bioactivity\nextraction systems. ∗Throughput measured under different\nconditions: HARVEST uses cloud-based LLM API with\n50-document parallelism; BioMiner reports 14s/paper on\n8×V100 GPUs. Only HARVEST performs full resolution to\nUniProt identifiers and standardized units.\nresent important progress but remain restricted to scien-\ntific articles or small-scale proofs-of-concept. HARVEST\nis the first system to demonstrate this combination of fi-\ndelity and scale across a full national patent corpus (Ta-\nble II).\nBeyond raw extraction, a significant advantage of\nHARVEST is its automated data normalization. The\nsystem resolves varied and often ambiguous protein de-\nscriptors to canonical UniProt identifiers [31] and stan-\ndardizes diverse activity units into a consistent numeric\nformat. By producing a dataset structured similarly to\nBindingDB, HARVEST provides a machine-actionable\nresource that is immediately ready for training machine\nlearning models. This eliminates the massive manual\npost-processing and data-cleaning efforts typically re-\nquired when working with “dirty” automated extractions\nfrom patent literature.\nThe cost profile of HARVEST fundamentally changes\nthe economics of medicinal chemistry data. Traditional\ncuration projects like BindingDB process approximately\n1, 500 patents over 2 years [18]; at that rate, processing\nour full corpus would require over 55 years of contin-\nuous expert labor. HARVEST completed this task in\nunder a week (Table II). This efficiency removes the fi-\nnancial barriers that have long confined large-scale bioac-\ntivity data to expensive commercial platforms such as\nReaxys [32, 33] or GOSTAR [34]. For approximately\nthe cost of six months’ worth of a commercial sub-\nscription, any research organization can now generate\na proprietary-scale dataset in a matter of weeks. This\ndemocratization of data allows academic groups to com-\npete in a landscape previously dominated by well-funded\ncommercial providers. Critically, this cost structure also\nchanges the update cycle for bioactivity data. Bind-\ningDB’s manual curation introduces a lag of years be-\ntween patent publication and database inclusion [18].\nHARVEST can in principle be rerun on new USPTO\nweekly releases continuously, maintaining a near-real-\ntime mirror of the public patent bioactivity landscape\nFIG. 7:Cost breakdown for processing the USPTO patent\ncorpus with HARVEST, including development and testing\nexpenses. The production cost of$0.11 per document\nexcludes these one-time development costs.\n– something no manual system can achieve regardless of\nfunding.\nA striking finding is that despite recovering 1,108 pro-\ntein targets entirely absent from BindingDB, these novel\ntargets account for only 2.4% of total extracted interac-\ntions. The vast majority of new data deepens coverage of\nestablished targets rather than revealing entirely new bio-\nlogical associations. This pattern likely reflects two con-\nverging forces: the pharmaceutical industry’s strategic\nfocus on validated targets where biological risk is lower,\nand the temporal lag between patent filing and scientific\npublication.\nThe release of H-Bench addresses a fundamental eval-\nuation problem in AI-driven drug discovery. Because\nmodern models are often trained on the same core pub-\nlic datasets, it is difficult to distinguish genuine gener-\nalization from memorization [10, 11]. H-Bench provides\na structurally guaranteed held-out resource because its\nrecords are derived from patent literature currently ab-\nsent from BindingDB. Our three-way evaluation of Boltz-\n2 [9] – across novel targets, known targets with novel\nchemistry, and known targets with familiar chemistry –\nreveals that the generalization gap is two-dimensional:\nmodels degrade both when chemistry is novel and when\ntargets lack prior bioactivity data. This is a more pre-\ncise characterization of model limitations than binary\ntrain/test splits on BindingDB alone can provide, and\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint \n\n8\nit points directly toward what the field needs to improve:\ntraining data that covers undercharacterized targets, and\nevaluation frameworks that separately stress-test chemi-\ncal and biological generalization. H-Bench provides both.\nThe success of HARVEST relies heavily on the high\nquality of structured data provided by the USPTO. Be-\ncause the USPTO includes chemical structures as em-\nbedded ChemDraw files within its XML archives, they\ncan be reliably converted into canonical SMILES. In con-\ntrast, processing raw PDF documents would require com-\nplex OCR and significantly more intensive processing,\nwhich often introduces errors. This highlights the criti-\ncal importance of maintaining publicly available data in\nwell-organized, machine-readable formats to enable au-\ntomated discovery.\nThe 2025 announcement by the CNIPA in China to\npromote XML for electronic patent submissions [35]\nopens a direct pathway to extend HARVEST to Chinese\npharmaceutical literature, which would add a massive\nand currently inaccessible reservoir of medicinal chem-\nistry data. More broadly, the value of open, structured\ndata formats for enabling downstream scientific compu-\ntation cannot be overstated: the difference between a\nPDF and an XML archive is the difference between dark\ndata and actionable knowledge.\nThe principles demonstrated by HARVEST extend far\nbeyond bioactivity extraction. Much of humanity’s ex-\npert knowledge remains practically inaccessible: techni-\ncally public in patents, regulatory filings, and clinical\nrecords, yet effectively “dark” due to the prohibitive cost\nof human curation. The multi-agent architecture intro-\nduced here, which decomposes complex document un-\nderstanding into specialized sequential agents grounded\nby structured resolution pipelines, is directly applicable\nto any domain facing this barrier. Recent examples in-\nclude agentic systems for multilingual pharmaceutical as-\nset scouting [36, 37] and automated extraction from clin-\nical and regulatory documents [38, 39]. As LLM capabili-\nties improve and inference costs decline [24], the marginal\ncost of converting massive document corpora into struc-\ntured knowledge is approaching zero. The critical ques-\ntion is no longer whether this conversion is feasible, but\nwhich datasets are most vital to recover and how to en-\nsure the resulting knowledge remains a public good rather\nthan a proprietary asset.\nIV. LIMITATIONS\nHARVEST’s current scope defines a clear roadmap for\nfuture development. Four categories of patent data re-\nmain to be addressed. First, the system cannot yet\nprocess Markush structures. These are highly complex\ndiagrams used to represent large families of molecules\nsimultaneously via variable parts. Automatically “un-\npacking” these structures into a list of specific, individ-\nual compounds is a significant technical challenge that\nis not yet implemented, although these cases account for\nonly 8% of the data points we excluded (Section IX D).\nSecond, graphical data such as dose–response curves re-\nmain inaccessible, preventing the extraction of param-\neters from assays reported only in figure form. Third,\nprotein target resolution is limited by the scope of the\ncurated Swiss-Prot database [40] and by the inherent\nambiguity of multi-subunit complexes like integrins or\nIL-23. Patents often reference these targets using incon-\nsistent subunit, heterodimer, or domain-level descriptors,\nmaking canonical mapping difficult. Finally, LLM safety\npolicies occasionally caused the system to refuse patents\ntargeting high-risk pathogens such as Ebola virus. This\nhas resulted in systematic coverage gaps in certain antivi-\nral research areas, where the model interprets the data\nextraction as a violation of safety guidelines.\nOur cross-validation is grounded in the 5,668 patents\nshared with BindingDB, representing 14% of the HAR-\nVEST corpus. For the remaining 86%, no indepen-\ndent reference currently exists – establishing such a ref-\nerence, through prospective experimental validation of\nHARVEST-derived predictions or through community\ncuration efforts, is a priority for future work. Based on\nthe consistency of activity residual distributions between\nvalidated and unvalidated subsets, we estimate the error\nrate for uncharacterized records at level not exceeding\n10-15%, comparable to known error rates in manually\ncurated databases [18].\nV. CONCLUSION\nWe have presented HARVEST, a multi-agent LLM\npipeline that converts the dark bioactivity data of phar-\nmaceutical patents into open, computable scientific in-\nfrastructure. By decomposing complex document un-\nderstanding into specialized sequential agents, HAR-\nVEST achieves BindingDB-level extraction fidelity at\n3,500 times the throughput of manual curation – pro-\ncessing 164,877 patent archives in under a week at$0.11\nper document, recovering 365,713 structural clusters and\n1,108 protein targets entirely absent from public reposi-\ntories. At this cost, comprehensive patent mining is no\nlonger a luxury for well-funded commercial providers; it\nis accessible to any research group with a compelling sci-\nentific question.\nThe accompanying H-Bench benchmark addresses an\nequally critical problem: the lack of genuinely held-\nout evaluation data for protein–ligand interaction mod-\nels. Our three-way evaluation of Boltz-2 [9] reveals that\nthe generalization gap is two-dimensional – model per-\nformance degrades both when chemistry is novel and\nwhen protein targets lack prior public bioactivity data.\nThis finding exposes a fundamental limitation of models\ntrained exclusively on manually curated repositories, and\nestablishes H-Bench as a stringent, leakage-free resource\nfor driving the development of more robust, physics-\naware models.\nTogether, HARVEST and H-Bench represent a prac-\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint \n\n9\ntical answer to the data bottleneck that currently limits\nAI-driven drug discovery. The hundreds of billions of dol-\nlars in R&D investment embedded in the global patent\ncorpus was always legally public; it was never compu-\ntationally accessible. As LLM inference costs continue\nto fall [24], the same approach can be extended to new\npatent jurisdictions, to regulatory filings, and to any do-\nmain where expert knowledge remains trapped in un-\nstructured text. The era of dark data in medicinal chem-\nistry is ending.\nVI. ACKNOWLEDGMENTS\nWe thank Mikhail Batin, Alexey Strygin, and Vita\nStepanova, the organizers of the Agentic AI Against Ag-\ning (AAAA) hackathon, for providing the venue that fa-\ncilitated the inception of this work. We are also deeply\ngrateful to Vladimir Manujlov, an employee of Gero, for\nhis assistance and guidance during the hackathon and for\nhis valuable edits and review of this manuscript. We ex-\ntend our gratitude to Daniel Kravtsov for providing his\ntechnical expertise in modern agentic systems, which of-\nfered helpful insights during the design of the HARVEST\npipeline. Finally, we acknowledge the use of the Gemini\nLarge Language Model for assistance in the drafting and\nlinguistic refinement of this manuscript.\nVII. CONFLICTS OF INTEREST\nK.A., L.M. and P.F. are employees and equity hold-\ners of Gero PTE. LTD., a company developing AI-driven\ndrug discovery tools. Gero proposed the patent mining\nchallenge at the AAAA hackathon and may benefit com-\nmercially from the methods and datasets described in this\nwork. V.S., A.M., K.C., and N.A.Z. were selected as the\nwinning implementation team by the independent AAAA\norganizing committee and subsequently contributed to\nthis work under a contract with Gero. All authors de-\nclare no other competing financial interests.\nVIII. DATA AVAILABILITY\nThe H-Bench benchmark and the full HARVEST\ndataset are available at https://github.com/gero-s\ncience/HARVEST under the Creative Commons Attri-\nbution 4.0 International (CC BY 4.0) license. The\nChemDraw binary file reader is available at https:\n//github.com/gero-science/cdx_reader.\nThis work also utilizes data from the BindingDB open-\nsource database (September 2025 release), which can be\naccessed at https://www.bindingdb.org. The raw\npatent application data used for extraction was obtained\nfrom the USPTO Patent Application Full Text Data with\nEmbedded TIFF Images (APPDT), available at https:\n//data.uspto.gov/bulkdata/datasets/appdt.\nDetailed dataset statistics, protein family distribu-\ntions, and activity label balance metrics are provided in\nthe Supplementary Information.\nIX. MATERIALS AND METHODS\nA. Patent Data Sources\nWe evaluated several repositories for large-scale bioac-\ntivity extraction, prioritizing structured data formats,\nAPI stability, and cost. Alternative sources such as\nSureChEMBL [21], Google Patents [41], Google Big-\nQuery Patents Public Data [42] and Lens.org [43] were\nexcluded due to limitations in their data formats (see\nSupplementary Section S2 for details).\nUSPTO Bulk Data [44] was selected as the primary\ndata source, which provides mostly unlimited downloads\nsupported by an API without any additional costs. The\nstructured XML format of the provided documents pre-\nserves table hierarchies and chemistry tags, which is es-\nsential for accurately linking compounds to their biolog-\nical targets and activity values. This high-density dis-\nclosure is a major advantage of the patent corpus: US\npatents contain an average of 160 measurements per doc-\nument compared to only 40 per scientific article [18].\nWe retrieved weekly archives from the USPTO Ap-\nplication Data (APPDT) using the Bulk Datasets API.\nTo ensure the corpus was rich in extractable PLI data,\nwe applied an initial filter to retain only documents\ncontaining either embedded chemical structure attach-\nments (CDX/MOL) or specific bioactivity keywords (e.g.,\n”IC50”, ”Ki”, ”Kd”, ”EC50”) identified via regular ex-\npression matching. The keyword filter was optimized\nfor high recall, minimizing false negatives at the cost\nof increased false positives that are subsequently fil-\ntered by downstream agents. This pre-filtering is sup-\nported by our empirical observation that patents lacking\nCDX/MOL attachments rarely contain structured quan-\ntitative bioactivity data.\nB. Patent Family Deduplication\nA single invention can generate multiple legally dis-\ntinct application records, such as pre-grant publications,\ngranted patents, continuation or divisional applications,\nand reissue or reexamination documents. These filings\nfrequently contain identical bioactivity tables and chem-\nical examples, which can lead to redundant data extrac-\ntion and inflated record counts. To ensure each unique\nexperimental observation is represented only once, we ag-\ngregated related filings into continuity clusters.\nWe constructed these clusters by building a directed\ngraph of parent–child relationships retrieved via the\nUSPTO API. To specifically identify content duplication,\nwe restricted the graph edges to priority claims and conti-\nnuity links: “is a continuation of,” “is a divisional of,” “is\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint \n\n10\nFIG. 8:The HARVEST pipeline architecture. USPTO patent documents are parsed by an XML extractor and passed to a\nthree-stage LLM extraction module: Agent 1 identifies biological targets and assay conditions; Agent 2 extracts quantitative\nbioactivity measurements; and Agent 3 resolves compound aliases to IUPAC names or embedded chemical identifiers. The\nextracted records are then processed by two resolution agents operating in parallel: Agent 4 converts chemical identities to\ncanonical SMILES via embedded structure files and py2opsin name-to-structure conversion; Agent 5 maps protein names to\nUniProt identifiers using UniProt FASTA. The final output is the normalized HARVEST dataset of Document–Assay–Result–\nCompound–Protein records.\na national stage entry of,” and “is a reissue of.” We ex-\nplicitly excluded “continuation-in-part” (CIP) relation-\nships from this automated deduplication, as CIP filings\noften introduce new substantive data not present in the\nparent application.\nFor each resulting cluster, we selected the most re-\ncent document by publication date as the representative\nrecord. This strategy ensures the capture of the most\ncomplete version of the disclosure, as later filings often\ninclude corrected tables or refined IUPAC nomenclature.\nThis deduplication process yielded a 17.7% reduction in\nthe total document volume, resulting in a final corpus of\nunique inventions for bioactivity extraction.\nC. Multi-Agent Architecture\nTo extract PLIs from linguistically complex patent\ndocuments, we developed a multi-agent architecture con-\nsisting of five specialized agents operating in sequence\n(Fig. 8). We adopted this multi-stage decomposition be-\ncause LLMs often exhibit performance loss when required\nto identify heterogeneous data types simultaneously [23].\nOur preliminary experiments on 500 patents confirmed\nthis limitation, as a monolithic single-prompt strategy\nresulted in three systematic failure modes.\nFirst,attention dilution caused the model to lose con-\nsistent compound–target associations in documents ex-\nceeding 200,000 tokens, leading to misattributed activity\nvalues. Second, task interference substantially increased\nSMILES hallucination rates when chemical names, bio-\nlogical targets, and numeric values were requested simul-\ntaneously. Third, output truncation caused premature\nresponse termination in patents containing over 500 ac-\ntivity records, losing a substantial fraction of data.\nDecomposing the extraction into sequential agents ad-\ndresses these three failure modes by narrowing the se-\nmantic scope of each LLM call. This approach is con-\nsistent with recent findings that task decomposition\nand agent specialization improve performance in fron-\ntier models [45–48]. The sequential architecture also im-\nproves system traceability and enables targeted prompt\noptimization for each subtask. Furthermore, this design\nprovides an early termination mechanism: if Agent 1\nidentifies no biological targets, subsequent steps are\nskipped to avoid unnecessary computation on irrelevant\npatents.\nThe pipeline utilizes three LLM-based agents\n(Agents 1–3) for semantic extraction, followed by two\nresolution agents (Agents 4–5) for chemical and protein\nstandardization. All agents operate within an asyn-\nchronous processing framework with configurable worker\npools, enabling high-throughput parallel processing of\nthe patent corpus.\nAgents 1–3: Semantic Extraction\nAgent 1 (Target Extraction) identifies biological\ntargets such as proteins, enzymes, and receptors, along\nwith test organisms and assay conditions. This extracted\ncontext is injected into the subsequent Agent 2 prompt\nto reduce misattribution errors in documents describing\nmultiple targets.\nAgent 2 (Activity Extraction) focuses on quanti-\ntative measurements. It extracts compound aliases (e.g.,\n“Example 1”, “Compound 42”), binding metrics (IC 50,\nKi, K d, EC50), numeric values, and measurement units.\nBy embedding the output of Agent 1, Agent 2 performs\ntarget-aware extraction, ensuring that each numeric re-\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint \n\n11\nsult is correctly linked to its respective assay and protein.\nAgent 3 (Compound Mapping) resolves the com-\nmon patent practice of referencing molecules by internal\naliases. It maps these aliases to IUPAC nomenclature or\nembedded chemical identifiers provided in the text. Iso-\nlating this task into a dedicated agent allows handling\nthe complexity: compound identities can be scattered\nacross hundreds of pages, requiring full-document con-\ntext awareness.\nThe pipeline utilizes three LLM-based agents to han-\ndle the initial semantic extraction from patent text.\nThese agents are built on google/gemini-2.5-flash,\nselected for its 1-million-token context window and effi-\ncient prompt caching.\nAgents 4–5: Resolution\nAgent 4 (Chemical Structure Resolution) con-\nverts resolved chemical identities into standardized\nSMILES strings. While USPTO patent XML embeds\nboth MOL and CDX (ChemDraw binary) files for all\nchemical compounds; we use CDX exclusively due to sys-\ntematic corruption discovered in the distributed MOL\nrepresentations. This corruption typically occurs dur-\ning default CDX-to-MOL conversion, where atom type\naliases or common substitutions (e.g., “Me”, “Et”, “HN”,\nor “CN”) are erroneously interpreted as carbon atoms.\nBy parsing the ChemDraw binary files directly, we avoid\nthese substitution errors. The impact of CDX-based res-\nolution on extraction fidelity is shown in Supplementary\nTable S2. The parser is implemented in Python and\nis released as an open-source tool (see Data Availabil-\nity, Section VIII). For patents containing only IUPAC\nnomenclature without embedded structural files, we use\npy2opsin [49] as a fallback.\nAgent 5 (Protein Target Resolution) maps ex-\ntracted protein names to UniProt identifiers [31] and\nretrieves their associated amino acid sequences. Due\nto the prevalence of non-standard nomenclature, unoffi-\ncial aliases, and context-dependent target specification in\npatent documents, this task requires LLM assistance. We\nemploy openai/gpt-5-2025-08-07 for this stage, as its\nsuperior reasoning capabilities relative to smaller Gemini\nFlash enable more accurate interpretation of ambiguous\nbiological context. When species is unspecified in the\npatent text, the agent defaults to Homo sapiens, consis-\ntent with the therapeutic focus of pharmaceutical patents\nand the composition of BindingDB, where 83% of binding\ndata derive from human proteins [18].\nD. Dataset Inclusion Criteria and Normalization\nHARVEST targets quantitative protein–ligand bind-\ning data. The following criteria, enforced through agent\nprompts and post-processing, define the records included\nin the final dataset.\nBinding metrics. Only IC50, Ki, Kd, and EC50 mea-\nsurements are retained. These four metrics account for\nover 80% of records extracted by the agents. Metrics\nsuch as percent inhibition are excluded, as extracting\nmeaningful binding constants from single-point or curve-\ndependent measurements would require additional pro-\ncessing logic not justified by their low prevalence in the\ncorpus. All retained values are normalized to nanomolar\n(nM).\nDefined protein targets. Only targets mappable to\none or more UniProt identifiers are retained. This in-\ncludes multi-subunit complexes, which are represented\nas semicolon-separated accessions. Approximately 87%\nof extracted records were successfully mapped. Records\nwhere the extracted target refers to a cell line or pheno-\ntypic endpoint (e.g., “HeLa cytotoxicity” or “A549 cell\nviability”) rather than a specific protein are excluded.\nDirect binding assays. Records derived from phe-\nnotypic or cell-based readouts, such as cytotoxicity, cell\nviability, proliferation, are excluded. These reported ac-\ntivity values reflect aggregate cellular responses rather\nthan direct protein–ligand interactions. This filter re-\nmoved 230,306 records (6.3%).\nMarkush structure exclusion. Patents containing\nonly Markush structures or generic R-group representa-\ntions are excluded. Enumerating individual compounds\nfrom these combinatorial representations requires spe-\ncialized chemical reasoning not currently implemented in\nthe pipeline. Because most patents additionally report\nconcrete compound examples, this criterion removed only\n8.0% of records.\nE. Architectural Design Considerations\nSeveral design decisions shaped the final HARVEST\narchitecture and are described here as they may inform\nsimilar extraction systems.\nContext window and the chunking problem.\nOur initial architecture chunked patent text to fit within\n256K-token context windows. Tables were split into row\ngroups with preserved headers and several surrounding\nparagraphs for local context. This approach introduced\nsystematic errors: chemical names spanning multiple ta-\nble cells were truncated at chunk boundaries. More criti-\ncally, patents frequently define compound aliases (e.g.,\n“Compound 1”) in one section and report bioactivity\nmeasurements in distant tables, creating unresolvable\ncross-references in isolated chunks. The availability of\nmodels with 1M+ token context windows resolved these\nissues entirely, with the additional cost offset by prompt\ncaching, resulting in a 3–5× reduction in per-patent in-\nference cost.\nExplicit grounding constraints. A persistent fail-\nure mode was the generation of plausible but fabricated\nIUPAC names for compound aliases. Agent 3 was there-\nfore instructed to return chemical identifiers only when\nfound verbatim in the patent text, defaulting to TIFF\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint \n\n12\nfilename references from chemistry tags when no system-\natic name was available. Combined with downstream\nvalidation via py2opsin and CDX file cross-referencing,\nthis constraint reduced false positive chemical identifica-\ntions in the final dataset.\nTemperature tuning.Setting the model tempera-\nture to 0 was adopted to minimize hallucinated content.\nEven modest values (e.g., temperature = 0.1) produced\nnoticeably higher rates of fabricated compound names\nand activity values during our internal validation.\nFew-shot example leakage.When a patent con-\ntained no extractable bioactivity data, the model occa-\nsionally output values copied from few-shot prompt ex-\namples. We addressed this by engineering the prompts\nto make “no data found” an explicit valid response and\npost-processing validation to filter records that matched\nour prompt examples.\nF. H-Bench Construction\nProtein selection.We selected 53 proteins span-\nning diverse target families–kinases (e.g. EGFR, JAK1,\nJAK2, KDR), GPCRs (CNR2, GRM5, DRD2, HTR2A,\nS1PR1), ion channels (P2RX3, P2RX7, GABRA1),\nlipid transfer proteins (CETP), nuclear receptors (ESR1,\nNR3C1, RORC), and epigenetic regulators (HDAC6,\nBRD4, EZH2). Of these, 44 are “overlap” targets shared\nwith BindingDB (each with>30 structural clusters and\n>1,000 BindingDB compounds) and 9 are “novel” tar-\ngets present only in HARVEST (>100 compounds). Di-\nversity across families was ensured by requiring 2–3 rep-\nresentatives per ChEMBL L2 target class and removing\nredundant proteins via sequence clustering (MMseqs2,\n40% identity threshold). An activity-balance filter re-\ntained only proteins for which 20–55% of Valid-subset\ncompounds are active.\nCluster-based splitting.For each protein, com-\npound Morgan fingerprints (radius 2, 2048 bits) were\ncomputed and clustered via complete-linkage hierarchical\nclustering at a Tanimoto distance threshold of 0.2. Clus-\nters were initially labeled by data source: A (HARVEST-\nonly) or B (BindingDB-only). A cluster connectivity\ngraph was then constructed using centroid Tanimoto sim-\nilarity (threshold 0.225, ≤2 hops), and an integer linear\nprogram (ILP) identified the minimum set of A clus-\nters to relabel as buffer C, maximizing structural sepa-\nration between the Valid (A) and BindingDB-only (B)\nsubsets. The final partition yields three non-overlapping\nsets: Valid (structurally distant from BindingDB), Com-\nmon (buffer zone, included in H-Bench), andBindingDB-\nonly (excluded from H-Bench).\nBoltz evaluation. Boltz-2 predictions were gener-\nated in affinity mode with 5 recycling steps, MSA server\nqueries, and molecular-weight correction enabled. Up to\n120 ligands per protein were sampled, balanced between\nactive and inactive classes and drawn from diverse clus-\nters. For compounds in the Common set, we selected only\nmolecules with Tanimoto similarity below 0.75. Three\noutput scores were evaluated: affinity\npred value,\naffinity prob binary, andconfidence score. Per-\nprotein AUC-ROC was computed for proteins with ≥5\ncompounds and both activity classes present, yielding\n6,469 predictions across 53 proteins (Valid) and 5,399\npredictions across 44 proteins (Common).\n[1] M. Bregonje, Patents: a unique source for scientific tech-\nnical information in chemistry related industry?, World\nPatent Inf 27, 309 (2005).\n[2] M. E. Valentinuzzi, Patents and scientific papers: Quite\ndifferent concepts: The reward is found in giving, not in\nkeeping [retrospectroscope], IEEE Pulse 8, 49 (2017).\n[3] C. Southan, P. Varkonyi, K. Boppana, S. A. R. P. Jagar-\nlapudi, and S. Muresan, Tracking 20 years of compound-\nto-target output from literature and patents, PLOS ONE\n8, e77142 (2013).\n[4] J. L. Watson, D. Juergens, N. R. Bennett, B. L. Trippe,\nJ. Yim, H. E. Eisenach, W. Ahern, A. J. Borst, R. J.\nRagotte, L. F. Milles, B. I. M. Wicky, N. Hanikel, S. J.\nPellock, A. Courbet, W. Sheffler, J. Wang, P. Venkatesh,\nI. Sappington, S. V´ azquez Torres, A. Lauko, V. De Bor-\ntoli, E. Mathieu, S. Ovchinnikov, R. Barzilay, T. S.\nJaakkola, F. DiMaio, M. Baek, and D. Baker, De novo\ndesign of protein structure and function with RFdiffu-\nsion, Nature 620, 1089 (2023).\n[5] N. R. Bennett, J. L. Watson, R. J. Ragotte, A. J.\nBorst, D. L. See, C. Weidle, R. Biswas, Y. Yu, E. L.\nShrock, R. Ault, P. J. Y. Leung, B. Huang, I. Goreshnik,\nJ. Tam, K. D. Carr, B. Singer, C. Criswell, B. I. M.\nWicky, D. Vafeados, M. Garcia Sanchez, H. M. Kim,\nS. V´ azquez Torres, S. Chan, S. M. Sun, T. T. Spear,\nY. Sun, K. O’Reilly, J. M. Maris, N. G. Sgourakis, R. A.\nMelnyk, C. C. Liu, and D. Baker, Atomically accurate de\nnovo design of antibodies with RFdiffusion, Nature 649,\n183 (2026).\n[6] F. J. Ferreira and A. S. Carneiro, Ai-driven drug dis-\ncovery: a comprehensive review, ACS omega 10, 23889\n(2025).\n[7] J. Jumper, R. Evans, A. Pritzel, T. Green, M. Fig-\nurnov, O. Ronneberger, K. Tunyasuvunakool, R. Bates,\nA. ˇZ´ ıdek, A. Potapenko, A. Bridgland, C. Meyer, S. A. A.\nKohl, A. J. Ballard, A. Cowie, B. Romera-Paredes,\nS. Nikolov, R. Jain, J. Adler, T. Back, S. Petersen,\nD. Reiman, E. Clancy, M. Zielinski, M. Steinegger,\nM. Pacholska, T. Berghammer, S. Bodenstein, D. Silver,\nO. Vinyals, A. W. Senior, K. Kavukcuoglu, P. Kohli, and\nD. Hassabis, Highly accurate protein structure prediction\nwith AlphaFold, Nature 596, 583 (2021).\n[8] J. Abramson, J. Adler, J. Dunger, R. Evans, T. Green,\nA. Pritzel, O. Ronneberger, L. Willmore, A. J. Bal-\nlard, J. Bambrick, S. W. Bodenstein, D. A. Evans,\nC.-C. Hung, M. O’Neill, D. Reiman, K. Tunyasuvu-\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint \n\n13\nnakool, Z. Wu, A. ˇZemgulyt˙ e, E. Arvaniti, C. Beat-\ntie, O. Bertolli, A. Bridgland, A. Cherepanov, M. Con-\ngreve, A. I. Cowen-Rivers, A. Cowie, M. Figurnov, F. B.\nFuchs, H. Gladman, R. Jain, Y. A. Khan, C. M. R. Low,\nK. Perlin, A. Potapenko, P. Savy, S. Singh, A. Stecula,\nA. Thillaisundaram, C. Tong, S. Yakneen, E. D. Zhong,\nM. Zielinski, A. ˇZ´ ıdek, V. Bapst, P. Kohli, M. Jaderberg,\nD. Hassabis, and J. M. Jumper, Accurate structure pre-\ndiction of biomolecular interactions with AlphaFold 3,\nNature630, 493 (2024).\n[9] S. Passaro, G. Corso, J. Wohlwend, M. Reveiz, S. Thaler,\nV. R. Somnath, N. Getz, T. Portnoi, J. Roy, H. Stark,\nD. Kwabi-Addo, D. Beaini, T. Jaakkola, and R. Barzilay,\nBoltz-2: Towards accurate and efficient binding affinity\nprediction, bioRxiv 10.1101/2025.06.14.659707 (2025),\npreprint.\n[10] P. Tossou, C. Wognum, M. Craig, H. Mary, and\nE. Noutahi, Real-world molecular out-of-distribution:\nSpecification and investigation, Journal of Chemical In-\nformation and Modeling 64, 697 (2024).\n[11] H. Fooladi, T. N. L. Vu, M. Mathea, and J. Kirchmair,\nEvaluating machine learning models for molecular prop-\nerty prediction: Performance and robustness on out-of-\ndistribution data, Journal of Chemical Information and\nModeling 65, 9871 (2025), pMID: 40947919.\n[12] D. van Tilborg, L. Rossen, and F. Grisoni, Molecular\ndeep learning at the edge of chemical space, ChemRxiv\n10.26434/chemrxiv-2025-qj4k3-v3 (2025), preprint.\n[13] N. Pallikkavaliyaveetil and S. Chandrasekaran, Small\ndata, big challenges: Machine- and deep-learning strate-\ngies for data-limited drug discovery, Advanced Drug De-\nlivery Reviews 10.1016/j.addr.2025.115762 (2025).\n[14] K. Debnath, P. Rana, and P. Ghosh, A survey on\ndeep learning for drug-target binding prediction: mod-\nels, benchmarks, evaluation, and case studies, Briefings\nin Bioinformatics 26, bbaf491 (2025).\n[15] A. Gangwal, A. Ansari, I. Ahmad, A. K. Azad, and\nW. M. A. W. Sulaiman, Current strategies to address\ndata scarcity in artificial intelligence-based drug discov-\nery: A comprehensive review, Computers in Biology and\nMedicine 179, 108734 (2024).\n[16] J. Dagdelen, A. Dunn, S. Shen, et al., Structured infor-\nmation extraction from scientific text with large language\nmodels, Nat Commun 15, 1418 (2024).\n[17] J. Michels et al., Natural language processing methods\nfor the study of protein–ligand interactions, J Chem Inf\nModel 65, 2191 (2025).\n[18] T. Liu, L. Hwang, S. K. Burley, C. I. Nitsche, C. Southan,\nW. P. Walters, and M. K. Gilson, Bindingdb in 2024:\na fair knowledgebase of protein-small molecule binding\ndata, Nucleic acids research 53, D1633 (2025).\n[19] D. M. Jessop, S. E. Adams, and P. Murray-Rust, Mining\nchemical information from open patents, J Cheminform\n3, 40 (2011).\n[20] M. Krallinger, F. Leitner, O. Rabal, M. Vazquez,\nJ. Oyarzabal, and A. Valencia, Chemdner: the drugs and\nchemical names extraction challenge, J Cheminform7, S1\n(2015).\n[21] G. Papadatos, M. Davies, N. Dedman, et al., Surechembl:\na large-scale, chemically annotated patent document\ndatabase, Nucleic Acids Res 44, D1220 (2016).\n[22] Y. Gadiya, S. Shetty, M. Hofmann-Apitius, P. Gribbon,\nand A. Zaliani, Exploring surechembl from a drug dis-\ncovery perspective, Scientific data 11, 507 (2024).\n[23] S. Qiao, R. Fang, Z. Qiu, X. Wang, N. Zhang, Y. Jiang,\nP. Xie, F. Huang, and H. Chen, Benchmarking agen-\ntic workflow generation, arXiv preprint arXiv:2410.07869\n(2024).\n[24] G. Appenzeller, Llmflation: Llm inference cost going\ndown fast, https://a16z.com/llmflation-llm-infer\nence-cost/ (2024), discusses Moore’s-law-like declines\nin LLM inference cost at fixed quality.\n[25] A. L. Hopkins and C. R. Groom, The druggable genome,\nNature reviews Drug discovery 1, 727 (2002).\n[26] D. Stumpfe, H. Hu, and J. Bajorath, Evolving concept\nof activity cliffs, ACS Omega 4, 14360 (2019).\n[27] S. R. Heller, A. McNaught, I. Pletnev, S. Stein, and\nD. Tchekhovskoi, Inchi, the iupac international chemical\nidentifier, J Cheminform 7, 23 (2015).\n[28] H. M. Berman, J. Westbrook, Z. Feng, G. Gilliland, T. N.\nBhat, H. Weissig, I. N. Shindyalov, and P. E. Bourne,\nThe protein data bank, Nucleic Acids Research 28, 235\n(2000).\n[29] J. Yan, J. Zhu, Y. Yang, Q. Liu, K. Zhang, Z. Zhang,\nX. Liu, B. Zhang, K. Gao, J. Xiao, et al., Biominer:\nA multi-modal system for automated mining of protein-\nligand bioactivity data from literature, bioRxiv , 2025\n(2025).\n[30] Z. Wang, F. Fu, W. Zhang, L. Yan, Y. Meng, J. Wu,\nH. Wu, G. Xu, and S. Chen, Biocheminsight: An open-\nsource toolkit for automated identification and recogni-\ntion of optical chemical structures and activity data in\nscientific publications, arXiv preprint arXiv:2504.10525\n(2025).\n[31] The UniProt Consortium, Uniprot: the universal pro-\ntein knowledgebase in 2023, Nucleic Acids Res 51, D523\n(2023).\n[32] Elsevier, Reaxys, https://www.elsevier.com/product\ns/reaxys (2026), accessed: 22 February 2026.\n[33] Elsevier, Overview of reaxys data, https://www.elsevi\ner.support/dataasaservice/answer/overview-of-r\neaxys-data (2026), accessed: 22 February 2026.\n[34] Excelra, Gostar™ small molecules, https://www.excelr\na.com/databases/gostar (2026), accessed: 22 February\n2026.\n[35] China National Intellectual Property Administration,\nPromoting the use of extensible markup language (xml)\nformat for submitting electronic patent application doc-\numents, https://www.cnipa.gov.cn/art/2025/5/26/a\nrt_75_199841.html (2025), accessed: 2026-01-25.\n[36] A. Vinogradova, V. Vinogradov, L. Greenwood, I. Yasny,\nD. Kobyzev, S. Kasbekar, K. Nguyen, D. Radkevich,\nR. Doronin, and A. Doronichev, Hunt globally: Wide\nsearch AI agents for drug asset scouting in invest-\ning, business development, and competitive intelligence,\narXiv preprint arXiv:2602.15019 (2026), preprint.\n[37] A. Vinogradova, V. Vinogradov, D. Radkevich, I. Yasny,\nD. Kobyzev, I. Izmailov, K. Yanchanka, R. Doronin, and\nA. Doronichev, LLM-based agents for competitive land-\nscape mapping in drug asset due diligence, arXiv preprint\narXiv:2508.16571 (2025), preprint.\n[38] L. Liu, V. Blake, M. Barman, B. Gallego, T. Churches,\nG. Kennedy, S.-Y. Ooi, G. P. Delaney, and L. Jorm, Us-\ning natural language processing to extract information\nfrom clinical text in electronic medical records for popu-\nlating clinical registries: a systematic review, Journal of\nthe American Medical Informatics Association 33, 484\n(2026).\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint \n\n14\n[39] Z. Yang, H. Yuan, R. Sayeed, A. L. M. Tan, E. Cai,\nM. Moro, X. Li, H. Ying, N. Brown, G. Weber, et al. ,\nCLINES: Clinical LLM-based information extraction and\nstructuring agent, medRxiv , 2025 (2025), preprint.\n[40] E. Boutet, D. Lieberherr, M. Tognolli, M. Schneider, and\nA. Bairoch, Uniprotkb/swiss-prot, Methods in molecular\nbiology (Clifton, N.J.) 406, 89 (2007).\n[41] Google, Google patents, https://patents.google.com/\n(2026), accessed: 22 February 2026.\n[42] Google, Google patents public datasets: connecting pub-\nlic, paid, and private patent data, https://cloud.goog\nle.com/blog/topics/public-datasets/google-paten\nts-public-datasets-connecting-public-paid-and\n-private-patent-data/ (2017), accessed: 22 February\n2026.\n[43] The Lens, Lens.org, https://www.lens.org/lens/user\n/subscriptions (2026), accessed: 22 February 2026.\n[44] USPTO, Uspto bulk data directory, https://data.usp\nto.gov/bulkdata/datasets (2026), accessed: 22 Febru-\nary 2026.\n[45] T. Zhou, J. Xu, G. Liu, J. Liu, H. Wang, and E. Wu, An\napproach for systematic decomposition of complex LLM\ntasks, arXiv preprint arXiv:2510.07772 (2026), intro-\nduces ACONIC framework showing decomposition im-\nproves agent performance on complex tasks.\n[46] S. Montazeri, Y. Feng, and K. Sha, PublicAgent: Multi-\nagent design principles from an LLM-based open data\nanalysis framework, arXiv preprint arXiv:2511.03023\n(2025), shows specialization provides value independent\nof model strength with 97.5% agent win rates.\n[47] Y. Yang, Q. Peng, J. Wang, Y. Wen, and W. Zhang,\nLLM-based multi-agent systems: Techniques and busi-\nness perspectives, arXiv preprint arXiv:2411.14033\n(2024), discusses dynamic task decomposition and or-\nganic specialization advantages over single-agent sys-\ntems.\n[48] U.-e. Nisa et al., Agentic ai: The age of reasoning—a\nreview, Journal of Artificial Intelligence (2025), a survey\nof agentic AI systems and their reasoning capabilities.\n[49] D. M. Lowe, P. T. Corbett, P. Murray-Rust, and R. C.\nGlen, Chemical name to structure: Opsin, an open source\nsolution, J Chem Inf Model 51, 739 (2011).\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint \n\nS1\nSupplementary Information\nHAR VEST: Unlocking the Dark Bioactivity Data of Pharmaceutical Patents via\nAgentic AI\nAppendix S1: Supplementary tables\nTarget Class Proteins Novel Compounds Novel Clusters\nEnzyme\nKinase 433 235,302 117,642\nTransferase 256 56,736 28,143\nProtease 221 44,680 23,166\nOxidoreductase 206 35,676 18,692\nOther 154 35,461 18,335\nHydrolase 244 30,281 15,282\nUnspecified 115 15,267 7,960\nPhosphatase 61 7,345 3,884\nLyase 45 1,233 714\nMembrane receptor\nFamily A G protein-coupled receptor 387 122,178 59,877\nOther 64 28,526 13,787\nUnspecified 68 13,054 6,321\nIon channel\nVoltage-gated ion channel 98 23,129 11,354\nLigand-gated ion channel 53 18,263 8,982\nOther 18 12,322 5,732\nUnclassified protein 313 46,928 22,499\nTranscription factor\nNuclear receptor 64 24,212 11,638\nUnspecified 70 18,099 8,796\nEpigenetic regulator 108 39,917 20,460\nUnclassified 755 33,066 15,806\nSecreted protein 80 26,798 12,500\nTransporter\nElectrochemical transporter 95 13,335 6,372\nOther 33 5,134 2,429\nOther 82 10,391 4,897\nOther cytosolic protein 54 10,359 5,357\nTotal 4,077 907,692 450,625\nTABLE S1: Distribution of novel protein–ligand interactions (PLIs) and structural clusters contributed by HARVEST,\ngrouped by ChEMBL target classification. For each Level 1 (L1) target class (bold font) containing multiple Level 2 (L2)\nfamilies, L2 subcategories are listed with indentation. Novel compounds/clusters include all ligands associated with\nHARVEST-only proteins or ligands not present in BindingDB for shared proteins. Structural clusters were defined using\nhierarchical complete-linkage clustering of Morgan fingerprints (radius = 2, 2048 bits) with a Tanimoto distance threshold of\n0.2, followed by aggregation of adjacent clusters\nAppendix S2: Alternative Patent Data Sources\nSeveral alternative patent data sources were evaluated but found unsuitable for corpus-scale bioactivity extraction:\nSureChEMBL (https://www.surechembl.org) provides free access to patent-extracted chemical structure data\nthrough an open API [21]. However, tabular data in the processed output are concatenated without clear delimiters,\nmaking it difficult to distinguish decimal separators from column boundaries. The API also exhibits instability\nunder sustained high-throughput querying, returning HTTP 500 errors at the request rates required for corpus-scale\nextraction.\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint \n\nS2\nGoogle Patents(https://patents.google.com) andGoogle BigQuery Patents Public Data were both\nevaluated but found to be unsuitable: Google Patents implements aggressive rate limiting and IP-based blocking during\nprogrammatic access, while BigQuery, though well-structured, proved cost-prohibitive for extracting the complete\npharmaceutical patent corpus with full-text descriptions in academic research budgets.\nLens.org provides a more cost-effective alternative with institutional subscription options, but patent description\nfields are provided in plain text rather than structured markup, complicating the disambiguation of tabular data and\nchemical structure references essential for Document–Assay–Result–Compound–Protein relationship extraction.\nAppendix S3: Impact of CDX-based structure resolution\nTABLE S2: Cross-validation on shared patents without CDX-based chemical structure resolution (using only MOL files and\npy2opsin). Without the CDX parser, the full-match rate drops from 80.3% to 72.7% for HARVEST records and from 70.5%\nto 63.6% for BDB records (cf. Table I), while compound mismatches approximately double (from 7.9% to 15.5% and from\n18.0% to 25.0%, respectively). These results demonstrate the improvement in extraction fidelity achieved by parsing chemical\nstructures directly from the original ChemDraw binary files.\nHARVEST\n(n=603,929)\nin BDB\nBDB\n(n=690,907)\nin HARVEST\nFull match 439,113 (72.7%) 439,113 (63.6%)\nTarget mismatch 57,079 (9.5%) 49,151 (7.1%)\nCompound mismatch 93,471 (15.5%) 172,515 (25.0%)\nNo overlap 14,266 (2.4%) 30,128 (4.4%)\nAppendix S4: Cross-validation case studies\nTABLE S3: Manual review of the largest cross-validation mismatches (a): Target mismatch — the same compound was\nfound in both databases but assigned to different proteins.\nPatent T arget UniProt ID\n(HAR VEST)\nUniProt ID\n(BindingDB)\nV erdict\nUS20140275087A1 GlyT1SC6A9 HUMAN MGAT1 HUMANBindingDB incorrectly recognized a protein\nUS20150376212A1 FLAPAL5AP HUMAN FEN1 HUMANBindingDB incorrectly recognized a protein\nUS20240246937A1 IL4I1OXLA HUMAN IL1RA HUMANBindingDB incorrectly recognized a protein\nUS20250051309A1 MK2, p38αMAPK2 HUMAN,\nMK14 HUMAN\nMK01 HUMAN,\n4EBP1 HUMAN\nBindingDB incorrectly recognized a protein\nUS20170226089A1 NR2BNMDZ1 HUMAN;\nNMDE1 HUMAN;\nNMDE2 HUMAN;\nNMDE3 HUMAN;\nNMDE4 HUMAN;\nNMD3A HUMAN;\nNMD3B HUMAN\nRXRA HUMANHARVEST and BindingDB incorrectly\nrecognized a protein\nUS20150191464A1 IRAK4IRAK4 HUMAN,\nTLR2 HUMAN\nIRAK4 HUMANHARVEST incorrectly recognized assay\nUS20140057889A1 Bcl-XL, Bcl-2B2CL1 HUMAN,\nBCL2 HUMAN\nB2CL1 MOUSE,\nBCL2 MOUSE,\nBCL2 HUMAN\nNo information about the organism\nUS20120035149A1 PKR1, PKR2PKR1 HUMAN,\nPKR2 HUMAN\nG3HIE5 CRIGR,\nG3H407 CRIGR\nNo information about the organism\nUS20240059703A1 KRAS G12C,C118A (aa 1-169), His-tagged RASK HUMAN SOS1 HUMANHARVEST and BindingDB do not process\nprotein modifications\nUS20170334917A1 PG-K1 (aa101-181), His-tagged PLMN HUMAN TPA HUMANHARVEST and BindingDB do not process\nprotein modifications\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint \n\nS3\nTABLE S4:Manual review of the largest cross-validation mismatches (b): Compound mismatch — the same protein was\nfound in both databases but associated with different compound sets.\nPatent H B Ovlp V erdict\nUS20230255941A1 44 953 33 HARVEST incomplete extraction\nUS20200095247A1 57 804 56 HARVEST incomplete extraction\nUS20130172317A1 68 2136 52 HARVEST incomplete extraction\nUS20220175775A1 79 1080 79 HARVEST incomplete extraction\nUS20140094448A1 205 1116 179 HARVEST incomplete extraction\nUS20240208941A1 89 332 89 HARVEST incomplete extraction\nUS20190322661A1 32 462 30 HARVEST incomplete extraction\nUS20240294506A1 2 499 2 HARVEST incomplete extraction\nUS20250129020A1 201 454 197 HARVEST incomplete extraction\nUS20240294524A1 84 783 68 HARVEST incomplete extraction\n(activity table not present in the XML package)\nUS20240254137A1 37 892 4 Mixed case: HARVEST incomplete extraction\nand non-matching source documents\nUS20190241588A1 754 766 59 HARVEST chemical structure error\nUS20250017938A1 781 1 1 BindingDB incomplete curation\nUS20160304519A1 515 48 48 BindingDB incomplete curation\nUS20220041601A1 284 8 4 BindingDB incomplete curation\nUS20120058988A1 562 94 92 BindingDB incomplete curation\nUS20180118724A1 332 259 146 HARVEST extraction error\nUS20130079303A1 322 318 115 HARVEST chemical structure error\nUS20180273529A1 443 298 40 HARVEST chemical structure error\nUS20230286970A1 393 405 392 Different inclusion threshold for inactive measurements\nTABLE S5: Manual review of the largest cross-validation mismatches (c): No overlap — neither compounds nor targets\nmatched between databases for the same patent.\nPatent H B V erdict\nUS20180118720A1 105 42 BindingDB patent mapping error\nUS20210009607A1 138 25 BindingDB patent mapping error\nUS20240343719A1 117 20 BindingDB patent mapping error\nUS20200157106A1 5 800 BindingDB patent mapping error\nUS20160115131A1 97 516 BindingDB patent mapping error\nUS20210053946A1 8 612 BindingDB accession or identifier error\nUS20190343826A1 37 608 BindingDB accession or identifier error\nUS20210188777A1 163 7 BindingDB incomplete curation\nUS20240209001A1 103 10 BindingDB incomplete curation\nUS20220315606A1 429 3 BindingDB incomplete curation\nUS20240400523A1 1 619 HARVEST incomplete extraction\nUS20250064789A1 16 540 Mixed case: identifier mismatch and\nHARVEST incomplete extraction\nUS20160060224A1 63 505 Technical identifier limitation\nUS20230365594A1 214 8 Different assay focus across sources\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted March 18, 2026. ; https://doi.org/10.64898/2026.03.15.711910doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}