Deep Learning Driven Phenotypic Profiling for High Content Screening of FDA Approved Drugs for Cardiac Protection
Faculty Mentor
Ramswamy Kannappan
Major/Area of Research
Pharmaceutical Sciences, Pharmacology
Description
INTRODUCTION: This research presents the development and validation of an automated deep learning framework to accelerate drug discovery through quantitative analysis of multimodal cellular micrographs. Conventional analysis methods are slow and fail to capture the subtle, nonlinear morphological signatures of cardiotoxicity. To address this, we developed a convolutional neural network pipeline optimized for CUDA-accelerated compute environments, capable of rapidly detecting patterns of cardiotoxicity in high- content microscopy datasets.
METHOD: We built a custom AI model from the ground up, running on our own system, using Meta's Llama and CUDA 13.1, optimized for our local setup. This model is designed to analyze complex laboratory datasets and adapt to our research needs, and was trained on H9C2 cardiac cells to recognize a high-dimensional "fingerprint" that differentiates untreated controls from Doxorubicin-induced cell death. After training to detect toxicity marked by changes in nuclear morphology, cytoskeletal structure, and oxidative stress, the AI was used to screen FDA-approved drugs for "phenotypic rescue," identifying compounds that restore damaged cells to a healthy phenotype.
RESULTS: Our results demonstrate that this AI-driven approach significantly outperforms established benchmarks in detecting nuanced cellular patterns, providing a robust, high-throughput platform for identifying potential cardioprotective agents.
DISCUSSION/CONCLUSION: This end-to-end analytical pipeline establishes a strategic roadmap for transitioning from qualitative observation to precision bioimage informatics in high-content screening.
Deep Learning Driven Phenotypic Profiling for High Content Screening of FDA Approved Drugs for Cardiac Protection
INTRODUCTION: This research presents the development and validation of an automated deep learning framework to accelerate drug discovery through quantitative analysis of multimodal cellular micrographs. Conventional analysis methods are slow and fail to capture the subtle, nonlinear morphological signatures of cardiotoxicity. To address this, we developed a convolutional neural network pipeline optimized for CUDA-accelerated compute environments, capable of rapidly detecting patterns of cardiotoxicity in high- content microscopy datasets.
METHOD: We built a custom AI model from the ground up, running on our own system, using Meta's Llama and CUDA 13.1, optimized for our local setup. This model is designed to analyze complex laboratory datasets and adapt to our research needs, and was trained on H9C2 cardiac cells to recognize a high-dimensional "fingerprint" that differentiates untreated controls from Doxorubicin-induced cell death. After training to detect toxicity marked by changes in nuclear morphology, cytoskeletal structure, and oxidative stress, the AI was used to screen FDA-approved drugs for "phenotypic rescue," identifying compounds that restore damaged cells to a healthy phenotype.
RESULTS: Our results demonstrate that this AI-driven approach significantly outperforms established benchmarks in detecting nuanced cellular patterns, providing a robust, high-throughput platform for identifying potential cardioprotective agents.
DISCUSSION/CONCLUSION: This end-to-end analytical pipeline establishes a strategic roadmap for transitioning from qualitative observation to precision bioimage informatics in high-content screening.