Computational Prediction of Single-Cell Perturbations
Predicting the effects of drugs regarding dosage, timing, possible drug combinations and other types of intervention such as gene knockouts can be experimentally difficult and very time-consuming. Nonetheless, this is one of the most important tasks in drug development and pharmaceutical research. A team of researchers including Dr Mohammad Lotfollahi and Carlo de Donno (co-first authors) and Prof Fabian Theis from Helmholtz Munich in collaboration with Meta AI developed the first open-source computational model based on generative AI to predict, interpret, and prioritize perturbations in cells. This can speed up the testing process and can serve as a guide for experimental validation.
Finding effective combinations of drugs to treat cancer is one of the most challenging problems faced in pharmaceutical research nowadays. The number of possible combinations becomes near-infinite as different drugs and dosages are considered and cannot be feasibly explored experimentally. Researchers at Helmholtz Munich in collaboration with Meta AI have developed a novel deep learning model that can predict the effect of combinatorial perturbations across different dosages or time points on the single-cell transcriptome. This model, called Compositional Perturbation Autoencoder (CPA), makes use of self-supervised learning and generative methods and decomposes gene expression into perturbation and covariate components that can then be combined to generate out-of-distribution predictions.
CPA successfully predicts the effects on mRNA transcript level of perturbations under different biological scenarios in single-cells and is the first model capable of generating predictions for combinatorial perturbations, paving the way for in silico hypothesis generation, that can lead to the identification of effective drugs and the development of new treatments. CPA is available as open-source software and is tailored for easy use and uptake from the single-cell genomics community. The model has the potential to accelerate dramatically the exploratory pipeline for identifying combinations of drugs with desirable outcomes, thus leading to targeted experiments and a reduction of time and monetary costs.
About the scientists
Prof Dr Dr Fabian Theis, Head of the Computational Health Center at Helmholtz Munich; Professor of Mathematical Modelling of Biological Systems at the Technical University of Munich (TUM), TUM School of Life Sciences Weihenstephan, TUM School of Computation, Information and Technology; Associate Faculty at the Wellcome Sanger Institute.
Mohammad Lotfollahi, team leader at Fabian Theis’ lab at Helmholtz Munich
Carlo de Donno, doctoral student at TUM School of Life Sciences at the Technical University of Munich
Original publication
Lotfollahi et al. (2023): Predicting cellular responses to complex perturbations in high-throughput screens. Molecular Systems Biology. DOI: https://doi.org/10.15252/msb.202211517
More Information
Link to the open-source software Compositional Perturbation Autoencoder (CPA): https://github.com/facebookresearch/CPA
Funding information
F.J.T. acknowledges support by the BMBF (grant number L031L0214A, grant number 01IS18036A and grant number 01IS18053A), by the Helmholtz Association (Incubator grant sparse2big, grant number ZT-I-0007) and by the Chan Zuckerberg Initiative DAF (advised fund of Silicon Valley Community Foundation, 2018-182835 and 2019-207271). This work was further supported by Helmholtz Association's Initiative and Networking Fund through Helmholtz AI [grant number ZT-I-PF-5-01]. M.L acknowledges financial support from Joachim Herz Foundation.
