Rivercell's €22M Bet: Build the Lab Before the Model

Eric Durand joined Rivercell in 2026 for a room that did not exist yet. The company had a thesis about how cells respond to drugs, and a plan to film those responses at scale. It did not have instruments that could film them.

In the first week of October, Rivercell came out of stealth with a €22 million seed round — about $25 million — led by HV Capital. HCVC, Alven and Bpifrance Digital Venture joined. The money is earmarked for a bigger wet lab in Paris, the data platform inside it, and a modelling programme that depends on both.

Yann Fleureau started the company in summer 2025. He had co-founded Cardiologs, the cardiac AI diagnostics firm Philips acquired in 2021. Durand arrived as co-founder and chief scientific officer after seven years leading oncology data science at Novartis, a spell as chief data science officer at Owkin, and a co-founding role at Bioptimus, which raised $76 million for biology foundation models.

Most AI drug discovery pitches begin with a model and hunt for data. This one begins with a laboratory. That order is the investment case, and the main reason the seed is worth arguing about.

What a virtual cell is supposed to predict

The pitch is narrower than the name suggests. A virtual cell, in Rivercell's framing, takes a starting cellular state, applies a defined perturbation — a drug, a genetic change, an environmental shift — and predicts a measurable result.

Public atlases mostly record associations. They show which genes move together in untreated cells. They rarely show what happens after an intervention, over time, across several measurement layers. Rivercell's manifesto lists four such gaps: association instead of intervention, one modality instead of many, single time points instead of trajectories, too few cell types.

Its answer is a proprietary loop. Live-cell imaging runs through each experiment. Endpoint gene expression is read afterwards. The model proposes the next experiment it is least sure about, the lab runs it, and the result updates the next prediction.

That is why the company calls its hardware the “GPU for the bio data center”. It is also why the seed is not really a model round.

Why the data, not the model, is the moat

AlphaFold is the obvious comparison, and the company makes it. Protein structure prediction worked because decades of public crystallography data already existed. There is no equivalent archive for how human cells respond to treatment over time.

So Rivercell plans to manufacture the archive. That is expensive, slow, and hard to copy if it works. It is also the part of the plan investors are underwriting.

CHRONICLE: Rivercell
────────────────────────────
2021       · 💰 Philips acquires Cardiologs
2025 SUMMER · 🧪 Rivercell founded in Paris
2026       · 👤 Eric Durand joins as co-founder and CSO
2026 OCT   · 🚀 €22M seed led by HV Capital
◉ NOW      · 🔬 Wet lab expansion, data platform
🔥 NEXT     · First public benchmark or model release

The missing piece is data: how cells change over time and under treatment, seen through multiple lenses and at multiple layers, captured at scale.— Yann Fleureau, co-founder and CEO, Rivercell

The advisory board carries weight for a seed-stage company. Fabian Theis directs the Computational Health Center at Helmholtz Munich. Charlotte Bunne runs a group on AI in molecular medicine at EPFL. Both work on the exact problem Rivercell is selling: predicting perturbation responses rather than describing steady states.

Angel investors come from operator roles at Twist Bioscience, DNA Script, Agilent and Abivax/Ipsen. Their presence says something about instrument and assay credibility. It does not substitute for a benchmark.

The $1.8 billion programme next door

The same week, a coalition led by Biohub committed nearly $1.8 billion to open, AI-ready biological datasets. The Virtual Biology Initiative is anchored by $500 million from Biohub: $400 million for measurement technologies such as cryo-electron tomography and high-throughput tissue microscopy, and $100 million for external research.

The US Department of Energy adds more than $500 million over five years through its Genesis Mission, including exascale computing and autonomous laboratory infrastructure. NIH contributes datasets built with more than $500 million of prior federal money. Google DeepMind, Isomorphic Labs and Meta put in $300 million combined. NVIDIA supplies accelerated compute.

Announced in the first week of October 2026. Figures as disclosed.

Read together, the two announcements cut both ways. The coalition validates the research direction and lowers the cost of talent and tooling. It also finances the open alternative to a private moat.

The competitive field is crowded already. Bioptimus raised $76 million for biology foundation models. Cellular Intelligence, formerly Somite and backed by the Chan Zuckerberg Initiative, acquired Novo's Parkinson's cell therapy programme on may 2026. Chai Discovery announced a GSK collaboration on October 9 after a wet-lab evaluation. Talus Bio shipped a structure-free model for proteome-wide screening on October 1.

Rivercell's differentiator is the closed loop between its own instruments and its own model. That is a manufacturing claim as much as a science claim.

What €22 million buys before the first benchmark

Three things are undisclosed: the size of the training dataset, any prediction benchmark, and a launch date for the virtual cell. That is normal at seed. It also means the round is priced on the team and the plan, not on evidence.

Biotech has a long memory of automation promises that arrived before working assays. A seed cheque for a data platform is a bet that this loop closes faster than the last one.

Rivercell's structure is less exposed to that specific failure. It is not selling a diagnostic device or a clinical candidate. It is building a data-generation engine for itself, with pharma partnerships as the eventual revenue test.

The nearer risk is technical. Reproducible responses across cell types and treatments are hard. Live-cell imaging generates enormous files. Assay drift can poison a dataset quietly. A €22 million seed funds a lab build-out and a first proof point, not a decade of atlas construction.

As we wrote in October, Reflector Bio's phenotypic screening claims rested on five compounds and an 80% hit rate — narrow evidence presented carefully. Rivercell is at an earlier stage: no readout at all, just a loop that has to start producing one.

For an investor, the Series A trigger is not a bigger model. It is a published benchmark showing that predictions trained on Rivercell's own data beat public baselines on held-out perturbations.