Extended Data Fig. 2: Model architecture.
From: In silico biological discovery with large perturbation models
a. Graphical model shows the dependencies between random variables previously described in Section 4.1. Dashed lines indicate implicit bi-directional dependencies that enable transfer learning across datasets. Symbolic perturbation, readout, and context descriptors (P,R,C) are first embedded (ZP, ZR, ZC), then used to generate output Y that represents the value of the readout R. b. Embeddings are implemented as learnable look-up tables. P, R, and C identify indices in the corresponding tables. c. Concatenated embeddings are forward propagated through a multilayer perceptron to predict the output Y
