Rxn·Lab Beta
Open-source retrosynthesis

Run, steer, and evaluate ML models for retrosynthesis.

RxnLab is an open-source platform to predict, steer, and evaluate retrosynthesis with modern ML models — from single-step disconnections to full multi-step routes. Built for chemists and ML researchers working on reaction-level generative modeling.

Once you've tried the platform, the feedback form takes ~2 minutes. Anonymous, focused on the UX of using RxnLab.

graph diffusion

DiffAlign (align-absorbing)

Graph diffusion model for single-step retrosynthesis.

USPTO-50k Single-step Multi-step
Open
DiffAlign (align-absorbing)
template based

LocalRetro

Template-based graph model for single-step retrosynthesis.

USPTO-50k Single-step Multi-step
Open
LocalRetro
seq2seq transformer

R-SMILES (RootAligned)

Template-free SMILES-to-SMILES transformer for single-step retrosynthesis.

USPTO-50k Single-step Multi-step
Open
R-SMILES (RootAligned)
graph edits

MEGAN

Graph-edit (semi-template) model for single-step retrosynthesis.

USPTO-50k Single-step Multi-step
Open
MEGAN