In silico biological discovery with large perturbation models.

Area of Science:

• Computational Biology
• Genomics
• Systems Biology

Background:

• Perturbation experiments yield crucial biological insights but are challenging to integrate due to diverse data types and contexts.
• Existing methods struggle to consolidate information from heterogeneous perturbation datasets.

Purpose of the Study:

• To develop a novel deep-learning framework for integrating multiple, heterogeneous perturbation experiments.
• To enhance biological discovery by enabling in silico analysis of complex biological relationships.

Main Methods:

• Introduction of the large perturbation model (LPM), a deep-learning approach.
• Representing perturbation, readout, and biological context as disentangled dimensions within the model.
• Training LPM on diverse, pooled perturbation datasets.

Main Results:

• LPM demonstrates superior performance over existing methods in multiple biological discovery tasks.
• Accurate prediction of post-perturbation transcriptomes for unseen experiments.
• Identification of shared molecular mechanisms between chemical and genetic perturbations.
• Facilitation of gene-gene interaction network inference.

Conclusions:

• LPM effectively integrates heterogeneous perturbation data, learning joint representations of perturbations, readouts, and contexts.
• The model accelerates biological insight derivation from pooled experiments.
• LPM facilitates in silico studies of biological relationships, aiding therapeutic development.