Pushing the boundaries of AI-driven chemistry

From retrosynthetic planning to chromatographic analysis, our research combines deep learning with domain expertise to build tools that work in practice.

Publications

Our research contributions to
computer-aided chemistry.

Larmor: Equivariant Graph Networks for First-Principles-Quality NMR Chemical Shift Prediction
New
Rasyn AI Research·2026

Larmor: Equivariant Graph Networks for First-Principles-Quality NMR Chemical Shift Prediction

Larmor-50M / Larmor-100M predict 1H and 13C NMR shifts at 0.131 ppm / 1.04 ppm MAE on nmrshiftdb2 - surpassing DFT B3LYP/PBE0 + GIAO at ~10,000× the speed. E(3)-equivariant attention over 3D geometry, trained on 1.2M experimental shifts.

Read announcement
Marcus: A Hierarchical Transformer for Reaction Condition Prediction
New
Rasyn AI Research·2026

Marcus: A Hierarchical Transformer for Reaction Condition Prediction

Marcus-100M / Marcus-1B jointly predict solvent, catalyst, base, temperature, and yield. 92.1% top-3 solvent accuracy, R²=0.72 on yield. Trained on 4.7M Reaxys + USPTO + ORD reactions.

Read announcement
Introducing Arcade: AI-Powered Institutional Memory for Chemistry Labs
Rasyn AI·2026

Introducing Arcade: AI-Powered Institutional Memory for Chemistry Labs

Arcade automatically indexes every experiment, route, dataset, and protocol your lab produces. Hybrid retrieval (BM25 + vector + numeric filters) with grounded, cited answers.

Read paper
Rasyn: A Hybrid AI Framework for Single-Step Retrosynthetic Analysis Combining Graph Neural Networks, Transformers, and Large Language Models
Rasyn AI Research·2026

Rasyn: A Hybrid AI Framework for Single-Step Retrosynthetic Analysis Combining Graph Neural Networks, Transformers, and Large Language Models

69.7% Top-1 accuracy on USPTO-50K with 45.5M parameters. Detailed ablation of 6 architectural innovations, token-sequence accuracy analysis, and improvement roadmap.

Read paper
ChromPeakNet: Deep Learning for Chromatographic Peak Detection in Untargeted Mass Spectrometry
Rasyn AI Research·2026

ChromPeakNet: Deep Learning for Chromatographic Peak Detection in Untargeted Mass Spectrometry

Multi-task 1D U-Net (1.97M parameters) achieves F1=0.943 on simulated and F1=0.889 on real LC-MS data, surpassing MZmine, OpenMS, and XCMS with <6% sim-to-real degradation.

Read paper

Benchmarks

State-of-the-art accuracy
on USPTO-50K.

USPTO-50K - Schneider split, 5,007 test reactions
Top-1 Accuracy
RetroTransformer v2 + R-SMILESOurs
69.7%
GraphRetro (2021)
53.7%
LocalRetro (2021)
53.4%
GLN (2019)
52.5%
Neuralsym (2017)
44.4%

Our Best Model

RetroTransformer v2

R-SMILES 20x augmentation

Top-1 Accuracy69.7%
Top-5 Accuracy89.3%
Top-10 Accuracy93.1%
Coverage100%
45.5M parameters

Ready to accelerate your synthesis?

Start with a free account. Input any target molecule and get AI-generated routes, protocols, and vendor sourcing in seconds.

Start for freeView Pricing

No credit card required

3D molecular structure