Publications

The models behind the lab, published.

Two papers, both with the benchmark and the baseline stated. Each links to the full text, the ablations and the figures.

  1. 01 / Rasyn AI Research / 2026

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

    Three architectures working together on one-step retrosynthesis: a graph neural network that proposes bond disconnections, an encoder-decoder Transformer with a copy mechanism that writes the reactant SMILES, and a fine-tuned language model that does edit-conditioned retrosynthesis in context. The paper also explains why the v1 model scored 85% token accuracy and 0.9% exact match, which is arithmetic rather than a bug.

    USPTO-50K Top-1
    69.7%C-SMILES 67.2%
    Parameters
    45.5M100% coverage
    Test reactions
    5,007Schneider split
    Read the paper

  2. 02 / Rasyn AI Research / 2026

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

    A multi-task 1D U-Net that learns peak detection, apex localisation and baseline estimation together, in place of the hand-tuned signal processing that MZmine, OpenMS and XCMS rely on. It keeps the ADAP chromatogram builder so the only thing being compared is the detector. The interesting result is the sim-to-real gap: MZmine loses about half its F1 moving off the simulated benchmark, and this does not.

    Real data (sep1) F1
    0.889MZmine 0.476
    Simulated F1
    0.943MZmine 0.937
    Parameters
    1.97M32/32 real peaks
    Read the paper