Unbiasing Retrosynthesis Language Models with Disconnection Prompts

Abstract

Data-driven approaches to retrosynthesis are limited in user interaction, diversity of their predictions, and recommendation of unintuitive disconnection strategies. Herein, we extend the notions of prompt-based inference in natural language processing to the task of chemical language modeling. We show that by using a prompt describing the disconnection site in a molecule we can steer the model to propose a broader set of precursors, thereby overcoming training data biases in retrosynthetic recommendations and achieving a 39% performance improvement over the baseline. For the first time, the use of a disconnection prompt empowers chemists by giving them greater control over the disconnection predictions, which results in more diverse and creative recommendations. In addition, in place of a human-in-the-loop strategy, we propose a two-stage schema consisting of automatic identification of disconnection sites, followed by prediction of reactant sets, thereby achieving a considerable improvement in class diversity compared with the baseline. The approach is effective in mitigating prediction biases derived from training data. This provides a wider variety of usable building blocks and improves the end user's digital experience. We demonstrate its application to different chemistry domains, from traditional to enzymatic reactions, in which substrate specificity is critical.

Multistep retrosynthesis combining a disconnection aware triple transformer loop with a route penalty score guided tree search

“…In the present work, we removed the tagging information, and reactions were remapped and retagged using our new SMILES tagging strategy and syntax. The same dataset split for training, validation, and test (90 : 5 : 5), as shared by Thakkar et al 30 was used across all models resulting in 1 139 608, 63 672 and 63 454 reactions respectively.…”

“…A given template can contain multiple disconnected sets of reactive atoms. Finally, the transformer model AutoTag reported by Thakkar et al 30 was trained from untagged SMILES to the corresponding tagged molecule to provide additional tagging examples.…”

Exploring Chemical Reaction Space with Machine Learning Models: Representation and Feature Perspective

“…They retrieve synthons from a predefined library and then employ a transformer decoder to complete the full molecule reactants. Similarly drawing inspiration from graph models, Thakkar et al initially predict disconnection bonds and incorporate additional features. Moreover, prior research consistently finds that providing atom mapping is beneficial for chemical reaction modeling .…”

“…However, in real cases, multiple sets of precursors could be transformed into target structures. By using a prompt-based method, the model incorporates human intervention to limit the prediction to a constrained subspace.…”

Do Chemformers Dream of Organic Matter? Evaluating a Transformer Model for Multistep Retrosynthesis

“…24 Since the first implementations of a retrosynthesis transformer model, 11,24 several enhancements and developments have been presented, including a diversity-enhanced transformer, 18 triple-transformer validation loop, 17 and disconnection-prompted transformer. 20 Although earlier studies show great promise for retrosynthesis transformers, there are still questions to address before introducing a transformer model like Chemformer into the AiZynthFinder production platform. 31 First, most studies consider models trained on the United States Patent and Trademark Office (USPTO) data.…”