Deep Learning Models for High-Content Phenotypic Screening
Title: |
Deep Learning Models for High-Content Phenotypic Screening |
DNr: |
Berzelius-2025-227 |
Project Type: |
LiU Berzelius |
Principal Investigator: |
Giovanni Volpe <giovanni.volpe@gu.se> |
Affiliation: |
Göteborgs universitet |
Duration: |
2025-07-04 – 2026-02-01 |
Classification: |
10610 |
Keywords: |
|
Abstract
High-content screening (HCS) is a cornerstone of modern drug discovery and biological research, combining automated microscopy with image-based profiling to capture complex cellular responses to chemical and genetic perturbations. While recent advances in deep learning and computer vision have improved the extraction of biologically relevant features from HCS data, most existing approaches remain limited by their modality-specific design and task-specific tuning, which restricts their generalization and reusability across experimental contexts.
This project aims to develop a scalable multimodal deep learning framework that integrates cellular imaging with complementary data types, including chemical structure priors and transcriptomic readouts, into a unified representation space. The primary objective is to enable robust generalization across diverse experimental conditions while reducing sensitivity to batch effects—that is, technical variations introduced during data acquisition that can obscure biological signals. By learning a holistic representation that captures both morphological and molecular features, the model will support a wide range of downstream applications, including toxicity prediction, mechanism-of-action inference, and compound clustering for drug repurposing.
The resources requested through this application will support model training and validation on the JUMP-Target-2-Compound subset of the JUMP-Cell Painting dataset—a diverse and standardized benchmark designed to evaluate large-scale phenotypic profiling approaches. This provides a representative challenge space for assessing the model’s robustness, scalability, and ability to extract biologically meaningful signals across experimental batches and perturbation types. We envision this project as an initial proof of concept that will serve as a pathway for the future development of foundation models in phenotypic screening, with the potential to significantly accelerate predictive biology and data-driven therapeutic discovery.
This research initiative is a collaboration between the University of Gothenburg and AstraZeneca, with each partner contributing complementary expertise. The University of Gothenburg provides academic leadership and contributes advanced AI capabilities, including deep learning model development. AstraZeneca brings critical pharmaceutical domain knowledge and supports experimental validation. Together, this partnership is committed to delivering a transformative framework for interpreting high-content screening data.