Grounding a language model in cause-and-effect relationships makes its answers more complete, repeatable and checkable. That was the argument our Chief Scientific Officer, Ben Sidders, made at BioTechX in Basel this October, with benchmarks and a worked target-discovery example to back it.

Ben has since re-recorded the talk so you can watch it in full.

‍

Why ground a model?

An ungrounded language model often returns results that are not verifiable. You can't trace them, and you can't easily prove that it's providing correct and real answers. The model will confidently summarise a field, but it won't provide a really comprehensive review, and scientists are very keen to have all of the information, not just some of it. Most frustratingly, ask the same question twice and you'll get a different answer almost every time.

Adding a grounding layer that constrains the facts the model can reason over overcomes the majority of these challenges. That layer is the Galactic AI™ knowledge graph: over 300 million cause-and-effect relationships, curated by a suite of language models fine-tuned on over ten years of Biorelate's own manually curated data.

‍

What grounding changed

The talk compares the same model, on the same platform, one with the graph and one without. With the graph, it returned 6,490 entities downstream of EGFR, roughly an order of magnitude more, and 85–90% of its facts carried a verifiable reference.

Grounded in the graph, the model returns the exact same 6,490 entities every single time the question is asked. Without the graph, there's a real swing in what comes back. For drug repurposing, the graph provides a layer of truth to check against: if a drug–disease relationship doesn't exist in the graph, novelty can be asserted.

There's still work to do. On the most complex prompt, for novel target identification, the model decided to not cite anything rather than cite incorrectly, and the team will continue to work on very complex queries.

‍

A worked example: novel targets for kidney fibrosis

To show this in action, Opus 5 was grounded in the Galactic AI™ V3 knowledge graph and set to find novel targets for kidney fibrosis. The agent's system prompt was looking for novelty, clear causal rationales and explicit disease mechanism linkage.

The agent first identified 841 candidates with a direct and two-hop connection to the disease. It was able to drop 540 of these because they already had known biology for the disease; here again, the graph provides that layer of truth. If there is a publication on a target and kidney fibrosis, it isn't included in the analysis.

It then further reduced these to 220 with a correctly polarised causal path, so the activity of the target increases disease rather than any other effect. This was whittled down to 80 with high path confidence: a significant number of publications supporting each step in the causal path, multiple routes between the target and the disease, the correct direction of effect and no contradictory evidence. Finally, the agent condensed these into four mechanistic clusters for review.

One cluster centres on SLC39A14, a metal transporter: SLC39A14 regulates iron, which increases apoptosis, which increases renal fibrosis. Potentially this is a cluster of eight or so different genes that all have the same mechanistic hypothesis. The agent lists a set of considerations for each: whether the direction is correct, the level of uncertainty and any particular gaps.

Another is SOX18, a transcription factor which acts through PROX1 to regulate lymphangiogenesis, which has a role in renal fibrosis. This is quite a novel hypothesis with very little supporting evidence, and the model has flagged that.

The results link back to Galactic Web, where a scientist can see the path in action, click on each independent link and view the specific paragraph in the publication that supports each relationship. It really gives you the ability to dig into the mechanistic rationale and verify the output, returning compelling target hypotheses which are ready for further evaluation.

‍

Watch the full talk

Cause-and-effect data is a great grounding foundation for AI and for agents. It drives factual recovery and consistency, and fundamentally increases trust and utility for biology.

If you're prioritising targets or biomarkers, or building agents that need an evidence layer your scientists can verify, watch the full recording above. To see how grounding works on your own questions, book a demo or email us at info@biorelate.com