DeepSA: a deep-learning driven predictor of compound synthesis accessibility.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Chemical informatics

Background:

• Advancements in artificial intelligence (AI) have led to numerous computational models for novel molecule generation.
• Assessing the synthetic accessibility of generated compounds is crucial for practical applications.
• Predicting synthesis difficulty aids in selecting viable molecules for further development.

Purpose of the Study:

• To introduce DeepSA, a deep learning model for predicting compound synthetic accessibility.
• To provide a tool that assists researchers in choosing molecules that are easier to synthesize.
• To improve the efficiency of drug discovery and development pipelines.

Main Methods:

• Developed DeepSA, a chemical language model utilizing natural language processing (NLP) algorithms.
• Trained the model on a large dataset comprising 3,593,053 molecules.
• Evaluated DeepSA's performance using the area under the receiver operating characteristic curve (AUROC).

Main Results:

• DeepSA achieved an AUROC of 89.6% in distinguishing difficult-to-synthesize molecules.
• The model demonstrates superior performance compared to existing state-of-the-art methods.
• Analysis showed that SMILES representations alone can effectively capture informative molecular features, comparable to graph-based methods.

Conclusions:

• DeepSA offers a valuable computational tool for predicting synthetic accessibility.
• The model can help reduce the time and cost associated with drug discovery.
• DeepSA highlights the efficacy of NLP techniques applied to chemical language for molecular property prediction.