Image Based Liver Toxicity Prediction.

Area of Science:

• Computational chemistry
• Toxicology
• Machine learning

Background:

• Drug-induced liver injury is a significant concern in pharmaceutical development.
• In silico methods, particularly deep learning, offer promising solutions for early toxicity prediction.
• Standard deep learning models often struggle with small toxicological datasets due to overfitting.

Purpose of the Study:

• To develop and validate an image-based deep learning model for predicting drug-induced liver toxicity.
• To address the challenge of small datasets in toxicological predictions using novel deep learning techniques.
• To improve the efficiency of identifying potentially harmful drug candidates during early drug discovery.

Main Methods:

• Utilized convolutional neural networks (CNNs) on 3D molecular conformations to capture geometric and chemical features.
• Employed the COVER method for data augmentation and class balancing (toxic vs. non-toxic).
• Validated the approach on the p53 endpoint from the Tox21 dataset and subsequently on a large liver toxicity dataset.

Main Results:

• Achieved competitive results on the p53 endpoint, comparable to winners of the Tox21 data challenge.
• Demonstrated applicability to liver toxicity prediction with a sensitivity of 0.79 and specificity of 0.52 on a large dataset.
• Showcased the effectiveness of image-based toxicity prediction using deep neural networks.

Conclusions:

• Image-based deep learning, combined with data augmentation, is a viable strategy for predicting drug-induced liver toxicity.
• This approach can aid in the early identification and elimination of problematic drug candidates.
• The methodology holds potential for broader applications in toxicological assessments within drug discovery.