molecular assembly on deep learning based on retrosynthesis

We propose retro explainer,formulizing the retrosynthetic action guided by deep learning.to guarantee a robust performance of our model,we propose 3 units: a multi scale graph transformer,structure aware contrastive learning and dyanamic adaptive multi-task learning.As a results, retro explainer is expected to offer valuable insights for reliable,high throughput and high quality organic synthesis in drug development. 

Interpretability:We introduced an energy-based molecular assembly process that offers transparent decision-making and interpretable retrosynthesis predictions. This process can generate an energy decision curve that breaks down predictions into multiple stages and allows substructure-level attributions; the former can help understand the "counterfactual" predictions to discover potential biases in the dataset, and the latter can provide more granular references (such as the confidence of a certain chemical bond being broken) to inspire researchers to design customized reactants