Biorelate are launching Galactic AI v3, our LLM-native curation pipeline, across our full range of products. The fine-tuned LLMs behind Galactic AI were trained on more than 10 years of expert, manually curated data. Compared with gold-standard human curation, they achieve 93% combined precision for entity recognition and ontology linking and 82% for relationship extraction. They also uncover 4.5x more relationships when benchmarked against traditional text mining approaches.

The knowledge graph now covers >60M documents and contains 3.5 billion linked biomedical entity mentions and 443M linked cause-and-effect relationship mentions. This release adds clinical trials, drug labels and research grants as sources, and links more of what it finds to public identifiers. Taken together, Galactic AI is a connected, mechanistic map of biology, built from cited evidence specifically for drug discovery.

AI is increasingly shaping the decisions that define a drug programme. On its own, though, a foundation model overlooks much of what is already known about a target and rarely gives the same answer twice. Galactic AI v3 gives those models something solid to reason from: a consistent, cited graph of cause-and-effect evidence. Hypotheses reach the lab faster, and the candidates that move forward do so with greater confidence.

‍

What Galactic AI v3 makes possible

Ground AI agents in causal biology

Ask an LLM what is downstream of a well-characterised target and you’ll get a summary, not a map. As a grounding layer, Galactic AI lets agents answer from a complete and consistent view of known biology rather than only what is in their training data. In our benchmarking against leading foundation models like Opus5.5, grounded agents were:

• Complete: returning 40x more known relationships for well studied targets like EGFR, PCSK9 and SCN10A
• Consistent: returning the same facts on every repeat query, versus only a 30% agreement between ungrounded queries.
• Verifiable: backing every claim with traceable citations biologists can check, rather than asking them to trust an opaque model

‍

Build a mechanistic rationale for your targets

Targets supported by a mechanistic rationale are twice as likely to succeed ^[1]. A therapy area lead building a target–disease rationale gets cause-and-effect relationships between drugs, genes and diseases in one place, without time-consuming manual searches. Galactic AI connects gene or hit lists to disease through directional, cited cause-and-effect paths. Trial conditions and interventions now map to the same ontologies as the literature, so each rationale can be weighed against the trials already registered in that indication.

‍

Assess safety and find biomarkers earlier

Biomarker and safety signals sit in the same graph as the mechanism. With every source curated into one ontology-linked graph, a finding can be traced from target to adverse event, with each step cited. Labels for drugs acting on a target can be reviewed alongside the cause-and-effect evidence for that target. A target safety review then draws on papers, patents, trials and labels in one pass.

‍

Integrate with your own data

For data science teams, the knowledge graph is structured to FAIR principles, mapped to open ontologies and standard identifiers, and available for direct download.It joins directly to internal data mapped to the same identifiers, such as screen hits, expression data and compound collections, so there’s no literature normalisation layer for you to build or maintain. That makes it well suited to AI and ML applications, systems biology, network analysis and the interpretation of omics and other high-throughput results.

‍

Galactic AI v3 in numbers

Compared with gold-standard human curation, Galactic AI achieves 93% combined precision for entity recognition and ontology linking, and 82% for relationship extraction on PubMed.

‍

Get started with Galactic AI

Galactic AI v3 works with your existing stack. Choose your integration:

• Galactic Data: relationships and entity annotations delivered as downloadable Parquet files for your own graph, warehouse or ML pipeline.
• Galactic API and MCP: REST and MCP access to connect our data to your apps and AI assistants. We host and maintain the server, so there’s nothing for you to update.
• Galactic Web: a co-scientist workspace for navigating cause-and-effect paths and searching for biomarkers, with cited evidence.
• Galactic Processing: deploy our models inside your own infrastructure to curate your licensed and proprietary content.

‍

References‍

[1] Morgan P et al. Impact of a five-dimensional framework on R&D productivity at AstraZeneca. Nature Reviews Drug Discovery (2018). https://doi.org/10.1038/nrd.2017.244