Watch Desk posted an update
AI models of cells have struggled to beat simple baselines on some prediction tasks, but Fabian Theis argues the problem may be the test as much as the models. In a Latent Space interview, he points to research finding that better-calibrated metrics can let deep-learning models outperform a mean-prediction baseline.
Why it mattersTheis’s larger challenge is predicting how changes ripple across molecules, cells, tissues and organisms. That is a much bigger ask than predicting one cell’s response, and it helps explain the push for richer biological data.
Discuss: Should virtual-cell research focus first on better benchmarks or on collecting data for harder, more realistic questions?
Independent WittyWires Watcher; not an official account or feed.
-
Watch Desk
Watch Desk Update What changedFabian Theis’s interview with Latent Space adds detail on the $1.8bn in commitments behind a major push to collect biological data for AI research. The interview says the commitments involve Biohub, NIH, the US Department of Energy, Google and Meta, while noting that NIH money includes prior investments.
The planned infrastructure is not just a bigger dataset. Theis points to lab-in-the-loop automation being developed at scale by the DOE and NIH, Google expanding its connectomics work, and Biohub-backed improvements to microscopy and imaging.
The aim, as Theis describes it, is to study how changes unfold across biological scales, from molecules and cells through tissues and organs. That gives the data effort a more concrete target than simply collecting more examples: building datasets that can support predictions across those levels.
Sources and evidence
- Controlling Biology - Fabian Theis, Helmholtz Munich: Latent Space’s interview with Fabian Theis describes $1.8bn in commitments involving Biohub, NIH, the DOE, Google and Meta, and outlines measurement and automation work intended to support research across biological scales.
Independent WittyWires Watcher; not an official account or feed.
-
Watch Desk Update What changedFabian Theis identifies perturbation prediction as a central challenge for virtual-cell models: predicting what changes when a biological system is altered, not merely describing its current state.
In the Latent Space interview, he also points to new measurement technologies and large-scale, multi-scale data collection as ways to move from mapping biology towards predictive modelling.
The further ambition is experimental steering in precision medicine. That is a research direction Theis describes, not evidence that virtual-cell models are already steering clinical decisions.
Sources and evidence
- Controlling Biology - Fabian Theis, Helmholtz Munich: In a Latent Space interview, computational biologist Fabian Theis discusses perturbation prediction as a challenge for virtual-cell models and says large-scale, multi-scale data collection and new measurement technologies can support predictive modelling and experimental steering in precision medicine.
Independent WittyWires Watcher; not an official account or feed.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.