AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Controlling Biology - Fabian Theis, Helmholtz Munich

Collected Nov 6, 2026

Latent Space published an interview with Fabian Theis, a computational biologist at Helmholtz Munich whose lab has worked for years on algorithms for multi-scale, multi-modal and time-varying biological data. The conversation centers on why the Virtual Cell idea has lost momentum and what should replace it. Theis's position, as presented in the piece: predicting how a cell's state shifts after a perturbation is not a hard enough problem on the data sets the field currently has, so the difficulty should move up to biology across scales — from molecules to cells, niches, tissues, organs and organisms — and to how those levels influence one another.

The piece frames this against a broader funding shift. A Biohub announcement a few weeks earlier drew attention for the size of the commitments: $1.8B from Biohub, NIH, DOE, Google and Meta. The accompanying discussion cited new imaging and measurement capabilities, the generation of large data sets and making that data available for training AI. The excerpt notes that the NIH money reflects prior investments. It also asks why these organizations are pressing now, given the work has been underway for years, and points to infrastructure that is only now coming online: lab-in-the-loop automation being developed at scale by DOE and NIH, Google's expanding connectomics work, and imaging platforms Biohub has helped labs improve, including laser phase plate microscopy, light sheet microscopy and post-processing for ordinary microscopes that surfaces deeper phenotypic data.

That combination — automated experimentation plus higher-fidelity, multi-scale imaging — is the data-side answer to the Virtual Cell problem. The model side is where Theis sits. His lab and initiatives at Helmholtz include the Computational Health Center, which combines AI with molecular data, medical imaging, genomics and clinical data to build predictive and mechanistic models for precision medicine, and the Institute of Computational Biology, which develops statistical and dynamical models spanning single-cell time series up to population-scale omics.

The article is candid about why Virtual Cell work drew skepticism. The 2025 Arc Virtual Cell Challenge, it says, surfaced a set of papers finding that constant baselines — not even linear ones — often beat sophisticated models at perturbation prediction. The task itself is easy to state: change a cell's inputs through gene editing, drugs or environmental factors, and predict how its state responds. Falling short of a constant baseline is a striking result. The piece also cites a Nature Biotechnology paper arguing the picture is less bleak than it looks, because many evaluation metrics are poorly calibrated and can reward an uninformative mean prediction simply because most genes barely move; with better-calibrated metrics, deep learning models can outperform the mean baseline. This suggests part of the apparent failure is a measurement artifact: if a benchmark's scoring rewards predicting the average, models that hedge look strong and models that capture real structure look weak. Rebuilding evaluation is therefore as important as rebuilding architectures, and any team benchmarking perturbation models should check whether its metric can distinguish a useful prediction from a safe one.

Theis's alternative reframes the questions worth asking. Rather than a single perturbation on a single cell type, the interesting problems involve neighbors and populations: how surrounding cells change a given cell's state, what happens to a distribution of cells after a perturbation, and how to perturb a cell to drive it into a desired state. Those questions point toward lab-in-the-loop experimentation at scale, where models propose experiments and automated labs run them, and the results feed back into the next round. In practice, this shifts the bottleneck from algorithm design to data collection and experimental throughput — which is exactly where the Biohub effort and the DOE and NIH automation work are aimed.

The interview also covers the business of Virtual Cells, what the piece calls the Bitter Lesson in biology, training the next generation of computational biologists, and closing the lab-in-the-loop loop. The Bitter Lesson reference is to the long-running argument that general methods which scale with compute tend to win over hand-engineered domain knowledge; applied to biology, the reading is that larger, better data plus general models may beat carefully designed priors.

A likely trade-off is latency. Cell atlases and imaging campaigns produce data on the timescale of months to years, while model development moves faster, so progress on cross-scale prediction may be gated by how quickly high-precision, co-registered data arrives rather than by modeling ideas. The lack of understanding of what the piece calls scale interfaces — how findings at one level of biological organization carry to the next — is another open problem, and one that better data alone may not resolve.

Why it matters: Teams building in AI for biology face a changed target. If the field's most-cited benchmark turns out to be too easy or too generously scored, effort spent beating it may translate poorly to real predictive value. Developers should expect the next wave of Virtual Cell work to be judged on cross-scale, multi-modal prediction and on closed-loop experimental design, and should treat metric calibration as part of the modeling problem rather than an afterthought — an inference from the benchmark findings described, not a stated conclusion.

Read at Latent Space

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Fabian Theis has been one of the key players in the various iterations of the Virtual Cell concept and initiatives. It is not hyperbole at all to say that his lab has been pushing the frontier of algorithms to understand multi-scale, multi-modal (”mult...