The knowledge graphs built up over years across different domains are constantly being rebuilt into more queryable, more interconnected versions — and it still takes roughly two weeks to manually connect the dots between interdisciplinary graphs. This is exactly where AI is reshaping the infrastructure of scientific knowledge.
In practice, it means a postdoc studying host–pathogen co-evolution could use AI to reason across host-species occurrence data, the relevant literature, and gene–phenotype associations for the pathogen — reconciling taxon IDs and matching gene identifiers across schemas, all from a single graph spanning ecology, genomics, and the literature.
Scientific Knowledge Graphs Were Built to Answer Known Questions
Knowledge graphs (KGs) are community investments that encode decades of relationships between concepts and entities. As more literature is published, database management only gets harder. The Global Biodiversity Information Facility (GBIF), one of the world’s largest open repositories of biodiversity occurrence data, holds a reported 2.8 billion occurrence records spanning 10 million species, curated from 1,600+ data publishers. It can tell you how many Panthera tigris occurrences were logged in Sumatra between 2010 and 2020. Monarch can return every known gene–phenotype association for a rare disease. PubMed KG 2.0 can surface the co-authorship network and citation clusters around a specific gene target. These are genuinely hard problems, solved well — to encode what was known and curated. Artificial intelligence now automates extraction, alignment, and multi-hop reasoning at scale.
A Quick Map of the Landscape Researchers Already Use
AI is changing what we can discover from the scientific graphs that already align with modern research, while retaining the central value of the graph ecosystem.
GBIF (Global Biodiversity Information Facility): Connects publishers across 197 countries, linking species and their observations to geographic locations, collection dates, taxonomic hierarchy, environmental metadata, and data provenance. Machine learning is already used for species ID and occurrence prediction; computer-vision models classify species from field images, and deep-learning models analyse bird songs and insect calls to improve biodiversity monitoring. More recent work aims to infer the hidden ecological relationships within billions of occurrence records — identifying migration shifts, predicting species distributions under climate change, and uncovering community-level patterns that would be impossible to detect manually. These models have helped scientists anticipate the next emerging ecological interaction.
Open Tree of Life (OTL) — building evolution’s largest connected graph: The value here isn’t only the evolutionary trees but a unified, queryable reference of 2.4 million species for researchers in comparative biology, conservation, and genomics. Natural-language-processing pipelines can now extract phylogenetic trees from published papers automatically, drastically reducing manual curation. The hard part for AI is handling conflicting tree topologies while preserving scientific uncertainty — something computational intelligence still can’t do without expert intervention.
Monarch Initiative: A biomedical informatics graph that associates 200,000 gene–disease–phenotype relationships across humans, mouse, zebrafish, fly, worm, and more. This cross-species design enables phenotypic reasoning rather than plain genetic lookup, letting researchers compare the observable characteristics of two equivalent genes to better understand the underlying biology. AI has strengthened semantic phenotypic embeddings to support rare-disease diagnosis, and as neural networks mature, Monarch increasingly surfaces non-obvious phenotypic matches across species — promising experimental directions before lab validation even begins.
PrimeKG: Unusually, PrimeKG was built as a machine-learning resource from the start, rather than a query tool later adapted for AI. It curates 20 biomedical databases to connect 10 major node types across drugs, diseases, genes, pathways, and side effects, drawing more than 4 million relationships. Its graph neural networks (GNNs), embedding models, and link-prediction algorithms generate candidates for the most promising repurposed drugs during pre-experimental design.
PubMed Knowledge Graph 2.0: To handle biomedical research at scale, this literature-derived graph transforms more than 30 million PubMed abstracts into roughly 72 million structured relationships connecting authors, publications, diseases, chemicals, genes, biological concepts, and institutions. It is heavily used for co-authorship mapping, citation-trajectory analysis, and concept co-occurrence. Large language models and biomedical NLP systems have enabled rapid expansion of structured, searchable relationships from newly published or unstructured abstracts.
These shifts mark AI becoming part of scientific knowledge itself — not just an analytical tool bolted on top.
How AI Is Changing Graph Construction: Scaling Extraction and Alignment
The data bottleneck: every graph above shares the same constraint — human curation can’t keep pace with scientific publishing, so graphs are always playing catch-up with new literature and data. AI addresses this at the core:
- Automated extraction: biomedical and large language models are already used in PubMed KG 2.0’s pipeline and in adjacent projects like SPOKE (Scalable Precision Medicine Oriented Knowledge Engine). They convert unstructured text into structured triples by identifying entities, verifying relationships, and mapping them to ontologies in a graph schema — flipping the historic process of domain experts reading papers by hand to establish relationships.
- Entity resolution across graphs: embedding-based matching and learned alignment models reconcile identifiers for the same biological entities across GBIF, Monarch, OTL, and other graphs (each with different taxon and node IDs). This is a prerequisite for any real cross-graph result at a tractable scale.
- Continuous ingestion: AI pipelines can run continuously over new literature and datasets, letting scientists spend more time interpreting results and less cleaning metadata.
The implication: graphs will evolve continuously to align with recent publications — but they stay exposed to reproducibility and opportunity risks until versioning and provenance are enforced.
How AI Is Changing Graph Reasoning: From SPARQL to Multi-Hop Inference
Traditional knowledge graphs tell you only what is already known; they fail to show the missing links that aren’t directly encoded, no matter how strongly the underlying data implies them.
Graph neural networks (GNNs) break through this by learning low-dimensional representations of nodes and edges to predict the missing links. Drug-repurposing predictions from GNNs on PrimeKG have held up against known pharmacology.
Natural-language interfaces break the SPARQL barrier. LLMs can now translate a natural-language research question into a structured query, execute it across different ontologies, and merge the results into an explanation. This is especially useful for postdocs and early-career researchers in interdisciplinary fields who have strong scientific intuition but limited graph-query expertise.
Multi-hop, cross-graph orchestration: AI orchestration layers, similar to GraphRAG in enterprise settings, make cross-graph traversal possible for a cross-domain question — without manual stitching.
What Becomes Possible for Researchers?
Applying AI in real research workflows is where the theoretical architecture starts to pay off.
Diagnosing rare diseases via cross-species phenotype matching is the most prominent case. Adding GNNs to the Monarch Initiative surfaces candidate genes through structurally similar gene–phenotype neighbourhoods, not just directly evident matches — meaningfully shortening the list of irrelevant hypotheses.
Drug repurposing at graph scale: GNN models trained on PrimeKG exploit structural patterns across drugs, targets, pathways, and side effects to surface compounds more likely to act on a given protein target or disease pathway — connections a literature search usually won’t yield, because they were never published explicitly.
Ecological forecasting from GBIF occurrence data: graph-based approaches model GBIF’s 2.8 billion records to understand how multi-species communities shift together, rather than fitting single-species models in isolation — often combined with climate and habitat data to train species-distribution models.
Automated literature synthesis: entering a new field, a postdoc or researcher accelerates sharply when they can generate structured maps of established findings, contested hypotheses, influential authors, emerging concepts, and unexplored gaps. Graph-grounded AI summaries stay verified and grounded, curbing the hallucinations you get from general LLMs.
The Open Problems: What AI Still Can’t Do Reliably
For all the progress, AI still has real limits.
Extraction errors: LLM-assisted relationship extraction is fast, but it makes mistakes that — if untracked — propagate into every downstream project trained on that graph. We need better tools to state confidently whether a relationship was established automatically or by a domain expert. Edge-level provenance (source, method, confidence) is essential.
Ontology alignment and identifier ambiguity at scale: alignment across graphs requires that an identifier refer to the same real-world entity everywhere. Embedding-based entity resolution is improving, but full cross-graph interoperability isn’t solved.
Hallucination in graph-grounded LLMs: grounding in a graph reduces hallucination substantially, but doesn’t eliminate it. When a query reaches beyond the graph’s knowledge, the model fills the gap with plausible-sounding, unverified inference — and those boundaries need to be clearly marked.
Access and reproducibility: for research to be reproducible, APIs must be stable, versioned, and queryable. Graph versions need continuous snapshotting and DOIs, as GBIF already demonstrates with DOI-based dataset versioning.
What the Next Five Years Look Like
Recent efforts — the Biolink Model, FAIR data initiatives, and platforms like Axy — are working toward cross-graph searchable datasets, so that GBIF, Monarch, Open Tree of Life, PrimeKG, and PubMed KG 2.0 can be queried as something closer to a unified system. AI accelerates that vision through relationship mapping and entity alignment, and it changes how early-career researchers navigate scientific knowledge across a queryable infrastructure.
So the next major discovery is unlikely to come from a new experiment or a bigger knowledge graph — it is more likely to come from connecting the missing dots in the ones we already have. That is AI’s real contribution to scientific infrastructure.
Axy is building toward exactly this: a connected layer over your lab’s existing knowledge-graph infrastructure. We’re inviting our first 500 researchers in small cohorts to help shape it. If that’s the future you want to work in, apply to join below.
