Deep neural nets as a method for quantitative structure-activity relationships.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Quantitative structure-activity relationships (QSAR)

Background:

• Neural networks (NNs) were initially popular for QSAR in the 1990s but were later replaced by methods like support vector machines (SVM) and random forests (RF) due to practical limitations.
• Recent advancements in preventing overfitting, improved training algorithms, and hardware have led to a resurgence of NNs, particularly deep neural networks (DNNs).

Purpose of the Study:

• To evaluate the efficacy of deep neural networks (DNNs) against random forest (RF) models for prospective quantitative structure-activity relationship (QSAR) predictions.
• To determine if a standardized set of DNN parameters can achieve high performance across diverse QSAR datasets.

Main Methods:

• Utilized large, diverse QSAR datasets from Merck's drug discovery program.
• Compared the predictive performance of deep neural networks (DNNs) against random forest (RF) models.
• Assessed the generalizability of a single set of recommended DNN parameters across multiple datasets.

Main Results:

• Deep neural networks (DNNs) consistently achieved better prospective predictions than random forest (RF) models on the studied QSAR datasets.
• A single set of recommended DNN parameters demonstrated superior performance across most datasets, negating the need for dataset-specific optimization.
• The effectiveness of the chosen DNN parameters was validated on independent datasets.

Conclusions:

• Deep neural networks (DNNs) represent a powerful and effective tool for quantitative structure-activity relationship (QSAR) modeling, surpassing traditional methods like RF.
• Standardized DNN parameter sets can be successfully applied to diverse QSAR problems, simplifying model implementation in drug discovery.
• Despite computational intensity, the use of GPUs makes DNN training feasible for large-scale QSAR analysis.