Predicting drug polypharmacology from cell morphology readouts using variational autoencoder latent space arithmetic.

Area of Science:

• Computational biology
• Machine learning in drug discovery
• Bioimage analysis

Background:

• Variational autoencoders (VAEs) generate compressed, interpretable latent spaces from biomedical data.
• Standard VAEs often have entangled latent spaces, limiting their utility.
• Specialized VAEs like β-VAE and MMD-VAE offer improved disentanglement and interpretability.

Purpose of the Study:

• To evaluate the capacity of VAEs to learn cell morphology from images.
• To assess the generative capabilities of VAE variants (Vanilla, β-VAE, MMD-VAE) in predicting compound polypharmacology.
• To explore the utility of gene expression data as a complementary information source.

Main Methods:

• Training and evaluating Vanilla VAE, β-VAE, and MMD-VAE on cell morphology data.
• Utilizing latent space arithmetic (LSA) to explore generative capacity for predicting polypharmacology.
• Training VAEs on gene expression data from the same compound perturbations to assess generalizability.

Main Results:

• β-VAE and MMD-VAE demonstrated superior disentanglement of morphology signals, yielding more interpretable latent spaces.
• VAE models successfully simulated morphology and gene expression readouts for specific compounds.
• The models predicted cell states perturbed by compounds with known polypharmacology.

Conclusions:

• β-VAE and MMD-VAE effectively learn and represent cell morphology and gene expression data.
• Latent space arithmetic with VAEs can predict compound polypharmacology and infer cell states.
• This approach can aid in developing targeted therapeutics and understanding drug off-target effects.