Retro3D: A 3D-aware Template-free Method for Enhancing Retrosynthesis via Molecular Conformer Information

TL;DR

This paper introduces Retro3D, a template-free retrosynthesis transformer that incorporates 3D conformer information via Atom-align Fusion and Distance-weighted Attention to align 1D SMILES with spatial coordinates and focus attention on chemically relevant atom pairs. By constructing a 3D distance weight from coordinates and guiding cross-attention with a SMILES Alignment Map, Retro3D achieves state-of-the-art performance among template-free methods and competitive results versus template-based and semi-template approaches on USPTO-50K and USPTO-FULL. The approach demonstrates robustness to conformer generation quality and improves validity and round-trip accuracy, enabling accurate predictions for molecules with intricate stereochemistry and polycyclic structures. This work has practical implications for accelerating realistic synthetic route design in drug discovery and organic synthesis by providing a scalable, geometry-aware retrosynthesis predictor without reliance on predefined templates.

Abstract

Retrosynthesis plays a crucial role in the fields of organic synthesis and drug development, where the goal is to identify suitable reactants that can yield a target product molecule. Although existing methods have achieved notable success, they typically overlook the 3D conformational details and internal spatial organization of molecules. This oversight makes it challenging to predict reactants that conform to genuine chemical principles, particularly when dealing with complex molecular structures, such as polycyclic and heteroaromatic compounds. In response to this challenge, we introduce a novel transformer-based, template-free approach that incorporates 3D conformer data and spatial information. Our approach includes an Atom-align Fusion module that integrates 3D positional data at the input stage, ensuring correct alignment between atom tokens and their respective 3D coordinates. Additionally, we propose a Distance-weighted Attention mechanism that refines the self-attention process, constricting the model s focus to relevant atom pairs in 3D space. Extensive experiments on the USPTO-50K dataset demonstrate that our model outperforms previous template-free methods, setting a new benchmark for the field. A case study further highlights our method s ability to predict reasonable and accurate reactants.

Figures (7)

• Figure 1: Architecture overview of Retro3D based on Transformer. a) SMILES and molecular conformer are processed through Word Embedding and ComENet to yield SMILES sequence embedding and 3D position embedding. Following the Atom-align integration, fusion embedding is then input into the Encoder. b) In Encoder, 3D Distance weight is obtained for redistribution of self-attention by Gaussian Basis Kernel function, while also refined layer-by-layer. c) In Decoder, SMILES Alignment is employed as guidance to align the cross-attention between the product and reactants.
• Figure 2: Case study for intricate structures. Each subfigure shows predictions of different methods: Retro3D (Top), Transformer (Bottom).
• Figure 3: Visualization on self-attention map of multi heads independently.
• Figure 4: Visualization on cross-attention map on SMILES Alignment.
• Figure 5: More Case