Podcast.
中文

The world is too loud. Read what matters.

Latent Space

The virtual cell challenge upset: linear models matched most deep learning models

A top virtual cell competition proved that simple linear regression can match many deep learning models—the problem is not insufficient data information, but flawed evaluation metrics and training objectives themselves.

Virtual cellsPerturbation predictionFoundation modelsEuropean startup ecosystemBiological dataWorld models
High information density: from the virtual cell competition's upset to concrete numbers on Germany's biotech startup ecosystem, for anyone wanting to understand the real progress in this space—though you can skip the guest's opening minutes of modest preamble.

The argument · tap a timestamp to hear it

2:13

Virtual cell interest is heating up because of sequencing throughput, not algorithmic breakthroughs

Biohub just announced an $1.8 billion collaboration with DOE, NIH, Meta, and Google to generate the multi-scale biological data needed for next-generation virtual cells. Fabian sees the industry's renewed interest in virtual cells driven first and foremost by the explosion in data generation scale, not algorithmic breakthroughs—though he warns that many projects are just ‘slapping an AI label on themselves’ rather than actually solving the cell-type-specific screening problem pharma needs.

— Fabian Theis
6:25

One experiment measured 10 million transcriptomes

Past perturbation experiments had far more cell types than perturbation conditions—typically just ‘control vs disease’ sparse comparisons. Parse Bio's giga lab measured 10 million transcriptomes in a single run; now hundreds of cell lines and thousands of drugs are tested in full perturbation combinations simultaneously, making the perturbation space as rich as the cell space for the first time—and that's what finally gives large-scale perturbation prediction models something real to work with.

— Fabian Theis
16:21

Inductive bias doesn't improve prediction; data has the final say

On whether to add manual priors (inductive bias) to models for interpretability and robustness, Fabian's stance is clear: the literature repeatedly shows that too-strong priors sacrifice predictive power. He cites AlphaFold—early versions laboriously hand-coded the molecule's rotational invariance, but later versions let the model learn that symmetry from more data, and they became more flexible and effective. Long-term, models that rely more on data and less on manual constraints will win out; interpretability can only be added after the fact.

— Fabian Theis
20:56

Virtual cells' bottleneck is the training objective, not the data

Early foundation model papers didn't seriously demonstrate scaling laws, which is why simple models could match them. But on harder tasks (like NicheFormer requiring prediction of a neighbor cell's state from one cell's expression), nonlinear models show real advantages. Fabian's scConcept model now scales to 350 million cells, about 100 million parameters, covers 16 species, and shows clear scaling behavior—proving the information is in the data, but prior training objectives and representations weren't truly leveraging it.

— Fabian Theis
26:11

The virtual cell challenge upset: linear models matched many deep learning models

Arc Institute's Virtual Cell Challenge was like a cold splash: on basic tasks like cell type prediction, logistic regression was nearly optimal; on perturbation prediction, many ‘fancy’ deep models didn't clearly beat simple linear ones. The championship solution used flow matching and similar nonlinear methods, proving complex models help on harder tasks—but the competition's fixed metrics let participants game the leaderboard, so the Open Problems consortium now requires teams to submit additional evaluation metrics to truly expose model differences.

— Fabian Theis
33:39

Germany's startups don't lack early capital; they lack risk-bearing later-stage VCs

Fabian thinks Germany and Europe have no shortage of early-stage funding (early VC, non-dilutive grants)—new EU startups now match US numbers, and Munich is Europe's hub for startup numbers. But when you get to larger bets like Series B, European VC clearly lags; many teams turn to US capital, which is Europe's most obvious fault line in deep tech.

— Fabian Theis
38:59

Most ‘virtual cell platform’ companies quietly end up making drugs

Fabian sees a pattern repeat: many companies claiming to ‘decode biology’ and ‘build virtual cell platforms’ get big funding but can't clearly explain the business model, so eventually they pivot to ‘we're making drugs’—because investors understand that story better. By contrast, he sees Noetic taking a more solid path—moat isn't just the model but unique perturbation data generation, like their Cancer Organoid Perturbation Atlas co-built with pharma partners willing to pay directly for early data access.

— Fabian Theis
46:42

Let models reverse-engineer the experiments you should run

Fabian's team's new Lab Compass is an inverse generative model: given a forward world model, it reverse-engineers what lab work you need to hit a target cell state. The team has run three or four ‘experiment-model-experiment’ fast loops with a Cambridge collaborator doing hematopoiesis research, though Fabian admits the full loop is still far from robot automation—mostly active learning for now, with the ultimate goal of pulling experiments into a reinforcement learning loop.

— Fabian Theis

In their own words · checked verbatim

they had, like in a single experiment, ten million transcriptomes measured, right? I thought that was like, like next level stuff.

Fabian Theis6:25

if we build in a priori inductive biases too strongly that we might lose predictive power, right?

Fabian Theis17:39

the honest learning from, from big scale machine learning is that at some point data beats human-made additional concepts in these models.

Fabian Theis18:45

a field that is rushing towards the gold pot of the virtual cell, people sometimes maybe were not as careful with the baselines

Fabian Theis26:11

Deep Learning is not needed for cell type prediction.

Fabian Theis27:14

I saw so many of these platform companies pivoting at some point to, "We make a drug."

Fabian Theis39:00

the, the ultimate vision being that we want to RL everything. We want to do RL also in, in, in, in this loop. It's a bit out there, but you know, I think we're, we're slowly getting there.

Fabian Theis49:45

Figures

Biohub cross-institutional virtual cell data initiative$1.8 billion2:13
Parse Bio single-experiment sequencing scale10 million transcriptomes6:25
scConcept multi-species latent space coverage16 species25:10
Helmholtz Munich Computational Health Center scale~45 group leaders, 500-600 people31:27

Glossary

world model
A model that predicts next state from current state and action; here, predicting cell states after perturbation.
bitter lesson
The hard truth that methods with less manual prior and more data and compute tend to win long-term.
organoid
Multi-cellular clusters cultivated in vitro from stem cells, with partial organ structure and function.
flow matching
A generative method that learns continuous transformation paths to generate samples.
Lab Compass
Fabian's team's inverse generative model that reverse-engineers experiments needed to reach a target cell state.

How to listen

Who it's for

Researchers in virtual cells and perturbation prediction, biocompute founders, and investors wanting to understand Germany's biotech startup ecosystem.

Skip

The first two minutes of the guest's background and pleasantries can be skipped without losing the rest of the content.

More on this