Descriptor Free QSAR Modeling Using Deep Learning With Long Short-Term Memory Neural Networks.

Area of Science:

• Computational Chemistry
• Machine Learning
• Drug Discovery

Background:

• Quantitative Structure-Activity Relationship (QSAR) modeling traditionally relies on pre-computed molecular descriptors.
• Descriptor calculation, selection, and model fitting are standard steps in current QSAR practices.
• There is a need for QSAR models that can handle large, diverse datasets without extensive feature engineering.

Purpose of the Study:

• To explore the feasibility of building high-quality, interpretable QSAR models using Long Short-Term Memory (LSTM) neural networks without pre-calculated descriptors.
• To evaluate the performance of LSTM-based QSAR models on large and diverse chemical datasets.
• To enhance the interpretability of QSAR predictions, particularly for mutagenicity data.

Main Methods:

• Utilized different forms of Long Short-Term Memory (LSTM) neural networks trained directly on SMILES codes or a novel linear molecular notation.
• Modeled three biological endpoints: Ames mutagenicity, *P. falciparum* Dd2 inhibition, and Hepatitis C Virus inhibition.
• Employed an attention-based mechanism with bidirectional LSTM for interpretability and structural alert detection in mutagenicity prediction.
• Compared LSTM models against traditional fragment descriptor-based models.

Main Results:

• LSTM models achieved prediction accuracies comparable to traditional fragment descriptor-based models across three diverse endpoints.
• LSTM models exhibited superior performance in predicting the activity of test chemicals dissimilar to those in the training set.
• Attention-based LSTMs successfully identified structural alerts for mutagenicity, enhancing model interpretability.
• The study demonstrated that descriptor-less QSAR models built with LSTMs are interpretable and not "black boxes".

Conclusions:

• It is feasible to construct robust and interpretable QSAR models using LSTMs without relying on pre-computed traditional descriptors.
• LSTM-based QSAR models offer advantages in handling large datasets and generalizing to novel chemical structures.
• The developed approach holds promise for the mainstream adoption of large-scale, descriptor-less QSAR modeling in cheminformatics and drug discovery.