PrexSyn

Introduction

PrexSyn is an efficient, accurate, and programmable framework for synthesizable molecular design.

It uses a decoder-only transformer to autoregressively generate postfix notations of synthesis¹: a molecular representation based on chemical reactions and purchasable building blocks. Generation is conditioned on molecular descriptors.

PrexSyn was trained on a billion-scale datastream of postfix notations paired with molecular descriptors using two GPUs and 32 CPU cores in two days. This scale is enabled by PrexSyn Engine, a real-time, high-throughput C++ data generation pipeline.

Need the exact paper features?

Use the dev-v0 branch. General physicochemical-property conditioning and composite logical queries from the original paper are deprecated in v1. See Paper and v1 differences.

“Synthesizable” means constructible under the configured reaction templates and building blocks. It is not a guarantee of experimental success.

Capabilities

Capability	Input	Output
Chemical-space projection	Graph or SMILES	Ranked analogs and pathways
Fingerprint/descriptor-based generation	ECFP4 or FCFP4	Molecules and pathways
Molecular sampling	Scoring function	Optimized candidates

Performance

The following figures show the results reported in the PrexSyn paper. The projection benchmark is maintained in v1. Migration of the optimization benchmark to v1 is work in progress.

Capability	Result
Record-high accuracy and speed in chemical-space projection and fingerprint/descriptor-based generation
Record-high sample efficiency in molecular sampling against scoring functions

Start here

Install PrexSyn.
Run the projection example.
Use the Python API or define a chemical space.

Resources

Repositories

PrexSyn: https://github.com/luost26/prexsyn
PrexSyn Engine: C++ backend for high-throughput training data generation and synthesis detokenization. https://github.com/luost26/prexsyn-engine
Data and model weights: Preprocessed chemical spaces and trained model weights. https://huggingface.co/datasets/luost26/prexsyn-data/tree/main

Papers and documentation

PrexSyn paper: Efficient and Programmable Exploration of Synthesizable Chemical Space. https://arxiv.org/abs/2512.00384
PrexSyn documentation: https://prexsyn.readthedocs.io

Community

MIT Coley Research Group: https://coley.mit.edu/

Citation

@article{luo2025prexsyn,
  title   = {Efficient and Programmable Exploration of Synthesizable Chemical Space},
  author  = {Shitong Luo and Connor W. Coley},
  year    = {2025},
  journal = {arXiv preprint arXiv: 2512.00384}
}

Projecting Molecules into Synthesizable Chemical Spaces. https://arxiv.org/abs/2406.04628 ↩