Toxicity prediction and classification of Gunqile-7 with small sample based on transfer learning method

Area of Science:

• Pharmacology and Computational Toxicology
• Traditional Mongolian Medicine

Background:

• Drug-induced diseases are a significant aspect of iatrogenic illness.
• Assessing the toxicity of traditional medicines like Gunqile-7 is vital for patient safety.
• Traditional animal testing for drug toxicity is costly and yields small datasets.

Purpose of the Study:

• To develop an efficient computational method for predicting the toxicity of Gunqile-7.
• To overcome the limitations of small sample sizes in pharmacological trials using advanced machine learning.
• To enhance the safety assessment of traditional Mongolian medicine.

Main Methods:

• Employed data augmentation to expand the limited dataset for Gunqile-7 toxicity.
• Utilized transfer learning with a one-dimensional convolutional neural network for model training.
• Applied Support Vector Machine-Recursive Feature Elimination for effective feature selection.

Main Results:

• The proposed method demonstrated improved accuracy in predicting Gunqile-7 toxicity.
• Achieved up to a 9 percentage point increase in accuracy compared to models without transfer learning.
• Successfully reduced the number of required training samples through data augmentation.

Conclusions:

• The combination of data augmentation and transfer learning is effective for toxicity prediction with small datasets.
• This approach offers a cost-efficient and accurate alternative to traditional animal testing for drug safety.
• The study validates the utility of computational methods in evaluating traditional medicines.