Two-Dimensional Deep Learning Frameworks for Drug-Induced Cardiotoxicity Detection.

Area of Science:

• Biomedical Engineering
• Computational Biology
• Pharmacology

Background:

• Drug-induced cardiotoxicity poses significant clinical and economic challenges.
• Existing in vivo and in vitro models have limitations in accurately identifying cardiotoxic substances.
• Accurate cardiotoxicity assessment is crucial for patient safety and efficient drug development.

Purpose of the Study:

• To develop novel deep learning frameworks for improved cardiotoxicity assessment.
• To leverage induced pluripotent stem cell-derived cardiomyocytes (iPSC-CMs) for human-relevant data.
• To enhance the accuracy and reliability of drug safety evaluations.

Main Methods:

• Utilized induced pluripotent stem cell-derived cardiomyocytes (iPSC-CMs) and impedance measurements to record mechanical beating signals.
• Applied short-time Fourier transform (STFT) and synchro-squeezing transform (SST) to convert temporal signals into 2D representations.
• Developed and implemented STFT-CNN and SST-CNN deep learning frameworks for data analysis.

Main Results:

• SST-CNN achieved 98.55% accuracy in drug type classification and 99% in cardiotoxicity classification.
• The frameworks demonstrated 98.5% average accuracy in classifying various drug concentrations.
• Outperformed traditional models like recurrent neural network (RNN) and 1D-CNN in accuracy and reliability.

Conclusions:

• The proposed STFT-CNN and SST-CNN frameworks offer a paradigm shift in cardiotoxicity assessment.
• These deep learning models provide a more robust, efficient, and human-relevant approach to drug safety.
• The scalable nature of these frameworks promises to revolutionize drug safety evaluations and reduce patient harm.