Our main website is www.ewaldlab.org/ – please visit it for more detailed and up-to-date information.

Most chemicals in commercial use have never been assessed for their potential to cause harm to humans or ecosystems. Drug discovery programs routinely screen thousands to millions of compounds in computational and cellular models to understand how they perturb biological systems. Our group aims to re-purpose drug discovery technologies and data to investigate the potential toxicity of human-created compounds that are used in society and released into the environment. We develop computational approaches to:

• learn multi-omics signatures of compound-induced toxicity
• link perturbations at the molecular and cellular level to adverse outcomes at the organ and organism level.

Benchmarking multi-omics data for toxicity detection

With the OASIS Consortium (8 academic institutions, 2 non-profits, 7 government agencies, and 16 pharma/agchem companies), we are collecting multi-omic dose-response data in 5 cell lines for ~1500 compounds with pre-existing human or rat in vivo liver toxicity data. The compounds include pharmaceuticals, pesticides, and industrial compounds. The overall objective is to determine whether profiling data from cells can predict in vivo liver toxicity. Sub-objectives include understanding species-specific compound toxicity (human, rat), interpreting compound mode-of-action from omics data, and developing new methods for integrating image-based profiles with transcriptomics and proteomics data.

Extracting information from images of cells

Cell Painting is a high-content, image-based profiling assay that captures morphological and subcellular changes in cells using multiplexed fluorescent stains to measure phenotypic responses to perturbations (Figure 1). It captures detailed mode-of-action information while being at least 1000 times cheaper than other single-cell omics methods. Even though Cell Painting data has single-cell resolution, most analyses are done at the population level. We develop new methods for investigating single-cell heterogeneity in response to chemical perturbations. This includes using deep learning to learn representations of cells that reveal subtle cell states and developing custom image analysis workflows to identify and quantify phenotypes that are relevant to toxicity, for example lipid accumulation or multi-nucleation.

Inferring compound mode-of-action using reference atlases

There are large, public Cell Painting datasets that contain profiles for >100,000 compounds and ~20,000 genetic perturbations, which contain rich mode-of-action signals. Studies that profile environmental contaminants are often smaller, usually containing ~100-1000 compounds with unknown modes-of-action. We adapt computational methods for creating single-cell RNA-seq reference atlases for image-based profiles. The objective is to map Cell Painting data from environmental toxicity screens to large, well-annotated reference datasets so that we can infer the mode-of-action of uncharacterized compounds.