a16z and Menlo rarely co-lead seed rounds. Their $13.5M bet on Phylo explains why.

A Rare Seed Bet Places Phylo at the Center of AI-Driven Biology

Phylo disclosed its $13.5 million seed round on February 3, 2026, co-led by Andreessen Horowitz and Menlo Ventures through the Anthology fund linked to Anthropic. Zetta Venture Partners, Conviction, SV Angel, and Valkyrie also participated. The valuation remains undisclosed.

The capital funds what Phylo calls an Integrated Biology Environment, a single workspace where a large-language-model agent plans multi-step experiments, searches literature across databases, generates analysis code, and dispatches jobs to lab instruments. The company claims more than 300 connectors to scientific databases and tools including COSMIC, Addgene, PubChem, and ChEMBL. Part of the capital underwrites security audits to meet biosecurity guidelines.

The round's structure signals where top VCs are placing bets. a16z Bio + Health general partner Jorge Conde argued that scientific AI will transform discovery the way cloud transformed software. Both a16z and Menlo publicly praised Phylo's alignment with actual scientist workflows rather than generic chat interfaces. That distinction matters: the incumbents (Benchling, Dotmatics, Labguru) dominate electronic lab notebooks but offer no autonomous planning. Phylo is betting that gap is the opportunity.

The thesis rests on a specific claim about bottlenecks. Phylo's founders say biology slows not from a lack of model intelligence but from execution infrastructure, the manual stitching of literature search, protocol design, code generation, and instrument orchestration. Ginkgo Bioworks, an early adopter, said it cut 10 cell-painting analyses from weeks to hours. Phylo's internal benchmarks show a 20% higher task completion rate versus baseline biology agents, though the methods remain proprietary.

Scale is harder to verify. The company reports thousands of monthly queries from alpha users and claims 18 of the 20 largest pharmaceutical companies as customers, but has not named them. The 7,000-lab figure traces back to the earlier open-source Biomni stack, which spread organically across academia before Phylo incorporated. Independent security audits are pending, and the system's protocol-hallucination risk (AI agents generating erroneous experimental steps) remains an open question that no audit has yet closed.

What's not in dispute: a16z and Menlo rarely co-lead seed rounds. When they do, the bet is that a category is being born.

The Stanford Duo Rebuilding Scientific Discovery as a Software Problem

Phylo's co-founders, Kexin Huang and Yuanhao Qu, are not typical biotech founders. Huang completed his PhD in computer science at Stanford, where his graduate work focused on building an AI scientist, a project he'd pursued for nearly a decade before it had a name. Qu is a cancer biology PhD whose experimental and computational training gave him direct experience with the fragmentation of modern research workflows. Together they built Biomni, an open-source biomedical research agent that spread through academia and into hospitals, biotech, and pharma without a traditional go-to-market push.

That pairing, a CS researcher who thinks about scientific discovery as an orchestration problem and a working biologist who lives the pain of fragmented tools, is the core signal. Phylo's founding team also includes scientific co-founders Jure Leskovec, a Stanford professor of computer science, and Le Cong, a Stanford professor of pathology. Advisors include Feng Zhang, Carolyn Bertozzi, and Fabian Theis. The composition matters: this is not a biology company hiring a few ML engineers. It is a team where the default assumption holds that the science problem is a software problem.

Huang's public writing makes the philosophy explicit. In a LinkedIn article titled "The Integrated Biology Environment," he argued that biologists enter the field to understand disease, not to stitch together PDFs, spreadsheets, and clunky database portals. The a16z investment thesis echoes this directly: Jorge Conde and Zak Doric wrote that biology never got its equivalent of the IDE, the integrated development environment that made software engineering productive, and that Phylo is building exactly that layer.

Qu's own public comments reinforce this. In a LinkedIn post about a research collaboration with Scale AI on evaluating coding agents for drug discovery tasks, he wrote about the gap between how AI performs on benchmarks and how it performs on real-world scientific work. That is the exact problem Phylo is trying to solve: not whether an agent can answer a question, but whether it can run a workflow end-to-end in a production lab environment.

The team is growing. Phylo's launch post listed founding team members Serena Z., Tianwei She, Zixin Huang, Gregory Minevich, Margaret Hua, and Malay Gandhi, and the company is actively hiring. The roles it recruits for, spanning computational biology, AI, and research engineering, reflect the same thesis: the next generation of life-science companies will be built by people who can operate across both domains.

What an Integrated AI-Native Biology Environment Actually Looks Like

Phylo's Biomni Lab takes the form it does because the problem it solves is fundamentally about context. Biologists at top labs juggle 82 biology tools, 68 databases, and over 100 software packages daily. Moving data between them means reformatting files, debugging scripts, and wrangling dependencies, work that has nothing to do with science. Biomni Lab wraps all of that behind a single natural-language chat interface, so a researcher can upload a protein's chemical composition and ask the system to simulate its 3D structure without opening another tab.

The architecture borrows from a concept most engineers already know: the IDE. Integrated development environments collapsed the editor, compiler, and debugger into one workspace. Phylo's bet is that biology needs the same consolidation. The platform's agent can plan a multi-step workflow, select the right tool for each step, run the analysis on cloud infrastructure with GPU acceleration when needed, and generate visualizations (charts or 3D molecular models) from the output. For long-running or parallelizable tasks, it distributes work across multiple graphics cards.

What makes this different from a chatbot bolted onto a tool registry is what happens after the results come back. Biomni Lab runs an automated hallucination check on its outputs. When the agent wrote code to produce an analysis, the platform surfaces that script alongside a natural-language explanation of each issue it flags, so a human reviewer can verify the logic without reading raw output. Every step is traceable. The system also maintains personalized memory across sessions, adapting to individual researcher preferences over time.

The open-source Biomni project, released in June 2025, reached more than 4,300 organizations before the commercial product launched. Ginkgo Bioworks used the platform to accelerate more than 10 cell-painting and transcriptomic analyses that previously took weeks, completing them in hours with results its own scientists validated as publication-quality. That case study is the concrete proof point Phylo leads with.

The commercial Biomni Lab layers enterprise security on top: ISO 27001 certification, SOC 2 Type 2 compliance, sandboxed sessions for data privacy. The free tier has daily usage limits; paid editions add custom agent support and additional cybersecurity controls. For engineers evaluating the stack, the more interesting question is what Phylo's architecture implies about where agentic AI actually meets physical science — and where the hard limits still hold.

Why a16z and Menlo See Biology as the Next AI-Native Vertical

Phylo's seed round didn't happen in a vacuum. It landed in a venture ecosystem that has spent the last 18 months repositioning biology as the next frontier for AI-native infrastructure, and the two firms leading that bet have made their thinking explicit.

a16z's Bio + Health team frames its thesis around a concept it calls "Eroom's law," the observation that drug development costs have risen exponentially over decades, the inverse of Moore's law. Vijay Pande, the former Stanford professor who leads a16z's bio practice, has argued that AI inverts this curve: by turning biological research into a computational problem, the cost and time of discovery collapse toward zero. The firm's 2025 fundraise allocated $700 million to Bio + Health specifically, part of a larger $15 billion war chest the firm describes as a bet on "the architectures of the future."

Menlo Ventures moved in parallel. After leading Anthropic's early rounds, Menlo closed a $3 billion fund in 2025, the largest in its 50-year history, with a mandate spanning seed through growth-stage AI companies, including healthcare and biotech. Its partnership with Anthropic on a $100 million AI startup initiative gave it a direct pipeline into the models that companies like Phylo intend to deploy against biological workflows.

The logic both firms work from is straightforward. Drug discovery and lab operations remain stubbornly manual: PhDs spend hours pipetting, labeling, and cross-referencing results across disconnected instruments and software systems. If AI agents can orchestrate those workflows the way they now orchestrate code generation or customer support, the productivity gain is enormous, and the company that owns the orchestration layer owns the vertical.

This is why Phylo's pitch as an "operating system for AI-driven biology" resonates with these investors. It's not selling a single algorithm or a diagnostic tool. It's selling the infrastructure layer on which an entire AI-native biology stack runs, and that's the kind of platform bet a16z and Menlo have both signaled they want to back.

The FDA's January 2025 draft guidance on AI models in drug submissions added regulatory tailwind. The agency has reviewed more than 500 submissions containing AI components since 2016, and the new framework gives companies a defined path to credibility assessment. For VCs, that reduces a key risk variable: the regulatory environment is no longer a black box.

Can AI Agents Actually Automate Scientific Discovery?

The question hanging over Phylo's raise isn't whether AI can help biologists. It's whether AI agents can close the entire loop — hypothesis to experiment to analysis to new hypothesis — with enough reliability to trust them with real lab time and real money.

The research says: not yet, but the gap is narrowing faster than most people outside the field realize.

A multi-agent system called Robin, built at FutureHouse, became the first AI system to autonomously discover and validate novel therapeutic candidates in an iterative lab-in-the-loop framework. Published in Nature in May 2026, Robin identified ripasudil, a clinically used ROCK inhibitor never previously proposed for dry age-related macular degeneration, as a promising candidate, then designed and analyzed a follow-up RNA-seq experiment to explain the mechanism. Every hypothesis, experimental direction, data analysis, and figure in the main text was produced by the system. The humans ran the wet lab. The AI did everything else.

That's a milestone, but it comes with a caveat that matters for anyone evaluating companies like Phylo. Robin worked in a narrow, well-defined therapeutic area with established assays and clear success metrics. The system didn't need to invent new experimental methods or reason about biology nobody understood. It searched, proposed, tested, and interpreted within boundaries its designers set.

The broader survey literature draws the same boundary. A September 2025 Nature Machine Intelligence comment from researchers at Virginia Tech, Carnegie Mellon, and MIT laid out the foundations of "agentic science" and was blunt about current limitations: these systems can reason, plan, and interact with digital and physical environments, but their reliability drops sharply when tasks require deep domain-specific knowledge or when they encounter conditions outside their training distribution.

The ICLR 2025 survey on agentic AI for scientific discovery, covering dozens of systems across chemistry, biology, and materials science, found a consistent pattern. AI agents performed well in data preparation, experimentation, and report writing. But the literature review phase, the step where a system needs to synthesize what's actually known and identify genuine gaps, saw the highest failure rate across nearly every framework examined. Agent Laboratory, one of the more capable systems, saw significant performance drops during literature review. ResearchAgent could generate novel ideas but couldn't perform structured literature reviews to ground them.

This is the exact problem space Phylo targets with its integrated biology environment. If the company can build an AI system that reliably reads, synthesizes, and reasons over the biological literature, connecting it to experimental design and execution in a closed loop, it would solve the step where current agentic systems break down most often.

The Stanford Agents4Science conference, which ran its first edition in October 2025, exists precisely because the community doesn't yet know the answer. The conference requires AI to serve as both primary author and peer reviewer of research papers, creating a controlled environment to test where these systems succeed and where they fail. The advisory board includes the chief editor of Nature Biotechnology and Nobel laureate Guido Imbens, people who take the question seriously enough to build an institution around answering it.

The honest state of the field as of early 2026: AI agents can automate significant portions of the scientific process in well-scoped domains. They can generate hypotheses, design experiments, analyze data, and write papers. What they cannot yet do is operate as reliable autonomous scientists across the full breadth of a discipline, particularly when the task requires inventing new methods, reconciling contradictory evidence across fields, or knowing when an experimental result is wrong rather than just surprising.

Phylo is betting that biology, with its structured ontologies, massive public datasets, and assay-driven validation, is the domain where those limitations fall first. a16z's data shows the $13.5M says a16z and Menlo agree it's worth testing.

What Phylo's Hiring Reveals About the Emerging AI-Biology Workforce

Phylo's open roles, and the 155 applicants who piled into its computational biology listing in a single month, read like a spec sheet for an entirely new job category. The South San Francisco role sits at $200K–$300K base, senior level, and demands a PhD-level computational biologist who can also write production code, build evaluation pipelines, and face enterprise biotech customers. That's not a scientist. That's not a software engineer. It's both, and the market is paying a premium for the overlap.

The job description makes the hybrid explicit. One half is constructing "bio skills," curating the biomedical workflows, tools, and data integrations that power Phylo's agents. The other half is designing benchmarks and structured evals that measure whether those agents meet the standards of working scientists. The candidate has to move between a pharma customer's scientific review and an AI team's model-improvement loop without flinching.

This pattern shows up across the broader market. The following table summarizes comparable salary ranges and market data points referenced in the article:

The common thread: companies aren't looking for ML researchers who happen to read genomics papers. They're hiring people who can ship agentic software that scientists will actually trust with real workflows.

When a seed-stage startup prices its computational biology hire against Anthropic's research scientist compensation, the signal is clear: people who can operate at the intersection of agentic AI and wet-lab biology are among the most contested talent in the current market.

For engineers and scientists reading the tea leaves, the practical takeaway is specific. The roles that command these premiums reward three things in combination: a strong command of biomedical data and workflows, enough engineering skill to build and evaluate pipelines, and comfort working inside AI-native tooling. A PhD alone won't clear the bar. A software engineer who can't talk to a pharma customer about their scientific needs won't either. The hiring spree Phylo just funded is the proof point, and the talent window it's racing to fill won't stay open long.

Working in frontier tech? Zero G Talent tracks the openings: browse frontier tech jobs, openings at ASML, Stripe and Waymo, and the people building the field.