In a controlled benchmark, across four established drug targets, Galactic AI v3 captured up to 185x more biological relationships than GPT-5
A controlled, side-by-side benchmark reveals how much biology a general-purpose model actually misses.
As foundation models like GPT-5 become more powerful and more accessible, drug discovery teams are increasingly tempted to query them directly for target and disease biology insight. But broad, general-purpose training doesn't guarantee comprehensive or reliable domain coverage. Missing key biology at the target identification stage carries a real cost: overlooked mechanisms, incomplete risk assessment, and downstream attrition that erodes pipeline value. The open question is whether a general-purpose LLM can be trusted to surface the full picture of a target's biology – or only a fraction of it.
Biorelate designed a head-to-head benchmark, posing the same standardised query – "what does [target] regulate or affect?" – to Galactic AI v3 and to GPT-5 via the OpenAI API (three repeats per target, to test consistency), across four well-established drug targets: EGFR, PCSK9, GLP1R and SCN10A.
Galactic AI v3 captured 16x to 185x more relationships than GPT-5 across all four targets. For EGFR alone, that gap widens to 13,667 relationships – including well-evidenced biomarker and drug-binding relationships (EGFR's role in non-small cell lung carcinoma and glioblastoma, and its binding to cetuximab and osimertinib) that GPT-5 never surfaced in any of its three repeat runs. The table below breaks the gap down target by target.

Relationships like these are exactly what a biologist would otherwise have to find by hand – cross-referencing biomarker literature, binding data, and pathway regulation evidence across thousands of publications. A general-purpose model missing this evidence doesn't just return an incomplete answer; it returns an answer that looks complete, with nothing to signal that anything is missing. That's a harder failure mode to catch than an obviously wrong answer, and a costlier one to catch late.
To test consistency, GPT-5 was queried three times per target using the same standardised question. For GLP1R, only 13 (29%) of the 45 entities GPT-5 identified were present in all three runs, the remaining results were returned inconsistently. A model that returns materially different answers to the same question is unpredictable, and that variance makes it a difficult foundation for a repeatable analytical process. It also rules the model out for anything operating under regulatory scrutiny, where the ability to reproduce a result is a baseline requirement.
Book a demo to compare general-purpose AI output with Galactic AI’s structured, evidence-backed view of target biology.