Reproducibility package for "Hierarchical Phase Transitions in Bayesian Flow Networks Enable Training-Free Molecular Design" (JCIM 2026, ci-2026-015382)

Abstract

# PAIR v3: code and reproducibility assets This record accompanies **Hierarchical Phase Transitions in Bayesian FlowNetworks Enable Training-Free Molecular Design**, by Liangji Zhou, YuanqiuChen, Tongtong Chen, Yijuan Huang, Qihao Pan, Si Li, Chenyu Gou, Meiqiong Fanand Yu Gao (Journal of Chemical Information and Modeling, manuscriptci-2026-015382). PAIR (Phase-Aware Iterative Refinement) is a training-free inference strategyapplied to the frozen pocket-conditioned Bayesian Flow Network calledPocketBFN in the manuscript. The software distribution and exportedcheckpoint filename use PAIR. Source repository: https://github.com/ZLJLZX/PAIR Version DOI: https://doi.org/10.5281/zenodo.22693741 ## Contents - **PAIR-v3-code.zip**: source code, sampling configuration, installation tools, regression tests, numerical analysis programs and figure scripts. The archive restores a `PAIR/` directory. The PAIR entry point is `scripts/iterative_bfn.py`; `demo_pocket.py` provides a single-pocket example. `analysis/` contains the statistical and pose-audit programs; `figure_scripts/` contains the manuscript figure sources.- **PAIR-v3-assets.zip**: the pretrained checkpoint, per-pocket and per-molecule numerical results, generated poses, PoseBusters records, concentration measurements, redocking controls, recorded environments and timing data. It restores `data/checkpoints/PAIR.ckpt` and `data/research/`. The legacy environment description is retained as `data/research/data/environment_legacy.yml`.- **PAIR-v3-figures.zip**: the final manuscript and Supporting Information figure PDFs, including the TOC graphic. Figure text is black; colours used to encode plotted data are preserved.- **SHA256SUMS.txt**: checksums of the upload files. The source ZIP also includes a per-file source checksum manifest. The v3 source is a corrected, portable release. It includes strict checkpointloading, a paper-aligned demonstration, explicit input and dependency checks,bounded chemistry operations, and fixes found during CPU and figure tests.The data archive consolidates the existing deposited research assets;recorded measurements and checkpoint bytes are preserved. Legacy identifiersinside raw records are retained as provenance. Earlier source trees andmanuscript administration tools are not duplicated in this release. ## Reuse and reproduction Extract the source ZIP, follow its README to install a dedicated Python 3.11environment, and run: python tools/prepare_assets.py --archive PAIR-v3-assets.zip python tools/reproduce_analysis.py python tools/render_figures.py The analysis wrapper recomputes the summaries from deposited per-moleculerecords and saves logs. It does not repeat GPU generation or docking.The supplied model smoke tests exercise checkpoint loading, gradients,sampling and molecular reconstruction. Full CUDA/Vina benchmark executionwas not repeated for the v3 source release; historical timing results referto the recorded original environment, not the revised process-isolatedlauncher. Fresh pose audits require the matching CrossDocked2020 receptor structures,which are third-party assets and are not redistributed. Recorded receptoridentifiers accompany the generated poses. The original text8 checkpointis unavailable, so its deposited concentration measurements can bereanalysed but cannot be regenerated from this record alone. The source has mixed licences. The authors' original code is MIT-licensed;adapted components retain CC BY-NC-SA 4.0, GPL-2.0, MIT or Apache-2.0 termsas specified in `THIRD_PARTY_NOTICES.md` and the included licence texts.The source licence does not grant new rights to model weights orthird-party data. Their original rights and provenance remain applicable.