Predicting compound activity from phenotypic profiles and chemical structures

Abstract

Recent advances in deep learning enable using chemical structures and phenotypic profiles to accurately predict assay results for compounds virtually, reducing the time and cost of screens in the drug discovery process. The relative strength of high-throughput data sources - chemical structures, images (Cell Painting), and gene expression profiles (L1000) - has been unknown. Here we compare their ability to predict the activity of compounds structurally different from those used in training, using a sparse dataset of 16,979 chemicals tested in 376 assays for a total of 542,648 readouts. Deep learning-based feature extraction from chemical structures provided a remarkable ability to predict assay activity for structures dissimilar to those used for training. Image-based profiling performed even better, but requires wet lab experimentation. It outperformed gene expression profiling, and at lower cost. Furthermore, the three profiling modalities are complementary, and together can predict a wide range of diverse bioactivity, including cell-based and biochemical assays. Our study shows that, for many assays, predicting compound activity from phenotypic profiles and chemical structures is an accurate and efficient way to identify potential treatments in the early stages of the drug discovery process.

Machine learning for perturbational single-cell omics

“…To tailor a perturbation atlas to perturbation modeling, it should take into account perturbations and conditions that capture possible cellular states relevant to our objectives: for example, varying response across cell types due to covariates, different targets and mechanisms of perturbations, perturbation interactions (Yofe et al, 2020), and biologically active chemical or sequence features (Becker et al, 2020). Of particular relevance for machine learning and DL, an atlas should capture enough experimental variation so that an ML model can generalize to many unseen situations, in addition to providing a collection of controlled experiments with many shared covariates.…”

Morphology and gene expression profiling provide complementary information for mapping cell state

“…Scientists have used individual profiling modalities to advance a variety of drug discovery applications, including improving screening library diversity, predicting cytotoxicity, prioritizing compounds for follow-up study, and inferring the mechanism of action of chemicals [21][22][23][24][25][26][27][28] . Integrating gene expression and morphology profiles with chemical structures revealed that each data type provides complementary information for predicting a drug's mechanism of action 29,30 , for predicting the effects of perturbations 31 , and for identifying nuisance compounds that can lead to false hits 32 . As well, to some degree, gene expression and morphology datasets contain sufficient information to predict changes in each other 29,30 .…”

High-Dimensional Gene Expression and Morphology Profiles of Cells across 28,000 Genetic and Chemical Perturbations

“…Data modality fusion and integration techniques are an active area of research in machine learning 4 and could potentially yield a superior representation of samples for many different biological profiling tasks on datasets where multiple profiling modalities are available. For example, predicting assay activity might be more successful using information about the impact of that compounds on cells' mRNA levels and morphology, rather than either data source alone 1 . Likewise, predicting the function of a gene based on similarities to other genes' profiles might be more successful using both data types.…”

“…Such a dataset would enable multi-modal (also known as multi-omic) analyses and applications. Examples include integrating the two data sources to better predict a compound's activity in an assay 1 , predicting the mechanism of action of a drug based on its profile similarity to well-understood drugs 2 , or predicting a gene's function based on its profile similarity to well-understood genes 3 .…”