CellFusion: A Unified Platform Combining Various Deep Learning Algorithms and CellProfiler for increased reproducibility and efficiency in Cell Painting assay analysis.
Tuesday, February 6, 2024
2:00 PM – 2:15 PM EST
Room/Location: Ignite Theater Sponsored by Drug Discovery World (Booth 395) in Exhibition
BIOLIFY.AI & Poznan University of Medical Sciences, Wielkopolskie, Poland
Background: Cell Painting (CP) assays are becoming pivotal in early-stage phenotypic assays for identifying bioactives and aiding drug discovery. The core challenge lies in analyzing multi-channel image data to discern activity and correlate mechanisms of action.
Problem: Our research has identified significant variability in the results from different analytical methods used for the analysis of CP assays, including traditional computer vision techniques like CellProfiler and various deep learning approaches. For instance, the similarity in compound clusters, defined as the intersection over union between CellProfiler and deep learning methods, often falls below 25%. This inconsistency, particularly in mentioned compound cluster formations and HIT predictions, raises serious concerns about the reliability and reproducibility of current bioactive compound identification methodologies. We conclude that different analytical methods can provide diverse insights into compound classification and mechanisms of action. This highlights not only the importance of a multi-modal approach that combines image and structural data analyses but also the necessity of a multi-method approach to fully capture the broad spectrum of compound effects from a given modality.
Proposed solution: Recognizing these findings, we developed CellFusion, a platform that showcases the effectiveness and significantly increased reproducibility of multi-method ensembles for HIT identification. CellFusion automates the training of multi-modal deep learning algorithms, both supervised and unsupervised, and facilitates feature inference for new data using existing models and CellProfiler. It is designed for fast integration thanks to an intuitive cloud-based API, easily scaling to handle hundreds of plates for both encoding and training. Additionally, the platform integrates optimization, batch effect correction, and visual exploration of obtained features.
Validation: Validation of the CellFusion platform involved two distinct datasets: a proprietary dataset provided by IBCH and a subset of the open-source JUMP dataset, both involving different cell lines and compounds. Using the data the platform was tested on various tasks, including mechanism of action (MOA) prediction, where our system improved accuracy by a significant margin. The comparative analysis highlights the platform's adaptability and robustness in handling assays from diverse sources.
This project was performed in collaboration with the Centre for Chemical Biology IBCH PAS, financially supported by POIR.04.02.00-00-C004/19-00 project and the Polish Ministry of Education and Science (previously MNiSW, decision no DIR/WK/2018/06) for POL-OPENSCREEN project and the involvement in the joint international project EU-OPENSCREEN ERIC.