GoKawiilChemical synthesis remains a critical bottleneck in the discovery and manufacture of functional small molecules1−3. While AI-assisted synthesis planning has proliferated in recent years, a detailed understanding of its failure modes has not been achieved, and models still struggle with predicting less frequent, yet strategically critical reactions, as well as hallucinated, incorrect predictions misaligned with chemists’ expectations4−12. In this work, we analyze the failure modes of current AI models and propose RetroChimera: a frontier retrosynthesis model, built upon two newly developed components with complementary inductive biases, integrated via a novel, learning-based ensembling strategy. Through experiments across several orders of magnitude in data scale, we show RetroChimera outperforms leading baselines, demonstrating robustness outside the training data, as well as the ability to learn from very small numbers of examples per reaction class. Using both pairwise and pointwise setups, we find that organic chemists prefer predictions from RetroChimera over published reference reactions and over other AI models. Finally, we demonstrate zero-shot transfer and fine-tuning on internal datasets from two major pharmaceutical companies, showing robust generalization under distribution shift. Our work demonstrates the viability of deep learning for accurate synthesis prediction in increasingly challenging regimes.
Chemist-aligned retrosynthesis by ensembling diverse inductive bias models
Why This Matters
Chemical synthesis planning is a major bottleneck in drug and materials discovery, and existing AI retrosynthesis tools often fail on rare but strategically important reactions or produce chemically implausible predictions. RetroChimera addresses these gaps by combining models with complementary strengths through an ensembling approach, offering more reliable, chemist-preferred predictions. This matters because better AI-driven synthesis planning could accelerate drug discovery pipelines and reduce costly trial-and-error in pharmaceutical and chemical manufacturing.
Key Takeaways
- RetroChimera combines two models with different inductive biases via a learned ensembling strategy to outperform existing retrosynthesis AI baselines.
- The model generalizes well with limited training data and performs robustly on out-of-distribution and pharmaceutical company datasets via zero-shot transfer and fine-tuning.
- Organic chemists preferred RetroChimera's predictions over both published reference reactions and other AI models in evaluation studies.
Explore topics:
retrosynthesis
retrochimera
ai-assisted synthesis
deep learning chemistry
pharmaceutical datasets
Get alerts for these topics