Source-linked AI summary
RetroMPA: A Molecular Property-Aware Auxiliary Framework for Enhancing Retrosynthesis Prediction
Mianzhi Liu, Fan Xiao, Zhiliang Yu, Huayang Huang, Yuke Li, Yi Yang, Wenbo Liu, Yu Wu
TL;DR
Data-driven retrosynthesis tools assist synthetic planning, but integrating molecular chemical knowledge remains limited. RetroMPA adds a property-aware auxiliary refinement framework to existing models, with laboratory analyses finding predicted products consistent with NMR measurements.
Problem
Data-driven computational tools assist or automate retrosynthetic planning, while single-step retrosynthesis identifies plausible reactants for a given product.
Method
RetroMPA uses pretrained molecule–text embeddings and optional dictionary-guided Top-K refinement to inject molecular property knowledge into existing retrosynthesis predictions.
Results
Laboratory products from predicted reactions showed NMR spectra in complete agreement with the model’s predictions.
Takeaways & Limitations
NMR-validated products provide preliminary support for RetroMPA’s generalization to real-world chemistry experiments.
Abstract
from arXiv · showhide
Retrosynthesis is a cornerstone of drug discovery and organic synthesis. While data-driven deep learning models have shown remarkable progress, they autonomously learn reaction patterns from extensive datasets with limited integration of established chemical knowledge as priors. To address this limitation, we introduce RetroMPA, a molecular property-aware, post-hoc enhancement module that injects chemical knowledge into the retrosynthesis pipeline. Rather than functioning as an independent SMILES sequence generator, RetroMPA is a broadly applicable, model-agnostic chemical filter designed to recalibrate and optimize the predictive pathways of existing algorithms. This plug-and-play framework integrates seamlessly with a range of data-driven retrosynthesis methods, enhancing outputs without modifying model architecture or requiring resource-intensive retraining. By leveraging a property-aware latent embedding space, RetroMPA consistently improves top-1 accuracy across eight representative retrosynthesis models by an average of 5.50% on USPTO-50K. Furthermore, we validate its scalability on the large-scale USPTO-Full dataset, achieving an average improvement of about 2.03% across both template-based and template-free architectures. Wet-lab experiments provide preliminary support for the practical utility of the framework. These syntheses confirmed viable, previously unreported substrate combinations for classic reaction paradigms---specifically, Suzuki-Miyaura coupling, Bucherer reaction, and Friedel-Crafts acylation---suggesting that RetroMPA can operate beyond mere data fitting. The code is open-sourced at https://github.com/MengzhouLu/RetroMPA.
Introduction
The introduction motivates RetroMPA as a molecular-property-aware, model-agnostic way to inject chemical knowledge into retrosynthesis without restructuring or retraining established predictors. It frames the framework as improving chemically plausible predictions and supporting synthetically viable routes beyond training data.
- Retrosynthesis recursively decomposes target molecules into simpler precursors, with single-step retrosynthesis identifying plausible reactants for a given product.
- Existing retrosynthesis methods use SMILES, molecular graphs, reaction templates, or deep-learning architectures, but remain limited by representation difficulty, template coverage, atom-mapping dependence, or data-driven reasoning.
- Molecular properties influence reaction feasibility, pathways, and efficiency, motivating their integration as explicit chemical knowledge in retrosynthesis models.
- RetroMPA enhances existing retrosynthesis models without architectural modification or retraining by learning molecular-property representations from multimodal molecule–text knowledge and using property-space transformations as chemical priors.
- Across USPTO-50K and USPTO-Full evaluations, RetroMPA improves prediction accuracy, chemical plausibility, and model-agnostic adaptability, with potential to discover synthetically viable routes beyond training data.
Results and discussion
RetroMPA consistently improves retrosynthesis top-1 accuracy across diverse base models without retraining, with gains on both USPTO-50K and USPTO-Full. Its improvements are larger for weaker baselines and it remains applicable in larger, chemically diverse settings.
- USPTO-50K results: 5.50% average top-1 accuracy gain was achieved across eight USPTO-50K base models without retraining or fine-tuning.The models spanned template-based, semi-template-based, and template-free paradigms.
- USPTO-50K results: +8.19% was the highest absolute improvement, observed for the Transformer model lacking explicit chemical inductive biases.This result suggests that RetroMPA can compensate for missing chemical priors.
- USPTO-Full scalability: 2.80% absolute gain for LocalRetro and 1.26% improvement for EditRetro were obtained on USPTO-Full.The scalability evaluation covered one template-based and one template-free paradigm because of computational and model-weight constraints.
- Computational efficiency: 130 hours of comprehensive network training were required on four NVIDIA RTX 4090 GPUs and dual Intel Xeon Platinum 8352V CPUs.The resource-intensive stages were universal Mol-Former pretraining and Molecular Decoder optimization.
- Improvement analysis: r = −0.96 indicates a strong negative Pearson correlation between base-model performance and RetroMPA’s improvement on USPTO-50K.Weaker base models generally received larger absolute gains, whereas gains diminished as baseline accuracy increased.
Conclusion
RetroMPA is a plug-and-play, molecular property-aware framework that injects chemical knowledge into existing retrosynthesis models without retraining or architectural modification. It improves top-1 accuracy across USPTO benchmarks, while remaining subject to reactant-input and molecular-uniqueness limitations.
- Framework: RetroMPA constrains and refines retrosynthesis predictions through a chemically grounded molecular-property embedding space rather than relying solely on atom-level co-occurrence patterns.This design approximates aspects of property-based chemical reasoning used by chemists.
- Framework: The plug-and-play module integrates with template-based, semi-template-based, and template-free architectures without retraining or modifying existing base models.Its modularity enabled post-hoc integration across eight representative methods.
- Results: 5.50% average gain in top-1 accuracy was achieved on USPTO-50K without retraining any base model, while USPTO-Full showed approximately 2% average improvement.The results indicate consistent improvements across both benchmarks, including the more complex and diverse USPTO-Full dataset.
- Results: Wet-lab validation provided preliminary evidence that RetroMPA can propose synthetically viable, previously unreported substrate combinations.The experiments supported the framework’s practical utility beyond purely data-driven evaluation.
- Limitations: RetroMPA requires at least one predicted reactant and therefore does not directly apply to single-reactant reactions; nearest-neighbor retrieval may also return structurally similar rather than identical molecules.The inverse projection from continuous embeddings to discrete SMILES does not guarantee molecular uniqueness.
- Future directions: Future directions include iterative application to multi-step planning and resolving molecular uniqueness through improved methods.The framework is presented as a modular foundation adaptable to new base models and chemical databases without architectural overhaul.
Method
RetroMPA is a model-agnostic auxiliary framework that injects molecular-property knowledge into retrosynthesis through multimodal pretraining, molecular-level decoding, and chemically informed representation learning. Its method uses curated reaction inputs, leakage controls, and auxiliary prediction to refine base-model outputs.
- Data curation and evaluation: Evaluation uses USPTO-50K and USPTO-Full, while preprocessing discards reagents, canonicalizes SMILES, and removes atom mappings and reaction class labels to reduce noise and information leakage.The molecular dictionary is reserved for inference and is not used for training or gradient updates; no SMILES augmentation or atom-mapping-based filtering is applied.
- Framework overview: RetroMPA combines Mol-Former and a Molecular Decoder to predict reactant embeddings from chemically informed molecular-property representations and recover molecules through Inverse Projection.Mol-Former maps molecules into vectors with explicit molecular-property semantics, while the decoder predicts corresponding reactant embeddings.
- Auxiliary retrosynthesis: Auxiliary retrosynthesis predicts remaining reactants from the product and one reactant supplied by the base model, refining the base model’s candidate output.The task predicts {R2, ..., Rn} given product P and first predicted reactant R1.
- Chemical knowledge injection: Multimodal pretraining aligns molecular and textual representations so chemical terms and molecular structures jointly inform property embeddings used as retrosynthesis priors.Mol-Former comprises a Molecular Transformer, Text Transformer, and Molecular Encoder, with contrastive, matching, and language-generation objectives.
- Molecular-level prediction: Molecular-level generation produces complete reactant molecules sequentially, and teacher forcing conditions each step on true preceding reactants during training.Training uses ground-truth reactants rather than base-model predictions, supporting compatibility across diverse predictors.
- Representation learning: Dual-prior contrastive learning combines Molecular Contrastive Loss and Molecular Matching Loss to make predicted reactant representations accurate, discriminative, and chemically meaningful.MML minimizes L2 distance to ground-truth embeddings and uses hard negatives to prevent representation collapse and separate difficult negatives.
TOC Graphic
The TOC graphic presents RetroMPA as knowledge driven, yielding optimized predictions with improved reactant pairs by distinguishing less suitable from more suitable options.
- RetroMPA is presented as knowledge driven.
- The framework produces optimized predictions with improved reactant pairs.
- The graphic contrasts less suitable and more suitable candidates.
S1.Analysis of the chemical datasets used for pretraining
RetroMPA pretrains its Mol-Former component on three large-scale chemical-text datasets to learn molecular chemical properties and guide retrosynthesis. Collectively, these datasets provide rich molecular knowledge and semantic grounding, including descriptions of functional groups and chirality.
- Dataset composition and purpose: Three datasets—ChEBI-20-MM, Mol-Instructions, and PubChemSTM—are used to pretrain Mol-Former for molecular chemical-property learning and retrosynthesis guidance.Together, they provide rich molecular knowledge for the model.
- Dataset composition and purpose: The chemical-text sources provide semantic grounding by explicitly describing molecular attributes such as functional groups and chirality.The passage notes that such descriptions constitute a portion of the text.
S2.Experimental chemical analysis
Laboratory experiments assessed whether model-predicted products were obtained in real-world reactions. Product identities were evaluated by a third-party institution using NMR spectroscopy and compared with the predictions.
- Experimental validation: The study conducted chemistry-laboratory experiments to test the model’s generalization in real-world scenarios.The experiments were based on the model’s predictions.
- Experimental validation: A third-party testing institution analyzed the reaction products using nuclear magnetic resonance spectroscopy.Chemical substances were identified through spectral analysis.
- Experimental validation: The NMR analyses determined whether the reaction products matched the model’s predictions.Supplementary figures present the ¹H and ¹³C NMR spectra of all chemical products.
S3. Inference algorithm with Dynamic Molecular Dictionary
RetroMPA’s streamlined inference optionally updates a dynamic molecular dictionary, then refines a base model’s Top-1 reactant pair through a conditional candidate-ranking consistency check. The original pair is retained when the target reactant ranks within the Top-K candidates; otherwise, it is replaced by the highest-ranked refined candidate.
- Initial prediction: The workflow begins by retrieving the Top-1 reactant pair (b1, b2) from a base retrosynthesis model for product SMILES P.The final output is a reactant combination Rfinal, with the dictionary D optionally updated.
- Consistency-based selection: If the original target reactant ranks within C[1:K], the base pair is retained; otherwise, it is replaced with the highest-ranked candidate C[1].This consistency check produces Rfinal as either the original pair or the conditioned refined combination.
- Conditional refinement: K = 10, and the method randomly selects one base reactant as the condition while predicting the other reactant.Imp generates Top-K candidates for the target reactant conditioned on the selected reactant and product P.
- Dynamic dictionary: When enabled, the dynamic dictionary D is updated with each predicted reactant’s SMILES and Mol-Former embedding.The update is optional and occurs for both b1 and b2.