Transformer-based multitask learning for reaction prediction under low-resource circumstances.

Area of Science:

• Organic Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Deep learning excels at chemical reaction prediction but suffers from low accuracy on limited datasets.
• Retrosynthesis and forward reaction prediction are crucial for organic chemistry and drug discovery.

Purpose of the Study:

• To develop novel deep learning models that enhance chemical reaction prediction accuracy for low-resource datasets.
• To address the limitations of existing models in handling insufficient chemical reaction data.

Main Methods:

• Introduction of two multitask transformer models: retro-forward reaction prediction transformer (RFRPT) and multiforward reaction prediction transformer (MFRPT).
• Integration of multitask learning with transformer architecture to improve prediction for both forward and retrosynthesis tasks.
• Evaluation of model performance on low-resource chemical datasets.

Main Results:

• Both RFRPT and MFRPT models significantly outperformed the baseline transformer model in reaction prediction accuracy.
• RFRPT achieved an average top-1 accuracy of 76.9%, and MFRPT achieved 79.8%, compared to the baseline's 69.9%.
• Multitask learning effectively mitigated the impact of limited training data by capturing essential chemical knowledge.

Conclusions:

• Multitask learning frameworks, such as RFRPT and MFRPT, are effective in improving the predictive performance of deep learning models for chemical reactions.
• These methods overcome the restriction of limited training data, advancing organic chemistry and drug discovery research.
• The proposed models demonstrate a viable approach for accurate reaction prediction even with scarce chemical data.