Hani Goodrazi – Arc institute has been working on virtual cell.
Drugs fail due to overfit experimental models, You need screen drugs with better models of human biology.
Geneva is platfrom that brings tumor models into perturbation model – that is a transcriptomics assay that deconvolves into effect. Take multiple cell line and then treat with 2 drugs and can distinguish the cell lines that is very sensitive to one drug. This can be done with Patient xenografts. In vitro and in vivo programs are different. GENEVA platform paper – Nature cancer paper.
Can drug screening be moved into these complex models? But it is still a model and not human. This will not take you all the way to a human drug.
G12C tumors when treated with Sotorasib (a covalent binding drug) – even with this drug it shows a lot of heterogeneity in response. We can collect data from the patient but cannot run a test on the patient.
Virtual cell – in silico tumor models. One concept many manifestations. It is a framework not a singluar model. The bar is xenograft model that is used for drug discovery.
A virtual cell need to generalize beyond training universe. Gene regulatory network is complex but not infinite so it is in bounded complexity.
Lessons from alpha-fold:
GENEVA extended. Cell Bxio ties cellural function to cellular state. In DNA there is no structure to functional relationship for example. A cell state is a lower projection of multiomic measurement – UMAS. In cells in various states – abstractions of gene regulatory modules. Evolution is doing pertubration management.
Perturbation proves causation.
Tahoe company from GENEVA created multiple datasets. They have created datasets with different cells and drugs. They can show Kras G12C as an important determinant and can predict mechanism of action. Arc virtual cell atlas – exists in Github.
Perturbation problem – start with a population of cells and then figure out how they affect the cell state. STATE directly models the perturbation process. – generative process. Population of cells are modeled and should be the same, and they are covariate adjusted and they move into a new state.
ST – is a generalization of common approaches – Cell is integrating information from the neighbor.
Evaluated perturbation prediction models. Pearson correlation from Tahoe 100M, Parse-PMC, Plogle-Nadig and also from Differentially expressed genes.
STATE has improved over existing baselines.
Arc institute also has a virtual cell challenge. It is data limited and not compute limited.
Arc also has DNA data set which is compute limited.
