Mining millions of biomedical papers with large language models to surface novel, overlooked drug targets
Biomedical knowledge doubles roughly every few years, buried across tens of millions of papers, preprints, and patents that no human team could read cover-to-cover. Before any AI system can propose a new drug target, it must first ingest this corpus — normalizing formats, deduplicating, and staging text for downstream extraction at a scale that dwarfs any individual researcher's reading capacity.
PubMed alone indexes more than 37 million citations spanning over 5,600 journals, growing by roughly 1 million new records per year. Add PubMed Central full-text articles, preprint servers (bioRxiv, medRxiv, chemRxiv), regulatory filings, conference abstracts, and patent literature (USPTO, EPO, WIPO), and the addressable corpus for target discovery exceeds 50 million documents.
No human research team can read at this scale. A domain expert reading 5 papers a day, 250 days a year, would need over 40,000 years to read the current PubMed corpus alone — and that ignores the million new papers published every subsequent year. This is the fundamental motivation for automated literature mining: the knowledge needed to find the next drug target is very likely already published, scattered across disconnected papers that no single researcher has read together.
PubMed grew from roughly 12 million citations in 2004 to more than 37 million in 2025 — more than tripling in two decades, while the number of working biomedical researchers grew far more slowly.
Before any language model touches the text, an ingestion pipeline handles unglamorous but essential steps: PDF and XML parsing (JATS format for PMC full text), OCR fallback for scanned older articles, deduplication across preprint/published-version pairs, metadata harmonization (journal, date, MeSH terms, author affiliations), and chunking of long documents into passages sized for transformer context windows.
Modern pipelines also score source reliability — a peer-reviewed Nature paper is weighted differently than an unreviewed preprint or a low-impact journal — and track provenance so that every downstream claim can be traced back to its originating sentence and paper for auditability, a requirement increasingly demanded by pharma R&D governance.
Curated databases like UniProt, OMIM, and Open Targets rely partly on manual literature curation by trained biocurators — invaluable but slow. A single biocurator can review roughly 5–10 papers per day for structured data extraction; at that rate, keeping pace with 1 million new papers per year would require several hundred full-time curators dedicated solely to new publications, before even addressing the backlog.
LLM-based pipelines instead process thousands of papers per hour on modest GPU clusters, extracting structured entities and relations continuously. This does not replace curation — it triages it, surfacing the highest-value papers and candidate facts for human curators and target-discovery scientists to review, dramatically increasing the fraction of the literature that translates into usable biological knowledge.
A fine-tuned biomedical language model reads each paper and performs two linked tasks: named-entity recognition (NER), tagging spans of text as genes, proteins, diseases, pathways, or compounds; and relation extraction (RE), identifying how those entities relate — "gene X is upregulated in disease Y," "protein Z inhibits pathway W." This is the step that converts unstructured prose into structured, queryable facts.
Biomedical NER has a well-established lineage of domain-adapted transformer models. BioBERT (Lee et al., 2020) took the original BERT architecture and continued pretraining on PubMed abstracts and PMC full-text, substantially improving performance on biomedical NER, relation extraction, and question answering compared to general-domain BERT.
PubMedBERT (Gu et al., 2020, Microsoft Research) went further, training entirely from scratch on biomedical text rather than fine-tuning a general-domain checkpoint, and using a domain-specific vocabulary — showing that in-domain pretraining from the start outperforms domain adaptation of general models on the BLURB biomedical NLP benchmark. SciBERT (Beltagy et al., 2019) took a similar from-scratch approach across broader scientific text (not just biomedicine), useful for cross-domain literature such as chemistry and materials science that intersects with drug discovery.
These encoder-only models remain the workhorses for high-throughput, low-cost NER tagging — cheap enough to run over the entire PubMed corpus — while newer LLM-based approaches add reasoning and relation extraction on top.
BioBERT reduced the gap to human-level performance on biomedical NER by a wide margin over prior general-domain models — but even a 90%+ F1 score means roughly 1 in 10 entity mentions is mislabeled at web scale, which is why downstream confidence scoring and human review remain essential.
Where encoder-only models like BioBERT excel at fast, cheap tagging, decoder LLMs (GPT-4-class, Claude-class, and open models like Llama fine-tuned on biomedical corpora) add something new: they can read a full passage in context, resolve ambiguous references ("it" referring to a gene mentioned three sentences earlier), and extract nuanced relation types with attached qualifiers — directionality, effect size, tissue context, and confidence language ("suggests," "demonstrates," "is correlated with" versus "causes").
Retrieval-augmented generation (RAG) pipelines pair an LLM with a search index over the full corpus: given a candidate gene-disease pair, the system retrieves the most relevant passages across millions of papers and asks the LLM to synthesize and cite specific evidence, rather than relying purely on parametric memory (which can be stale or hallucinated). This hybrid of fast NER models for bulk tagging plus LLM-based synthesis for high-value passages is now the dominant architecture in production literature-mining systems.
Raw extracted text spans must be normalized to canonical identifiers before they are useful: "TNF-alpha," "TNFα," and "tumor necrosis factor alpha" must all resolve to the same gene entity (TNF, HGNC:11892). This entity linking step maps free text to standard ontologies — HGNC and Entrez for genes, UniProt for proteins, MONDO and MeSH for diseases, Reactome and KEGG for pathways, ChEMBL for compounds.
Without robust normalization, a knowledge graph built downstream would fragment into thousands of near-duplicate nodes for the same biological entity, destroying the very connectivity that makes a knowledge graph valuable. Normalization accuracy is therefore one of the most consequential — and least glamorous — quality bottlenecks in the entire pipeline.
Every extracted entity and relation is added as a node or edge to a continuously growing biomedical knowledge graph. Unlike a single paper, which reveals one fact, a knowledge graph spanning millions of papers reveals structure — clusters, hubs, and crucially, gaps: gene-disease pairs that share many indirect paths through shared pathways but have never been directly co-mentioned in any single paper.
A biomedical knowledge graph typically models several node types (gene/protein, disease, pathway, compound/drug, phenotype, tissue) connected by typed, directed edges (upregulates, associated_with, inhibits, expressed_in, treats). Each edge carries provenance metadata: the source paper(s), extracted sentence, publication date, and an LLM-assigned confidence score.
Critically, edges are not binary facts but weighted, evidence-aggregated claims — the same gene-disease pair might accumulate dozens of independent mentions across the corpus, each adding to a cumulative evidence score, while a spurious single-paper claim remains low-weight until corroborated elsewhere. This aggregation is what distinguishes a knowledge graph from a simple citation index.
The Open Targets Platform (a public-private partnership including GSK, EMBL-EBI, Wellcome Sanger, and others) is the most widely used open biomedical knowledge graph for target identification, integrating genetic association data, literature co-occurrence, expression data, and pathway data into a single target-disease association score, freely queryable by any researcher.
BenevolentAI (founded 2013, London) was among the first companies to commercialize LLM-adjacent NLP-driven knowledge graphs for drug discovery, notably contributing an early literature-mining-derived hypothesis (baricitinib as a JAK inhibitor with antiviral potential) that was later clinically validated during COVID-19 — one of the most cited real-world successes of AI-assisted literature mining, though the broader track record of AI-first biotechs converting hypotheses into approved drugs remains thin relative to the hype.
Insilico Medicine's PandaOmics couples multi-omics data with an internal biomedical knowledge graph and LLM-based text mining to score novel targets, feeding candidates into its Chemistry42 generative chemistry engine — the combined pipeline was used to nominate a novel target and generate a clinical-stage candidate for idiopathic pulmonary fibrosis in under 18 months, markedly faster than traditional discovery timelines, though later-stage clinical attrition still applies as it does to any drug candidate.
IBM Watson for Drug Discovery, once heavily promoted for literature-mining-driven target and hypothesis generation, was discontinued after partner hospitals and pharma collaborators reported that its recommendations often reflected obvious, already-known associations rather than genuinely novel insight — a widely cited cautionary lesson in evaluating AI literature-mining claims against real discovery value, not just processing volume.
The payoff of graph construction is topological: once entities are embedded in a connected graph, algorithms can measure indirect relatedness that no single paper states explicitly. Two proteins might share five pathway neighbors and a dozen co-mentioned diseases without ever being discussed together directly — a strong hint of an unstated but plausible mechanistic link, and exactly the kind of gap literature-mining systems are built to surface for further investigation.
Graph embedding techniques (node2vec, knowledge-graph embeddings like TransE and ComplEx, and increasingly graph neural networks) learn continuous vector representations of every node such that biologically similar entities cluster together in embedding space, allowing link-prediction models to rank plausible-but-unstated gene-disease edges across the entire graph simultaneously, rather than one hand-picked pair at a time.
| Product | Indication | Trial Design | Key Result |
|---|---|---|---|
| Manual literature review | Curator/scientist-selected papers | Domain expert reads, annotates, and synthesizes papers by hand | Highest per-paper judgment quality; gold standard for depth |
| Keyword / co-occurrence text mining | Full corpus, keyword-indexed | Counts term co-occurrence within abstracts/sentences (e.g. PubTator-style) | Fast, cheap, fully auditable, decades of tooling maturity |
| BioBERT / PubMedBERT NER pipelines | Full corpus, entity-tagged | Fine-tuned transformer NER + relation classifiers at scale | High-precision structured extraction at low inference cost |
| LLM + knowledge-graph platforms | Full corpus + omics + genetics | RAG-enabled LLM synthesis fused with graph embeddings and multi-omics scoring | Surfaces indirect, cross-source novel hypotheses with evidence citations |
A knowledge graph alone is not a discovery — it is a map. The scoring step is where the system decides which gene-disease edges deserve human attention: ranking under-explored links by a combination of literature novelty (low mention count, but growing), evidence strength (causal versus merely correlative language), and independent corroboration from GWAS, eQTL, and other omics data sources external to the text corpus itself.
The scoring challenge is a precision-recall trade-off familiar to any ranking system, but with unusually high stakes: rank too permissively and scientists waste months chasing spurious correlations; rank too conservatively and the system only ever rediscovers already well-known biology, the exact failure mode that undermined earlier AI literature-mining efforts.
Effective scoring models combine several signals: (1) co-occurrence frequency — how often gene and disease appear together, weighted so extremely high frequency indicates "already well studied" rather than "novel"; (2) causal language strength — LLM-extracted relation types are weighted, with "X causes Y" and "loss-of-function of X drives Y in a mouse model" scored higher than "X was measured alongside Y"; (3) recency and trend — links with accelerating mention frequency in the last 2–3 years may indicate an emerging but under-consolidated finding; (4) independent genetic validation, layered in from outside the text corpus entirely.
The single strongest external corroboration signal for a candidate target is human genetic evidence. Genome-wide association studies (GWAS) identify loci statistically associated with a disease across large populations; when a GWAS-implicated gene independently matches an LLM-surfaced literature hypothesis, the two orthogonal evidence types reinforce each other far more convincingly than either alone.
Mendelian randomization (MR) goes further, using naturally occurring genetic variants as an instrumental variable to test whether a gene's effect on a biomarker is causally related to disease risk — approximating a randomized trial using population genetics, without needing to actually dose a drug. A influential 2015 analysis (Nelson, Tyler, Cardon et al., Nature Genetics) found that drug targets with human genetic support were roughly twice as likely to succeed through clinical development, which is why nearly every modern target-scoring pipeline, whether LLM-based or not, weights genetic corroboration heavily.
Open Targets and similar platforms combine literature-derived evidence with genetic association, expression specificity, animal model, and pathway data into a single composite target-disease association score precisely because no single evidence type — including LLM-mined literature — is considered sufficient on its own.
A biologically compelling gene-disease link is not automatically an actionable drug target. Tractability assessment asks a separate, practical question: can this protein realistically be modulated by a drug molecule at all? Roughly 20,000 human protein-coding genes exist, but only an estimated 3,000 or so are considered part of the "druggable genome" — proteins with binding pockets suitable for small molecules, accessible extracellular domains suitable for antibodies, or other modalities (PROTACs, ASOs, gene therapy) that expand the addressable space further.
Scoring pipelines therefore fold tractability databases (ChEMBL bioactivity data, structural pocket predictions, existing chemical probe availability, safety/toxicity liability databases) into the final rank, downweighting biologically exciting but practically intractable targets in favor of those with a credible path to an actual therapeutic modality.
No literature-mining system, however sophisticated, is deployed as an autonomous decision-maker in pharma R&D. The final and non-negotiable step is expert review: human scientists interrogate the highest-scoring candidates, check the underlying evidence citations for hallucination or misreading, cross-reference tractability and safety databases, and select the small handful of targets worth committing real experimental budget to pursue.
Large language models can generate fluent, confident-sounding claims that are not actually supported by the source text they were asked to summarize — a failure mode directly relevant to literature mining, where a hallucinated "gene X causes disease Y" claim, if unchecked, could send a research team down a costly dead-end.
Production systems mitigate this with citation-grounded generation (every extracted claim must link to a specific retrievable sentence in a specific paper), confidence calibration, and — critically — a mandatory human verification step before any claim influences a real R&D decision. Reviewers spot-check the cited passages directly rather than trusting the LLM's summary of them, treating the system's output as a lead to verify, not a finding to accept.
The cautionary lesson from IBM Watson for Drug Discovery and similar earlier efforts was not that literature mining is worthless, but that unchecked, opaque AI recommendations without rigorous evidence traceability and expert sign-off failed to earn the trust — or deliver the results — needed for real R&D adoption.
Text-mined co-occurrence, even when extracted with a sophisticated LLM, fundamentally measures what has been written about together, not what is biologically true. A gene may be frequently mentioned alongside a disease simply because it was used as a control, a housekeeping comparator, or because two unrelated research trends happened to intersect in citation patterns — not because it plays a causal mechanistic role.
Expert reviewers apply exactly the causal-inference scrutiny that automated scoring can only approximate: is there a plausible mechanism connecting the gene to disease pathophysiology? Do animal models or CRISPR knockout/knockdown studies support a functional, not just associative, relationship? Does the genetic evidence (MR, GWAS colocalization) point in a direction consistent with a therapeutic hypothesis, e.g., loss-of-function protecting against disease, implying inhibition could treat it?
Once a shortlist of 3–10 candidate targets per disease area survives expert review, they enter conventional experimental target validation: CRISPR knockout or knockdown in relevant cell lines and organoids, target engagement assays, disease-relevant phenotypic readouts, and increasingly, single-cell and spatial transcriptomics to confirm the target is expressed in the right cell type and tissue context implicated by disease biology.
The honest framing of LLM-assisted literature mining, as used by serious practitioners in the field, is as a triage and hypothesis-generation accelerator rather than a replacement for the discovery process itself: it compresses months of manual literature synthesis into hours, widens the net of candidates a small team can plausibly consider, and surfaces cross-domain connections a single specialist would be unlikely to notice — while the actual burden of proof still runs through the same genetic, mechanistic, and experimental validation gauntlet that has always separated a plausible hypothesis from a validated drug target.