RETROSPECT: RETROsynthesis via Sequential Prediction, and Chemically Transformed-ranking

RETROSPECT introduces a modular retrosynthesis system combining a Transformer-based proposal model with LambdaMART reranking to improve chemical synthesis prediction. The system achieves 55% top-1 accuracy on USPTO-50K benchmarks, demonstrating that decomposing retrosynthesis into proposal generation and learned selection improves both ranking quality and candidate diversity.

RETROSPECT addresses a fundamental challenge in computational chemistry: predicting viable synthetic routes for target molecules. The research decomposes retrosynthesis into two complementary stages—proposal generation and intelligent reranking—rather than attempting end-to-end optimization. This modular architecture allows independent improvement of each component and facilitates integration into larger ensemble systems.

The ChemAlign Transformer generator employs several sophisticated training techniques including hybrid SMILES augmentation strategies, exponential moving average weights, and a differentiable atom-balance loss function. These design choices collectively push the proposal model to 55% top-1 accuracy while maintaining 99.86% chemical validity, a critical requirement for practical application. The reranking stage leverages structural descriptors and reaction-template statistics, with feature ablations revealing that upstream proposal scores and template frequency provide the strongest signal for candidate prioritization.

DFT-derived quantum mechanical descriptors show marginal improvement, suggesting that classical chemical features remain highly informative for retrosynthesis ranking. This finding has important implications for deployment efficiency, as computationally expensive quantum calculations may not justify their contribution to overall system performance.

The framework enables practical improvement across two dimensions: stronger base proposals feed richer candidate pools to the reranker, while learned selection compensates for proposal model limitations. By achieving 59.4% top-1 accuracy on the merged benchmark, RETROSPECT demonstrates that this decomposition approach outperforms single-stage alternatives. The modularity supports integration into RetroChimera and similar ensemble architectures, positioning this work as infrastructure for downstream chemistry applications rather than a complete end-to-end solution.

• →RETROSPECT achieves 59.4% top-1 accuracy using proposal-selection decomposition on USPTO chemistry benchmarks
• →Modular architecture enables drop-in integration into existing ensemble systems like RetroChimera
• →Upstream proposal scores and template statistics provide most reranking signal; expensive DFT features show marginal gains
• →System maintains 99.86% top-1 validity, critical for practical synthesis prediction applications
• →Hybrid training techniques including SMILES augmentation and differentiable auxiliary losses improve both accuracy and candidate quality