Fabian Theis
Related people
Episodes
Notes
Latent Space: The AI Engineer Podcast · Controlling Biology - Fabian Theis, Helmholtz Munich
-
How Does a Model Choose the Next Biology Experiment?
Fabian Theis says a model chooses the next biology experiment by inverting a forward model. LabCompass, an inverse generative model from his group, takes a forward world model and says what to do in order to reach a chosen state.
Mapping cells came first. Predicting the next state came next. Steering the experiment is the step after that, and it puts the scientific method back in the loop. Papers often stop at a model, a concept figure, and a little validation, then a new project starts. Stopping there is not continual learning. Molecule design and material design already run the loop. Big pharma has named the theme, but he does not see many cases of several loops that run an experiment, up
-
Why Do Linear Baselines Still Matter for Virtual Cell Models?
Fabian Theis says linear baselines still matter because a field rushing toward a virtual cell sometimes skipped them. On the plain task of naming a cell type from a new transcriptome, some models close to the state of the art are logistic regression. CellTypist does a good job. A paper from his group about five or six years ago said, in a plain title, that deep learning is not needed for cell type prediction. The claim holds inside one organ. Across organs the task gets subtler, because cell types show up again in different organs and immune cells change with the tissue.
Early perturbation analyses used classical linear models. Around 2019 his lab introduced linear shifts inside a latent sp
-
Why Did Later AlphaFold Models Learn Rotation Symmetry From Data?
Fabian Theis says later AlphaFold models found rotation symmetry in the data instead of building that rule into training. The first models went to extraordinary effort to make the molecules rotation invariant inside the learning process. Later versions let the model find the symmetry. A physical constraint put in by hand is not always rigid, so the later models were a bit more flexible, and letting the model learn from the data gained something.
He sets the AlphaFold story next to a longer result. Building a priori inductive biases in too strongly might cost predictive power. In cell modeling the number of cells is large, but the number of perturbations, the things done to a cell, is not, s
-
What Is an Organoid?
Fabian Theis says an organoid is a little piece of an organ grown in a dish. Cells taken out of a body can grow into small multicellular systems that connect and do things. He is not a trained biologist, and the first time he saw cells do that he was impressed.
People can put cells in a dish and grow mini brains, a set of neurons that start connecting and talking to each other. Those mini brains have been shown to compute. Heart cells can grow into little heart organoids that start beating, and collaborators of his have beating heart-cell clumps in a dish. The colon is a simpler system that works very well, and cancer systems exist too. When tissue is hard to acquire, as with some neurodege
-
What Is a Virtual Cell?
Fabian Theis says a virtual cell, in the way people use the phrase, is a model that predicts how a cell would behave under all kinds of perturbations. He calls that reading a bit too simplistic for the whole idea, and he calls the deeper question maybe a bit of a holy grail in computational biology.
A cell is an input-output system with memory, and the molecules involved are not only RNA. Understanding it means predicting the output given different inputs. Perturbation-prediction models are what the phrase often points at, and those models have taken off. His lab started the work in 2019 with a nonlinear model that did a linear transformation in a latent space. Many pseudo-linear versions s
-
How Competitive Are Startups in Germany?
Fabian Theis leaves a small doubt on the exact ranking and still puts Munich, at least in continental Europe, as the prime place now for VC capital and for deep tech, with bio and defense really exploding. Startup creation and implementation in Germany is more than competitive versus Europe.
The institutes hold university affiliations without undergraduate teaching. They work with PhD students and postdocs and stay close to the universities and the clinics so the science can move into translation. The model is similar to NIH-type setups in the US, spread across the country rather than gathered in one place. Government in Germany is realizing the country needs a much stronger transfer compon
-
What Is Perturbation Space for a Virtual Cell?
Fabian Theis calls perturbation space the input space of everything that happens to a cell. A virtual cell model should take the whole space, including combinations, because the inputs influence cell state. He is unsure whether disease and time belong on that list as well.
A full gene knockout is only the hardcore end. Most influences are subtler, though taking pieces out still helps explain the system. In ordinary development and in homeostasis, cells live in complex multicellular states and talk to each other all the time. A neighbor, a signaling gradient, or hormones coming in all count, and much of that outside input cannot be tuned up or down. In dish-based drug screens, changing media
-
What Is a Transcriptome?
Fabian Theis calls a transcriptome the messenger RNAs each cell reads off a genome that the Human Genome Project mapped for the body and that stays pretty much the same everywhere. Different cell types exist because each cell reads different material off that genome. The messenger RNAs encode downstream proteins and are expressed differently from cell to cell. A human has something like 20,000 different types, and splicing produces many more.
Each cell is a 20,000-dimensional vector: zeros where a messenger RNA is absent, and counts of how many of the messenger RNAs are present. The messenger RNAs are called transcripts, and together they are the transcriptome. The 20,000-dimensional space