Denoising Drug Discovery Data for Improved Absorption, Distribution, Metabolism, Excretion, and Toxicity Property...

Area of Science:

• Computational chemistry
• cheminformatics
• Drug discovery

Background:

• Accurate prediction of absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties is crucial for efficient drug discovery.
• Experimental data often contains inherent errors, posing a significant challenge for developing reliable predictive models.
• Existing denoising techniques are primarily designed for classification tasks and are not directly applicable to regression-based ADMET prediction.

Purpose of the Study:

• To develop novel deep learning-based denoising schemes tailored for regression tasks in ADMET property prediction.
• To identify effective metrics for detecting noise in experimental assay data used for model training.
• To enhance the performance and reliability of computational models used in drug discovery.

Main Methods:

• Development of deep learning denoising algorithms specifically for regression problems.
• Utilizing training error (TE) as a primary metric for identifying noisy data points.
• Comparative analysis of TE against ensemble-based and forgotten event-based metrics for noise detection.
• Fine-tuning existing ADMET models using data preprocessed by the developed denoising scheme.

Main Results:

• Training error (TE) was identified as a reliable metric for detecting noise in regression tasks, outperforming other tested metrics.
• Fine-tuning models with TE-denoised data led to the most substantial performance improvements.
• The proposed denoising method effectively improved models with moderate levels of noise.
• The method demonstrated robustness, not degrading performance in low or high noise scenarios.

Conclusions:

• The developed deep learning denoising scheme is the first to successfully improve model performance for ADMET data.
• Training error (TE) is a valuable metric for noise identification in regression-based cheminformatics models.
• This approach has broad implications for enhancing the accuracy and utility of predictive models for various types of experimental assay data.