Phyla: Grounding Drug Diligence in Cited, Contextual Evidence

Formation Bio acquires and develops clinical or IND-stage drug assets, and one of the most consequential decisions we make is whether to pursue a specific drug. To make this decision, biologists need to combine target validity, mechanism biology, modality fit, and the clinical landscape - evidence that lives in dozens of unrelated databases across inconsistent formats, identifiers, and blind spots. Historically, pulling that data together has taken entire teams of people manually querying sources, de-duplicating and harmonizing gene names across databases, and judging how much to trust each dataset given the conditions it was generated under.

This fragmentation is a consequence of how biological knowledge is created. Every dataset comes from a real-world experiment, and experiments are expensive to run at scale. Experiments are run under different conditions, recorded in different formats, and require expert interpretive judgment about context and quality, which is difficult for a human analyst to reason with at scale. It's arguably harder for an AI model, which can move quickly across sources, but has no inherent sense of which numbers deserve skepticism - and no habit of noting when a search comes up empty. For example, in biomedical literature, a negative result, i.e., "no GWAS hit for this gene", is the kind of insight that often goes unrecorded, which means that an LLM trained on that literature is likely to miss it.

What Phyla does

We built Phyla to systematically query biomedical datasets, standardize their outputs, and make gaps in the evidence explicit. It targets four concrete failure modes:

Fragmentation. Biomedical evidence is spread across databases that each speak their own API dialect. Phyla aggregates across them so a user isn't manually switching between tools.

Missing negative signal. Literature is systematically biased toward positive findings, so absence of evidence tends to disappear rather than get reported. Phyla explicitly records when a source was checked and came back empty, and identifies which source it was.

Identifier inconsistency. The same gene, disease, or compound can carry different names in different databases. Phyla has built-in ontology resolvers so a query can cross databases without requiring a human to translate by hand.

Hallucination. Generative models are especially prone to inventing or misattributing references. Phyla counters this by tying claims to a documented, checkable source.

For diligence, a biological signal is only useful in context. Whether a drug is safe and effective depends on where its target is expressed, how tightly it binds, how the drug is delivered and which tissues it actually reaches, and how it's metabolized. No single assay captures all of this, so diligence depends on combining evidence across multiple sources. Because so much depends on the accuracy of this information, source traceability is essential: reviewers need to know where each claim came from before acting on it.

How Phyla works

Phyla is delivered through ARK, Formation Bio's internal AI system. Users do not need to install or learn a separate tool; ARK invokes Phyla automatically when a question calls for it. A typical interaction starts as a plain-language request - for example, "Evaluate IL4R as a target for asthma."

ARK routes that into a sequence of underlying Phyla calls - such as pulling the gene record for IL4R, searching disease ontologies for the condition, and searching drug databases for relevant approved therapies. The result comes back as a cited answer. Users can access output either as structured JSON or, more commonly, as natural language, depending on what the task calls for. This means our teams can now quickly aggregate internal and external evidence that used to require manually querying each source and reconciling the results by hand.

Implementation

The moment a user asks a question or sends a request through ARK, Phyla begins to query across all available sources. Rather than organizing around data source type, Phyla is structured around the domains of asset diligence - target validity, mechanism of action, modality fit, and so on. This framing lets a reviewer audit the analysis in the language they actually think in, such as, "is the mechanism supported in patients?" rather than in the language of infrastructure, like "which API or dataset are we missing?" When choosing databases to integrate, we cast a wide net with a deliberate focus on questions relevant to drug development rather than just drug discovery, because of our focus on asset diligence.

A separate analysis layer does pure computation with no network calls, which makes it testable, usable from a notebook, and guaranteed to return the same numbers no matter where it runs. This is particularly important when we expose data to users who may not have domain expertise - it provides guardrails to ensure robust data interpretation. These kernels are what let Phyla answer questions like how specific a target's expression is, or whether that expression is concentrated in a distinct subpopulation, which in turn shapes whether patient stratification is realistic.

Different APIs, formats, and query languages are collapsed into one consistent grammar. Gene identifiers are translated between naming systems once and reused, rather than re-translated every query. Every tool returns results in the same schema, with sources attached, and a display layer aliases equivalent field names so two databases read as one shared vocabulary in a single answer. Every result is structured to include the database name, a direct link to the exact record, the version, the identifiers used in the query, and the date it was accessed. The final reports require a numbered citation on every factual claim, one provenance label per reference (literature, external database, internal database), and an evidence grade ranging from “direct” down to “absent”, meaning explicitly documented as not found. When two sources conflict, Phyla deliberately doesn't vote, average, or pick a winner. Collapsing disagreeing sources into a single number would destroy the exact information a diligence analysis needs. Instead, both answers are returned with their own attribution, and the human reviewer decides how to weigh them. The system balances response speed against staleness with a cache window that can vary by source. If one database is down, Phyla still returns everything else it found and flags what failed.

When building the architecture of Phyla, we opted to take a CLI-first approach instead of the more obvious alternative of one MCP server per data source. That would have meant loading dozens of tool schemas (roughly 90 tools in Phyla's case) before a user has even asked a question, significantly bloating the context. Instead, we leaned on the observation that LLMs are capable CLI directors when given clear grammar guidelines, so we shipped Phyla as a single tool that makes 90 underlying tools across 27 sources addressable in under 2,000 tokens.

How Phyla stacks up against other tools

Formation weighed building our own aggregation tool against adopting the growing ecosystem of external biomedical MCPs and AI workbenches. A few considerations pushed us toward building – we found that most public tools are oriented toward drug discovery rather than drug development, which is where our actual diligence questions live. We also needed internal data to sit alongside external sources, and building in-house meant we would be able to switch between frontier models as they advance, rather than being locked into one provider.

That decision shows up directly in the data that Phyla can draw upon. Phyla's connected data environment offers broader coverage for clinical development and commercial diligence. Roughly 50% of Phyla's coverage does not appear in the published inventories of the frontier AI platforms we benchmarked against. Phyla’s non-overlapping coverage adds clinical and real-world evidence, commercial intelligence, indication-relevant omics, and specialized genetics.

We evaluated the extent to which Phyla improves fact retrieval efficiency, as compared to the latest frontier models. We measured the speed and token efficiency for querying facts for a gene panel, like protein sequence length, tissue expression, and sub-cellular localization. We saw that Phyla enables 24% faster responses, while maintaining accuracy, and 39% fewer output tokens for the same facts.

For more complex research tasks, we compared Phyla with Claude Science and GPT-Rosalind on generating research reports for seven target–disease pairs. Phyla cited its statements more often and more consistently (99.7% vs 86.5–87.1%), never dropping below 99.3% in any case, while the other two platforms fell as low as 71% (GPT-Rosalind) and 79% (Claude Science). Those citations also held up better: 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. Phyla also cited more references – a median of 40, vs 29 (Claude Science) to 31 (GPT-Rosalind) and matched or exceeded the competitors on recent literature, with about 10% of its cited PubMed papers published in the past year versus 1% (GPT-Rosalind) to 9% (Claude Science).

Phyla also cited more references – a median of 40, vs 29 (Claude Science) to 31 (GPT-Rosalind) and matched or exceeded the competitors on recent literature, with about 10% of its cited PubMed papers published in the past year versus 1% (GPT-Rosalind) to 9% (Claude Science).

To understand how Phyla is being used in day-to-day work, we looked at who uses it and how often. In the six months since launch, 91 people at Formation - more than half the company - have used Phyla, and 63% of them came back in more than one month. The heaviest users are the people closest to acquisition decisions: nearly all of Business Development uses it to research and evaluate potential deals. Clinical Development uses it to pressure-test mechanistic hypotheses on programs already in development, and our engineers call it at high-volume when building screening tools.

Why it matters for Formation Bio

Diligence decisions are made based on evidence that's inherently scattered and often incomplete. Phyla wasn’t built to resolve this inherent uncertainty, but rather to make it visible and auditable. It preserves disagreement between sources instead of averaging it away, documents absence of evidence as carefully as it documents presence, and attaches a checkable citation to every claim it makes. This is what allows a human reviewer to rely on an AI-assisted analysis rather than second-guessing where each number came from.

Because Phyla is delivered through ARK, that capability isn't confined to a specialist tool a handful of people know how to use. It’s available company-wide, automatically, to anyone asking a relevant question. Because it's built specifically around Formation Bio's own diligence workflows and its unique internal data, it compounds: every new dataset the company builds up, and every diligence question the team asks makes the next answer smarter and more grounded.