DirectMultiStep: Direct Route Generation for Multistep Retrosynthesis.

Area of Science:

• Computational Chemistry
• Artificial Intelligence in Drug Discovery
• Organic Synthesis Planning

Background:

• Traditional computer-aided synthesis planning (CASP) methods face scalability challenges due to iterative single-step predictions and exponential search space growth.
• There is a need for more efficient and accurate computational tools to automate retrosynthetic analysis.

Purpose of the Study:

• To introduce novel transformer-based models for direct, multistep retrosynthetic route generation.
• To evaluate the performance of these models against state-of-the-art methods and assess their generalization capabilities.

Main Methods:

• Development of transformer-based models employing a mixture of experts approach.
• Direct generation of multistep synthetic routes as a single output string, with conditional predictions based on preceding steps.
• Utilized the PaRoutes dataset for model training and evaluation, including models with varying input constraints (e.g., number of steps, starting materials).

Main Results:

• The DMS Explorer XL model demonstrated significant improvements in Top-1 accuracy (1.9x and 3.1x on n1 and n5 test sets, respectively) compared to existing methods.
• Incorporating constraints like desired steps and starting materials reduced model size and enhanced accuracy, with DMS-Flex (Duo) achieving 25-50% higher Top-1 and Top-10 accuracies.
• Models showed strong generalization by successfully predicting synthesis routes for FDA-approved drugs not present in the training data.

Conclusions:

• The multistep-first approach using transformer models offers a promising direction for fully automated retrosynthetic planning.
• Model performance can be further optimized by incorporating specific constraints, balancing model complexity and predictive accuracy.
• While effective, the models' performance on less common reaction types may be limited by the diversity of the training dataset.