Identifying Chirality in Line Drawings of Molecules Using Imbalanced Dataset Sampler for a Multilabel Classification...

Area of Science:

• Computational Chemistry
• Machine Learning
• Cheminformatics

Background:

• Chirality is crucial in chemistry and biology, but its automated recognition from molecular structures is challenging.
• Existing machine-readable molecular representations may not fully capture chirality information.
• Deep learning advances offer potential for automatic chemical structure recognition, yet chirality discovery remains understudied.

Purpose of the Study:

• To develop and evaluate deep learning models for classifying molecule chirality (achiral/chiral) and chirality types (none/centre/axial/planar).
• To address data imbalance issues in multilabel chirality classification using a novel sampling technique.
• To investigate the interpretability of deep learning models in predicting chirality based on structural elements.

Main Methods:

• Pretraining deep neural networks on the ChEMBL+ dataset (79,641 molecules).
• Fine-tuning models for binary and multilabel chirality classification tasks.
• Implementing a Formulated Imbalanced Dataset Sampler (FIDS) to handle imbalanced label combinations in multilabel classification.
• Conducting 10-fold cross-validation on a manually curated CHIRAL dataset (1,142 molecules).

Main Results:

• Achieved up to 90% accuracy in binary chirality classification.
• The multilabel classification task, enhanced by FIDS, showed an overall performance increase from 87% to 89%.
• Accuracy per label combination in the multilabel task improved by up to 50% with FIDS.
• Heatmap analysis demonstrated the models' ability to predict chirality based on specific molecular locations.

Conclusions:

• Deep learning models can effectively classify molecular chirality and chirality types.
• The proposed FIDS method significantly improves performance in imbalanced multilabel chirality classification.
• The study highlights the potential of deep learning for interpretable chirality prediction, linking predictions to specific structural features.