Latent Space

Mini Brains, Beating Heart Organoids, and the AI That Predicts Them — Fabian Theis, Helmholtz Munich

Published 2026-11-06 · Duration 50:51

Summary

The field of computational biology is rapidly advancing from merely mapping cellular states (cell atlases) to predicting how cells will respond to external changes (perturbations). The ultimate goal is to achieve a 'lab-in-the-loop' system, where computational models actively guide and steer the next physical experiment. Key advancements include the development of 'world models' and inverse generative models, such as LabCompass, which allow researchers to determine the necessary input (perturbation) required to reach a desired cellular state. This shift is driven by massive data generation efforts (e.g., Biohub's $1.8B initiative) and the need for cell-type context-specific drug screening.

Download summary

Key takeaways

  1. Shift from Mapping to Control 2:15

    The progression in systems biology is moving from simply mapping cell states (mapping) to predicting outcomes (prediction), and finally to actively steering the next experiment (control). This is achieved through 'world models' and inverse generative models like LabCompass, which determine the required action to reach a target state. (0:00, 1:35, 3:00)

  2. The Importance of Perturbation Space 3:40

    The current frontier involves modeling the 'perturbation space'—the set of all possible inputs (e.g., CRISPR knockouts, drug cocktails, environmental changes) that can influence a cell's state. This space is exploding in data richness, moving beyond simple control vs. disease comparisons. (2:20, 3:30)

  3. Foundation Models and Scaling Laws 17:10

    While early models were often linear, the field is moving toward complex, flexible models. The success of modern AI models is increasingly dependent on demonstrating clear scaling laws across massive datasets (e.g., 350 million cells), suggesting that data scale is key to overcoming current limitations. (10:30, 11:40)

  4. The Lab-in-the-Loop Vision 1:00

    The ultimate vision is a closed-loop system where computational predictions directly inform and guide wet-lab experiments, enabling continual learning and accelerating the scientific method. This requires overcoming organizational and technical hurdles. (1:00, 2:00)

More digests on Computational Biology

Technical details

  • Virtual Cell Modeling 130s

    A virtual cell model aims to predict how a cell will behave under various perturbations (inputs), treating the cell as an input-output system with memory. This is considered a 'holy grail' in computational biology. (2:20)

  • Generative Models for Prediction 2351s

    Advanced techniques include Nicheformer (a foundation model for spatial omics) and scConcept/CellFlow, which are used to model complex relationships, such as predicting a cell's state based on its local environment (niche). (23:51, 25:45)

  • Inverse Generative Models (LabCompass) 60s

    LabCompass is an inverse generative model that takes a forward world model and determines the specific perturbation needed to drive the system to a desired target state. (1:00)

  • Data Representation and Omics 1300s

    Data is often represented as high-dimensional vectors (e.g., 20,000-dimensional transcriptome vectors) which are then encoded into a lower-dimensional latent space. The field is expanding to include spatial omics and multi-species comparisons. (13:00, 11:00)

Mentioned resources

  • Biohub (Initiative)
  • Scanpy / scverse (Open-Source Tools)
  • Nicheformer (Foundation Model)
  • scConcept / CellFlow (Computational Model)
  • LabCompass (Inverse Generative Model)

Channel & topics

Watch on YouTube · Back to latest

This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.