Domain-invariant features for mechanism of action prediction in a multi-cell-line drug screen

Abstract

Motivation: High-content screening is an important tool in drug discovery and characterization. Often, high-content drug screens are performed on one single-cell line. Yet, a single-cell line cannot be thought of as a perfect disease model. Many diseases feature an important molecular heterogeneity. Consequently, a drug may be effective against one molecular subtype of a disease, but less so against another. To characterize drugs with respect to their effect not only on one cell line but on a panel of cell lines is therefore a promising strategy to streamline the drug discovery process.

Results: The contribution of this article is 2-fold. First, we investigate whether we can predict drug mechanism of action (MOA) at the molecular level without optimization of the MOA classes to the screen specificities. To this end, we benchmark a set of algorithms within a conventional pipeline, and evaluate their MOA prediction performance according to a statistically rigorous framework. Second, we extend this conventional pipeline to the simultaneous analysis of multiple cell lines, each manifesting potentially different morphological baselines. For this, we propose multi-task autoencoders, including a domain-adaptive model used to construct domain-invariant feature representations across cell lines. We apply these methods to a pilot screen of two triple negative breast cancer cell lines as models for two different molecular subtypes of the disease.

Availability and implementation: https://github.com/jcboyd/multi-cell-line or https://zenodo.org/record/2677923.

Supplementary information: Supplementary data are available at Bioinformatics online.

Fig. 1.

MOA prediction is performed on an image via a phenotypic profile. The development of such a profile spans four ordered stages. Each stage may be accomplished by a variety of algorithms, the combination of which define a unique pipeline. Some stages may be omitted in certain pipelines, or subsumed to a common framework

Fig. 2.

Multi-task autoencoders used for dimensionality reduction over multi-cell-line data. Clockwise from top left: vanilla autoencoder, multi-task autoencoder and DAA. Coloring indicates separate treatment of each domain (cell line)

Fig. 3.

t-SNE embeddings of encodings from autoencoder (left) and DAA (right), with cell lines distinguished by color, and mean silhouette scores of 0.11 and 0.01, respectively

Fig. 4.

MDS embedding of drug effect profiles for MDA231 and MDA468 cell lines with DMSO centroid centered on origin. Detection of differential drug effects between cell lines with examples for each category below (MDA231 top and MDA468 bottom). From left to right: no drug effect in either cell line (negative control); drug effect in MDA231 cell line only; drug effect in MDA468 cell line only; similar drug effects in both cell lines; differentiated drug effects in both cell lines. Shown are example images, blue, DAPI; red, microtubules; green, DSB