Merging Bioactivity Predictions from Cell Morphology and Chemical Fingerprint Models Using Similarity to Training Data

Abstract

The applicability domain of machine learning models trained on structural fingerprints for the prediction of biological endpoints is often limited by the diversity of chemical space of the training data. In this work, we developed similarity-based merger models which combined the output of individual models trained on cell morphology (based on Cell Painting) and chemical structure (based on chemical fingerprints). Using a combination of a decision tree and logistic regression models on the structural versus morphological feature space of the training data, which leveraged the similarity of test compounds to training compounds, the similarity-based merger models used logistic equations to weigh individual model outputs. We applied these models to predict assay hit calls of 92 assays from ChEMBL and PubChem and 89 anonymised assays released by the Broad Institute, where the required Cell Painting annotations were available. We found that for the 181 assays used in this study the similarity-based merger model improved AUC in relative terms by 16.3% compared to the models using chemical structure alone (mean AUC of 0.75 vs. 0.64), and by 21.3% compared to the models using Cell Painting data alone (mean AUC of 0.62). Our results demonstrate that similarity-based merger models combining structure and cell morphology models can more accurately predict a wide range of biological assay outcomes and expand the applicability domain by better extrapolating to new structural and morphology spaces.

A Decade in a Systematic Review: The Evolution and Impact of Cell Painting

“…Other studies using similar ML methodologies compared bioactivity predictions based on Cell Painting and chemical structure information. 80,81 Prediction accuracies for targets like β-catenin (usually assayed using a specific stain) were better when using a BMF Macau model with Cell Painting profiles as side information (F1 score = 0.87) compared to a model using only chemical structural data (F1 score = 0.48). 80 Another study showed that the models that combined structural information and Cell Painting profiles, using similarities to training data, improved the AUC by 16.3% compared to models that only used chemical structure information.…”

“…80 Another study showed that the models that combined structural information and Cell Painting profiles, using similarities to training data, improved the AUC by 16.3% compared to models that only used chemical structure information. 81…”

Using Generative Modeling to Endow with Potency Initially Inert Compounds with Good Bioavailability and Low Toxicity

“… 47 We started from previous work 14 that employs random forest models trained on a combination of Morgan fingerprints, 23 Cell Painting, 24 and gene ontology features to classify molecules as toxic or nontoxic. Although similar approaches for combining data have been shown to improve accuracy on a range of bioactivity predictions, 48 this approach needs to be established individually for each bioactivity endpoint studied. When using this strategy to find compounds targeting αS, we failed to identify compounds that were both efficacious and nontoxic.…”

Leveraging Cell Painting Images to Expand the Applicability Domain and Actively Improve Deep Learning Quantitative Structure–Activity Relationship Models

“…To leverage the advantages of both these approaches and overcome their respective limitations, we propose a computational pipeline in the spirit of active learning. Our approach builds upon previous work showing the positive impact of employing multimodal compound descriptors for bioactivity prediction. , In a first step, initial models are built based on structural information and images, respectively. Then predictions from the image-informed and chemistry-informed ligand-based models guide the selection of molecules to be tested in an actual physical assay.…”