Summary: 10x Science raised a $4.8 million seed round led by Initialized Capital for a platform that automates mass-spectrometry characterization. Its hybrid chemistry algorithms and AI agents determine molecular composition and structure while retaining traceable records for regulatory compliance.
TechCrunch reported on April 22, 2026, that 10x Science — a startup founded in December 2025 — has raised a $4.8 million seed round. Initialized Capital led the round; Y Combinator, Civilization Ventures, and Founder Factor took part. The company skips candidate generation entirely; it takes on whatever comes after generation: characterization.
The context is simple. Protein-structure prediction — AlphaFold is the emblem, and the basis of the 2024 Nobel Prize in Chemistry — keeps inflating the number of candidates AI proposes. The Nobel committee divided that award between Google DeepMind’s Demis Hassabis and John Jumper for protein structure prediction and David Baker for computational protein design; both halves amount to engines for proposing molecules. Every candidate still must be measured, validated, and structurally confirmed before it can move into testing or mass production. Upstream generation is oversupplied, while downstream instruments and data handling never grew to match — and that mismatch is precisely the gap 10x Science is betting on.
Where the Pipeline Jams Once AI Has Proposed the Candidates
Co-founder David Roberts offers a funnel: “You can add as many candidates as you want to the top of the funnel” — yet characterization is a gate no molecule skips. Mass spectrometry is the standard way to establish composition and structure: it reads out a molecule’s mass and charge. It has long meant senior scientists comparing spectra by hand and iterating through trial and error — slow, grinding work. Raise upstream output tenfold with AI and the bottleneck lands entirely at this stage.
The Technical Route: Mass Spectrometry Paired with AI Agents
Under the hood, the platform is a hybrid: chemistry and biology algorithms that behave deterministically, working alongside AI agents trained on spectrometry data. A few design decisions stand out:
- Mass-spec data feeds straight into the platform; molecular composition and structure are determined automatically
- AI agents work across steps, chaining the analysis workflow together
- Analyses are traceable, preserving a full record for regulatory compliance
The first customers show the difference concretely. Matthew Crawford, a scientist with Rilas Technologies, sent one file up, and the platform recognized the protein from the file name alone, then pulled the matching sequence from online databases on its own initiative. That capacity to complete its own context is exactly what conventional mass-spec software lacks.
A $4.8M Seed and a Bertozzi Lab Lineup
The founders number three. David Roberts is the chemical biologist; Andrew Reiter, the biologist; and Vishnu Tejus, the serial founder with a background in computer science and AI. Their common root is the Stanford lab of Nobel laureate Carolyn Bertozzi — the group where the three studied how cancer cells and the immune system interact. The seed money funds engineering hires, model refinement, and a wider customer base; further out, the company wants protein structure data brought together with other cell data — Roberts calls it “a new way to define molecular intelligence.”
Who Uses It, and How It Charges
Rilas Technologies, where Crawford works, runs analyses on clients’ behalf, sparing them “several million dollars” in equipment spending; he has spent several weeks on the platform and says researchers who need mass-spec answers in a hurry get to “keep that can of worms closed.” The company also reports active work with multiple major pharmaceutical companies and academic researchers beyond Rilas.
On the business side, it is monthly SaaS. Zoe Perret of Initialized Capital is blunt about it: “This is a SaaS platform that pharma has to pay for, every single month.” Generation-side models bill per call; characterization, by contrast, runs as a continuous workflow built around compliance — revenue that sticks far harder.
Three Signals for AI-for-Science Teams
The case carries three signals. Value is drifting downstream: each step forward in AI makes candidate generation cheaper, and the scarce capability shifts to validation and characterization. The hybrid of deterministic algorithms and AI agents, meanwhile, survives regulatory review more readily than an end-to-end black box — in pharma, traceability is not a bonus but the price of admission. And $4.8 million is a small number by 2026’s AI funding standards, yet the “small capital, deep domain, subscription” route is a more practical template for AI-for-science teams unwilling to fight the giants over compute.
The founding team also teaches a hiring lesson: two domain scientists plus one AI-native builder, all from one Stanford lab, got a paid platform in front of pharma within months of the company existing. The model itself was never the scarce asset; what was scarce was having mass spectrometry and agentic software fluency at one table. Teams putting together AI-for-science bets should put that mix ahead of one more scaling engineer.