Chemical space projection
The v1 repository includes the projection benchmark used for Table 1 of the PrexSyn paper.
Run the full benchmark
The script evaluates the included 1,000-molecule Enamine REAL and ChEMBL subsets with 64, 128, and 256 samples per target. It performs five runs for every dataset/sample-count pair: 30 runs and 30,000 projections in total.
The default device is CUDA. Useful options are:
uv run python scripts/benchmarks/projection.py \
--device cuda \
--num-runs 5 \
--out ./outputs/benchmarks/analog
Each target result is cached in an LMDB file under the output directory. Rerunning the same settings resumes from cached entries. Delete or choose another output directory when you need an independent rerun.
After all runs finish, the script writes summary.csv and prints grouped similarity, reconstruction-rate, and timing statistics.
Paper settings
The paper reports results from one NVIDIA RTX 4090. Sampling is stochastic, and timing depends on the GPU, CPU, storage, and system load. Compare aggregate means and standard deviations rather than expecting identical rows.
The released configuration downloads the paper checkpoint and its matching Enamine/Rxn115 chemical-space cache automatically.