Formation Bio puts drug diligence inside ARK: Phyla attaches a checkable citation to every claim and records empty searches
Synopsis
Formation Bio has deployed Phyla, a drug-diligence tool invoked automatically by its internal AI system ARK, which queries 27 biomedical sources under one grammar, resolves identifiers once, and attaches a checkable citation to every factual claim. Assembling that evidence previously took teams manually querying sources and harmonizing gene names, while the literature systematically omits negative results. Across research reports for seven target–disease pairs, Phyla cited 99.7% of its statements, versus 86.5%–87.1% for Claude Science and GPT-Rosalind.
Interpretation
Evidence retrieval for drug diligence is now handled by Phyla, which Formation Bio's internal AI system ARK calls automatically, so a user asks in plain language and receives a cited answer. Combining target validity, mechanism biology, modality fit and the clinical landscape previously took whole teams manually querying sources, de-duplicating gene names across databases, and judging how far to trust each dataset given the conditions it was generated under. In a comparison generating research reports for seven target–disease pairs, Phyla cited 99.7% of its statements, against 86.5%–87.1% for Claude Science and GPT-Rosalind, and never fell below 99.3% in any case.
Phyla collapses different APIs, formats and query languages into one consistent grammar, translates gene identifiers once for reuse, and returns every tool's results in the same schema with sources attached. It is organized around the domains of asset diligence, such as target validity, mechanism of action and modality fit, rather than around data-source type, so a reviewer can audit the analysis in the language of the questions they actually ask. Phyla ships as a single tool that makes 90 underlying tools across 27 sources addressable in under 2,000 tokens, instead of one MCP server per source that would load roughly 90 tool schemas before a user has asked anything.
Every factual claim requires a numbered citation, each reference carries one provenance label, and an evidence grade runs from direct down to absent, where absent means explicitly documented as not found. The literature is systematically biased toward positive findings, so a negative result such as no GWAS hit for a gene tends to go unrecorded, and a model trained on that literature is likely to miss it. In an audit of claims, 92% of Phyla's verifiable claims were fully supported by their cited sources, compared with 80% for Claude Science and 67% for GPT-Rosalind, whose shortfall mostly came from sources backing only part of a claim.
Perspective
The demonstrated scope is Phyla as invoked through ARK inside Formation Bio, covering 27 sources and 90 underlying tools for asset-diligence questions such as target validity, mechanism of action and modality fit; the comparisons are limited to gene-panel fact queries and research reports for seven target–disease pairs.
The citation-rate and citation-support comparisons rest on seven target–disease pairs, and Phyla's citation rate never fell below 99.3% in any case; roughly 50% of its coverage does not appear in the published inventories of the frontier AI platforms it was benchmarked against, a statement about coverage rather than accuracy. When sources conflict, the system does not vote, average or pick a winner but returns both answers with their own attribution for a human to weigh; the cache window varies by source, and if one database is down it still returns everything else and flags what failed.