EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Expert-Augmented Deep Learning Approach for Synthesis Route Evaluation

Yujia Guo, Marco V. Mijangos, Tat Hong Duong Le, Samuel Genheden, Varvara Voinarovska, Giulia Bergonzini, Ola Engkvist, Samuel Kaski, Mikhail Kabeshov

Journal of Chemical Information and Modeling · 10.11.2026

DOI: 10.1021/acs.jcim.6c03207 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Chemical Information and Modeling, online erschienen am 10.11.2026. Verfasst von Yujia Guo und 8 weiteren Autor:innen. Die Arbeit nennt 27 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Research Council of Finland.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Selecting multistep synthetic routes remains a major bottleneck in computer-aided synthesis planning: high-quality negative data are scarce, and individual reported routes provide incomplete benchmarks for route quality. We introduce an expert-augmented framework that combines large-scale reference-based learning with expert-defined chemical feasibility assessment. Large-scale pretraining on literature-derived reference routes achieves 60.3% top 1 reference-route recovery on held-out targets. In expert assessment, the model’s top-ranked route is feasible in 88.4% of assessed cases, either matching a reference route or representing an alternative endorsed by experts. Among alternative top-ranked proposals, 75.4% are judged feasible, demonstrating that divergence from a literature reference does not necessarily indicate chemical infeasibility. The pretrained model captures a measurable expert-feasibility signal (Spearman |ρ| = 0.46), which increases to 0.78 ± 0.05 after expert fine-tuning. On an independent multi-expert dataset, agreement with the expert majority increases from 48.0% to 64.0% after fine-tuning, comparable to the 64.8% agreement between a held-out expert and the majority of their peers. Together, these results show that large-scale proxy supervision can provide scalable route prioritization, while limited expert supervision substantially improves alignment with chemical feasibility.

Förderung

  • Research Council of Finland

Zitieren

APA

Guo, Y., Mijangos, M. V., Le, T. H. D., Genheden, S., Voinarovska, V., Bergonzini, G., Engkvist, O., Kaski, S., & Kabeshov, M. (2026). Expert-Augmented Deep Learning Approach for Synthesis Route Evaluation. Journal of Chemical Information and Modeling. https://doi.org/10.1021/acs.jcim.6c03207

Vancouver

Guo Y, Mijangos MV, Le THD, Genheden S, Voinarovska V, Bergonzini G, et al. Expert-Augmented Deep Learning Approach for Synthesis Route Evaluation. Journal of Chemical Information and Modeling. 2026. doi:10.1021/acs.jcim.6c03207

BibTeX
@article{Guo2026,
  author = {Guo, Yujia and Mijangos, Marco V. and Le, Tat Hong Duong and Genheden, Samuel and Voinarovska, Varvara and Bergonzini, Giulia and Engkvist, Ola and Kaski, Samuel and Kabeshov, Mikhail},
  title = {Expert-Augmented Deep Learning Approach for Synthesis Route Evaluation},
  journal = {Journal of Chemical Information and Modeling},
  year = {2026},
  doi = {10.1021/acs.jcim.6c03207}
}
Kostenlos starten →