Newest tag from 2023, news log running to 2025, default branch dev

The tags stopped in 2023 while the announcements kept going

Three releases are visible: v1.0 on 2021-07-09, v1.1.0 on 2021-12-15, and v1.2.2 on 2023-08-01. The gap in that sequence is as informative as the entries, since 1.1.0 is followed by 1.2.2 with no 1.2.0 or 1.2.1 in sight. The announcements, by contrast, run to 2025.07.23, when HelixFold3.2 was released as an improvement on HelixFold3 in protein related tasks and structural quality, with its code at `apps/protein_folding/helixfold3`.

So the newest tag is roughly two and a half years older than the newest announcement, and the last recorded push to the repository is 2026-03-31. The default branch is `dev` rather than `main`, and one older entry links into it directly, the 2021.11.23 protein-protein interaction paper pointing at `tree/dev/apps/protein_protein_interaction`. Anyone following version numbers here is reading a series that stopped before most of the work on the page existed.

Open for academic research, paid API for everyone else

The licensing story changes inside the timeline. The 2024.08.15 entry releases the codes and model parameters of HelixFold3, described as biomolecular structure prediction replicating the capabilities of AlphaFold3, and states that the initial release is available as open source on GitHub for non-commercial academic research. The accuracy claim is scoped in the same sentence: comparable to AlphaFold3 in predicting the structures of conventional ligands, nucleic acids, and proteins.

Three months later the picture is different. The 2024.11.08 entry introduces a convenient paid API for academic and commercial applications, with a usage guide on the project site, described as a way to make HelixFold3 integration easier and to support high-throughput use. Between the two sits the 2024.08.30 note that an initial version of the HelixFold3 server is available on the project's website.

For anyone outside academic research, the usable path is therefore the hosted API and the hosted server, not the repository. The repository remains the place where the models were first published under a non-commercial label.

One paper's code link points at somebody else's repository

Almost every entry in the news log links into `apps/`, and those paths are the real index of the collection. Protein folding appears as `apps/protein_folding/helixfold`, `apps/protein_folding/helixfold-single`, and `apps/protein_folding/helixfold3`. Docking is `apps/molecular_docking/helixdock`. Compound pre-training is `apps/pretrained_compound/ChemRL/GEM` and the same path one level deeper for `GEM-2`. Drug target interaction is `apps/drug_target_interaction/batchdta`.

The exception is instructive. The 2024.05.13 entry for the motif-aware RNA language modelling paper accepted by Nature Machine Intelligence links its code to `github.com/CatIIIIIIII/RNAErnie`, a personal repository outside this project rather than a path under `apps/`. A few other entries point at papers with no code path at all, including the HelixADMET and HelixMO publications, so the log mixes code releases with paper announcements and does not mark which is which.

The page is a news log, and the guides are elsewhere

There is no install command on this page. Everything at the top is dated announcements, and the navigational work is pushed into files: `installation_guide.md` and `developer_guide.md`, each with a Chinese mirror named `_cn.md`, a `docs/` directory that a `.readthedocs.yml` builds, a `tutorials/` directory, and a second README in Chinese. The root also holds `competition/`, `research/`, `scripts/`, and a `c/` directory for native code with a `CMakeLists.txt` beside it.

That layout is the practical answer to a reader arriving from a search result. The project name is 螺旋桨, the bio-computing toolset, and the repository is a container for a dozen research efforts rather than one installable tool, which is why the documentation is split into a user-facing installation guide, a developer guide, and rendered docs. The Python package itself is `pahelix/`, and the `setup.py` at the root builds it.

The build script needs CMake and its comments describe another era

`setup.py` defines a `CMakeExtension` that takes a source directory instead of a list of source files, and a `CMakeBuild` command that runs `cmake --version` before building, raising a runtime error reading CMake missing, probably upgrade to a newer version of Pip. The extension output directory has `pahelix/toolkit/` appended to it, which the comment describes as required for auto-detection of auxiliary native libraries, and the configuration is Debug or Release depending on the build flag. A `CMAKE_GENERATOR` environment variable is honoured, with a note that Conda-Build can set it, and a table maps four Windows platform tags to CMake architectures: `win32` to Win32, `win-amd64` to x64, `win-arm32` to ARM, and `win-arm64` to ARM64.

The comments have drifted. One says that since CMake is forced via PEP 518 in the pyproject.toml this should never happen and the whole method could be removed, yet no pyproject.toml appears in the repository listing, which lists `setup.py` and no project file. Another explains that super() is avoided to support Python 2 because distutils classes are old-style, in a file that imports from setuptools and uses f-string free modern packaging.

One leaderboard rank and a wall clock figure are the two hard numbers

Most entries claim improvement without a figure on the page. Two do not. The 2022.08.11 HelixGEM-2 entry states that the model ranked first on the OGB PCQM4Mv2 leaderboard, with the leaderboard linked, which is the only third-party ranking claim in the log. The 2022.07.18 HelixFold entry gives wall clock numbers instead: the complete training time was optimised from 11 days to 5.12 days, and ultra-long monomer proteins of around 6600 amino acids became supported, with the paper and the code both linked.

Everything else in the timeline is a venue. The HelixFold-Single method is listed as accepted by Nature Machine Intelligence on 2023.10.09, the motif-aware RNA paper on 2024.05.13, HelixMO as accepted at BIBM 2022, and the multimodal protein-protein interaction paper as accepted at MLCB 2021, with docking-based virtual screening at BIBM 2021 and property-aware relation networks at NeurIPS 2021. Two announcements also describe running services rather than code: drug design for the HelixMO work, and online structure prediction for proteins and complexes.

The licence field resolves to nothing and a licence file exists

The licence recorded for this repository is NOASSERTION, which is the absence of a value rather than a licence. A `LICENSE` file is present at the root, so the metadata and the tree disagree about whether terms were detected, and nothing on the page says which applies to the code and which to the model parameters. Combined with the non-commercial wording attached to the HelixFold3 release and the paid API for commercial use three months later, the terms are worth reading directly rather than assuming a single open-source grant.

The archiving path is the other visible convention. A Zenodo badge carries DOI 314704349, which is how a release of this collection is meant to be cited when the repository itself does not tell you. The page also points at a project site under the paddlehelix.baidu.com domain for online structure prediction, drug design, and complex prediction services, several of which are announced alongside the papers.

Editorial conclusion

PaddleHelix is a reference collection rather than a product, and the value is in the individual papers and their code paths: a monomer folder, a complex folder, a docking folder, and the compound pre-training work are all separable pieces you can read on their own. Three things to check first. Pick a paper rather than a version, because the tags stopped at v1.2.2 in August 2023 while the announcements continue to July 2025 and the last recorded push is 2026-03-31. Read the licence question carefully, because the metadata resolves to no assertion while a licence file exists at the root, and the HelixFold3 release is described as open source for non-commercial academic research with commercial use routed to a paid API. And expect to build from source with CMake if you want the native toolkit, since the page carries no install command of its own and points at a separate installation guide instead.

Frequently asked questions

Official sources

1. Issues
2. PaddlePaddle/PaddleHelix on GitHub
3. README
4. Releases