MechRetro is a chemical -mechanism -driven graph learning framework for interpretable retrosynthesis prediction and pathway planning Yu Wang12 Chao Pang12 Yuzhe Wang12 Yi Jiang12 Junru Jin12 Sirui Liang12 Quan Zou3

2025-05-02 0 0 2.07MB 29 页 10玖币

侵权投诉

MechRetro is a chemical-mechanism-driven graph learning framework for

interpretable retrosynthesis prediction and pathway planning

Yu Wang1,2, Chao Pang1,2, Yuzhe Wang1,2, Yi Jiang1,2, Junru Jin1,2, Sirui Liang1,2, Quan Zou3*,

and Leyi Wei1,2*

1 School of Software, Shandong University, Jinan 250101, China

2Joint SDU-NTU Centre for Artificial Intelligence Research (C-FAIR), Shandong University,

Jinan 250101, China

3Institute of Fundamental and Frontier Sciences, University of Electronic Science and

Technology of China.

*Corresponding author:

Q.Z: zouquan@nclab.net

L.W: weileyi@sdu.edu.cn

Abstract

Leveraging artificial intelligence for automatic retrosynthesis speeds up organic pathway

planning in digital laboratories. However, existing deep learning approaches are

unexplainable, like "black box" with few insights, notably limiting their applications in real

retrosynthesis scenarios. Here, we propose MechRetro, a chemical-mechanism-driven

graph learning framework for interpretable retrosynthetic prediction and pathway planning,

which learns several retrosynthetic actions to simulate a reverse reaction via elaborate

self-adaptive joint learning. By integrating chemical knowledge as prior information, we

design a novel Graph Transformer architecture to adaptively learn discriminative and

chemically meaningful molecule representations, highlighting the strong capacity in

molecule feature representation learning. We demonstrate that MechRetro outperforms the

state-of-the-art approaches for retrosynthetic prediction with a large margin on large-scale

benchmark datasets. Extending MechRetro to the multi-step retrosynthesis analysis, we

identify efficient synthetic routes via an interpretable reasoning mechanism, leading to a

better understanding in the realm of knowledgeable synthetic chemists. We also showcase

that MechRetro discovers a novel pathway for protokylol, along with energy scores for

uncertainty assessment, broadening the applicability for practical scenarios. Overall, we

expect MechRetro to provide meaningful insights for high-throughput automated organic

synthesis in drug discovery.

Introduction

Retrosynthesis aims to identify a set of appropriate reactants for the efficient synthesis of

target molecules, which is indispensable and fundamental in computer-assisted synthetic

planning1-3. Retrosynthetic analysis is first formalized by Corey4-6 and solved by the

Organic Chemical Simulation of Synthesis (OCSS) program. Later, driven by sizeable

experimental reaction data and significantly increased computational capabilities, various

machine-learning-based approaches7, especially deep-learning (DL) models, have been

proposed and achieved incremental performance8. From the perspective of molecule

representations, existing DL-based retrosynthesis approaches can be generally

categorized into two classes: (1) Sequence-based retrosynthesis approaches9-11, and (2)

Graph-based retrosynthesis approaches12,13. The former utilizes sequential structures to

represent molecules and predicts by encoder-decoder-based neural language architecture,

while the latter employs graph structures to describe molecules and forecasts through

molding the graph changes between products and reactants.

Sequence-based approaches utilize serialized notations (e.g., SMILES14, SELFIES15) to

describe molecules and leverage the encoder-decoder framework to translate target

products into potential reactants. Liu et al.16 present a model, namely Seq2Seq, which

consists of a bidirectional long short-term memory (LSTM)17 encoder and decoder for

retrosynthetic translation. Later, with the development of neural machine translation

models, like Transformer18, sequenced-based retrosynthetic approaches achieve gradual

performance improvement. Karpov et al.19 first adapt Transformer architecture with

modified learning rate schedules and snapshot learning to retrosynthesis. Segler et al.9

introduce specialized data augmentation techniques on retrosynthetic Transformers,

demonstrating the combination of SMILES augmentation leads to better results. Afterward,

as the pretraining-finetuning paradigm rises, Irwin et al.20 propose MolBART with large-

scale and self-supervised pre-training, which significantly speeds up the convergence of

the retrosynthesis task. However, there exist the following two limitations. First, the direct

structural information in molecules is denoted by implicit symbols, leading to more

computational costs and information loss. Second, the popular molecular representation

approaches are grammatically strict and not guaranteed semantically valid, resulting in

frequent invalid syntaxes. To address the limitations, Ucak et al. propose RetroTRAE21,

alleviating the structural information loss by encoding atom environments; Zheng et al.22

propose SCROP, effectively avoiding illegal outcomes through coupling with a DL-based

syntax corrector.

Graph-based approaches usually adopt graph structures to represent molecules and

predict graph changes of the target molecule to infer the reactants. The graph changes are

most modeled by the two-stage paradigm that contains reaction center prediction (RCP)

and synthon completion (SC). Jin et al.12 propose WLDN, using the Weisfeiler-Lehman

isomorphism test23 for retrosynthetic graph learning. Later, with the development of graph

neural networks (GNNs), many GNN-based frameworks emerged and achieved a notable

improvement in performance. Shi et al.24 present the G2G framework, which utilizes R-

GCN25 for RCP and reinforcement learning for SC. Following the same paradigm, Yan et

al.26 and Somnath et al. 13 devise RetroXpert and GraphRetro, respectively. The former

applies a GAT27 variant for RCP and a sequence-based Transformer for SC, while the latter

designs two MPNNs28 for the two stages. Different from the above approaches, Dai et al.29

propose GLN, a novel method that leverages reaction templates to connect products and

reactants. Nevertheless, traditional GNNs excessively focus on local structures of

molecules, neglecting the effect of long-distance characteristics (e.g., Van der Waals' force).

To solve the problem, Ying et al.30 propose Graphormer, introducing a shortest-path-based

method for multi-scale topological encoding.

Although existing retrosynthesis approaches have achieved significant progress in

accelerating the data-driven retrosynthesis prediction, they still suffer from the following

intrinsic problems. (1) Sequence-based approaches suffer from the loss of topological

information of molecules susceptible to chemical reactions. Meanwhile, graph-based

approaches neglect sequential information and long-range characteristics. Both of them

are constrained in feature representation learning, limiting further performance

improvement. (2) Existing deep-learning-based retrosynthesis approaches face a common

problem; that is, the decision-making mechanism of the models remains unclear, which

restricts the model's reliability and practical applications. (3) Most existing approaches

focus only on the single-step retrosynthesis prediction that enables generating plausible

reactants but is perhaps not accessible. Therefore, the multi-step retrosynthesis prediction

with pathway planning from products to accessible reactants is much more meaningful for

experimental researchers in practical chemical synthesis.

In this work, we propose MechRetro, an innovative deep graph learning framework via

integrating chemical prior knowledge for interpretable retrosynthesis prediction and

pathway planning. To learn discriminative molecule representations, we propose a novel

Graph Transformer architecture that enables sufficiently capturing both topological

information and sequential information of molecules. Comparative results show that

MechRetro outperforms the state-of-the-art approaches with a large margin on benchmark

datasets. Via interpretable analysis, we showcase MechRetro adaptively captures the

discriminative sub-structures of molecules concerning different chemical reactions,

indicating the vital capacity in molecule feature representation learning. To simulate the

real chemical reaction mechanism and guarantee the chemical rules, we model three

reaction steps (e.g., bond changes, leaving group matching, and leaving group connection)

through a self-adaptive joint-learning strategy, leading to the transparency of the model

decision process and improving the model interpretability. We demonstrate that the self-

adaptive joint-learning strategy can learn shared knowledge across different chemical

reactions, improving the model's performance and robustness. Importantly, by integrating

a heuristic tree-search algorithm into MechRetro, we can identify efficient synthetic routes

for drug molecules in the multi-step retrosynthesis prediction, demonstrating the strong

ability in high-quality pathway planning. Interestingly, case study results on some well-

known drugs show that MechRetro discovers and verifies credible synthetic routes that

have not been reported in the literature before, highlighting the practicability in real

scenarios. We expect MechRetro to provide meaningful insights for high-throughput

automated organic synthesis in drug discovery.

Methods and materials

The framework of the proposed MechRetro

The overview of our MechRetro is illustrated in Fig.1a. As can be seen, our MechRetro

contains four major modules: (1) Graph Transformer module, (2) Self-adaptive learning

module, (3) Decision-making module, and (4) Prediction & pathway planning module. The

workflow of our MechRetro is described as follows.

In the Graph Transformer module (see Fig.1b), taking the processed molecule graph data

as input, we first propose a multi-sense and multi-scale bond embedding strategy for the

bonds of molecules while introducing topological and contrastive atom embedding

strategies for the atoms of molecules, to capture chemically important information.

Afterward, the model learns hidden molecular representations from two embedding entries

via a multi-head attention mechanism. Next, in the Self-adaptive learning module, the

resulting hidden molecular representations are parallelly passed into three deep neural

decision units: reaction center predictor (RCP), leaving group matcher (LGM) and leaving

group connecter (LGC), which are trained simultaneously by an elaborate self-adaptive

learning strategy to maintain the shared knowledge. Among the decision units, RCP

identifies the changes of bonds and atoms' hydrogen count (see Fig.1c); LGM matches

the leaving groups (see section Reaction center and leaving group for details) from the

collected database for products (see Fig.1d); LGC connects the leaving groups and

fragments from the product (see Fig.1e). In the Decision-making module, the product is

transformed into reactants after a decision process consisting of five retrosynthetic actions

(see Fig.1f) and energy scores for uncertainty assessment, which simulates a reversed

chemical mechanism process. In the last module, based on the single-step predictions, a

heuristic tree-search algorithm is integrated to discover efficient synthetic routes with

transparent decision-making processes while guaranteeing the accessibility of start

reactants. The details of the four modules are described as follows.

Fig.1. Overview of MechRetro. a. MechRetro contains four major modules: (1) Graph Transformer, (2) Self-adaptive

learning, (3) Decision making, and (4) Prediction & pathway planning. b. MechRetro introduces (1) multi-sense and

multi-scale bond embeddings and (2) topological and contrastive atom embeddings for more chemically important

information. c. RCP recognizes bond changes between product and reactants and hydrogen changes for each atom.

d. LGM picks out an appropriate leaving group from a collected leaving group set. e. LGC identifies which atom in the

reaction center should connect to the gate atom of the predicted leaving group by calculating a connection matrix

between the original product and the predicted leaving group. f. The product is predicted to the reactants via an

interpretable reasoning process in our retrosynthesis model. The reasoning process can provide more insights into the

retrosynthesis prediction in real scenarios.

Problem definition

For the convenience of description and discussion., we here briefly introduce the problem

definition.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

MechRetroisachemical-mechanism-drivengraphlearningframeworkforinterpretableretrosynthesispredictionandpathwayplanningYuWang1,2,ChaoPang1,2,YuzheWang1,2,YiJiang1,2,JunruJin1,2,SiruiLiang1,2,QuanZou3*,andLeyiWei1,2*1SchoolofSoftware,ShandongUniversity,Jinan250101,China2JointSDU-NTUCentreforArtificialI...

展开>> 收起<<

MechRetro is a chemical -mechanism -driven graph learning framework for interpretable retrosynthesis prediction and pathway planning Yu Wang12 Chao Pang12 Yuzhe Wang12 Yi Jiang12 Junru Jin12 Sirui Liang12 Quan Zou3.pdf

共29页,预览5页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

MechRetro is a chemical -mechanism -driven graph learning framework for interpretable retrosynthesis prediction and pathway planning Yu Wang12 Chao Pang12 Yuzhe Wang12 Yi Jiang12 Junru Jin12 Sirui Liang12 Quan Zou3

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: