PaddleHelix: A PaddlePaddle Toolkit for Protein Folding, Docking and Molecular Property Prediction
Bio-Computing Platform Featuring Large-Scale Representation Learning and Multi-Task Deep Learning “螺旋桨”生物计算工具集
At a glance
- What is it?
- PaddleHelix bundles several research models (HelixFold, HelixFold-Single, HelixDock, HelixGEM-2, HelixADMET) under one repository. The open code is real, but the newest models arrive as research snapshots and the most capable ones are also offered as paid or hosted services.
- Who is it for?
- PaddleHelix suits teams already running PaddlePaddle, or groups that want HelixFold-Single's MSA-free inference path and are willing to work from the repository's per-app instructions. It is the wrong choice if you need a single stable API across all models, or if you cannot accept that the newest HelixFold3 weights are released for non-commercial academic research only.
- Can I use it commercially?
- Check first. The repository uses a licence we do not classify automatically, so read its LICENSE file before any commercial use.
- Is it still maintained?
- Yes. The repository last received commits 169 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What PaddleHelix Actually Contains
PaddleHelix is not one model. It is a collection of applications under an apps/ directory, each with its own code, and in several cases its own paper. The topics list names the scope: drug-target interaction, drug-drug interaction, protein-protein interaction, protein folding, docking, RNA structure prediction, molecular design and representation learning. The README's news entries map those topics to named projects: HelixFold and HelixFold-Single for protein structure, HelixFold-Multimer for complexes, HelixFold3 and HelixFold3.2 for biomolecular structure prediction, HelixDock for protein-ligand docking, HelixGEM and HelixGEM-2 for molecular property prediction, HelixADMET for ADMET endpoints, BatchDTA for drug-target affinity, HelixMO for molecular optimization, and RNAErnie for RNA language modeling.
The audience is computational biology and cheminformatics engineers who work in Python and are comfortable reading a paper alongside its code. The repository is a research monorepo, not a product. There is no unified CLI and no single model registry. If you want to predict a protein structure, you go to apps/protein_folding/helixfold or apps/protein_folding/helixfold-single. If you want docking, you go to apps/molecular_docking/helixdock. Each directory carries its own setup narrative. That structure is the central fact about adopting PaddleHelix: you are adopting several projects that happen to share a GitHub organisation and a deep learning framework.
The Mechanism: PaddlePaddle Graphs Plus Per-App Training Pipelines
Every model here is built on PaddlePaddle, Baidu's deep learning framework, and the code is Python. The README states that HelixFold's inference pipeline reproduces AlphaFold 2 using PaddlePaddle, and that the full HelixFold release includes both training and inference pipelines. It also states that training time was optimised from 11 days to 5.12 days, and that ultra-long monomer proteins of around 6600 amino acids are supported. Those are the project's own claims, not measurements I can reproduce here.
The architectural split matters more than any single number. HelixFold follows the AlphaFold 2 pattern: multiple sequence alignments feed an attention-based structure module. HelixFold-Single removes the MSA step entirely, relying on a protein language model and the primary sequence alone; the README describes it as an MSA-free pipeline that can predict structures within seconds. That is a different data flow, not a speed setting. Without an MSA you lose the evolutionary signal that AlphaFold 2 depends on, and the project positions the trade-off as speed and independence from alignment databases.
HelixDock is described as a pre-training model on large-scale generated docking conformations, aimed at protein-ligand structure prediction. HelixGEM-2 is described as a molecular property prediction network that models full-range many-body interactions, and the README notes it ranked first on the OGB PCQM4Mv2 leaderboard at the time of its release. HelixADMET is described as endpoint extensible, incorporating self-supervised knowledge transfer. In each case the shared mechanism is representation learning: pretrain on large unlabelled or generated data, then fine-tune or attach a task head.
Getting It Running: Per-App Setup, Not a Single Install
There is no top-level pip install for the whole toolkit in the material provided. Each application directory carries its own instructions, and the README links to them by path. For HelixFold3 the README points to apps/protein_folding/helixfold3 and states that the code and model parameters were released on 2024.08.15, with the initial release available as open source on GitHub for non-commercial academic research. For HelixDock the pointer is apps/molecular_docking/helixdock. For HelixGEM-2 it is apps/pretrained_compound/ChemRL/GEM-2, and for the earlier GEM it is apps/pretrained_compound/ChemRL/GEM. BatchDTA lives at apps/drug_target_interaction/batchdta.
The badge in the README lists Python 3.6+ and Linux, Windows and macOS as supported. PaddlePaddle itself must be installed before any of these apps will import; the version constraint is stated per app rather than globally, so read the directory you intend to use. The practical sequence is: clone the repository, enter the app directory, follow that directory's README, install PaddlePaddle as it specifies, then download the model parameters the app expects. Because the default branch is dev, a plain clone gives you the development branch; if you want the released state you should check out the tag. The most recent tagged release listed is v1.2.2 from 2023-08-01, while the HelixFold3 news entry is dated 2024.08.15 and HelixFold3.2 is dated 2025.07.23. Code in dev is therefore ahead of the latest tag by more than two years of work, and that gap is worth understanding before you pin anything.
Where PaddleHelix Is the Wrong Tool
The licence is the first constraint. The repository metadata reports NOASSERTION, and the README states plainly that the initial HelixFold3 release is available as open source for non-commercial academic research. Those two facts together mean you should not assume a permissive licence across the whole repository. Different apps may carry different terms. A commercial team that wants HelixFold3 has a documented alternative: the README describes a paid API introduced on 2024.11.08 for academic and commercial applications, with a usage guide on the PaddleHelix site. That is a deliberate split between open research code and commercial access, and it is the single most important thing to check before planning a product around this repository.
The second constraint is consistency. Because each app is a separate project, there is no guarantee that HelixFold, HelixDock and HelixGEM-2 share a PaddlePaddle version, a data format or a checkpoint convention. Upgrading PaddlePaddle for one app can break another. There is no compatibility matrix in the material provided, and I cannot confirm one exists.
The third constraint is deployment weight. These are research training and inference pipelines. The README frames HelixFold's training optimisation in days, which tells you the intended scale. If you need a small, dependency-light predictor for a web service, an MSA-free model such as HelixFold-Single is the closest fit in this repository, but it is still a PaddlePaddle model with its own environment.
Alternatives and the Real Difference in Approach
The obvious comparison for the folding apps is AlphaFold 2 and its reimplementations. The README itself says HelixFold is a reproduction of the AlphaFold 2 inference pipeline using PaddlePaddle, so the algorithmic lineage is not in question. The difference is the framework and the surrounding ecosystem. If your team already runs PyTorch, OpenFold (the PyTorch reimplementation of AlphaFold 2) is the more natural fit, because you avoid introducing a second deep learning framework and its CUDA and driver constraints into your stack. If you are standardised on PaddlePaddle, or you are working in an environment where PaddlePaddle is already supported, HelixFold removes that friction.
For MSA-free prediction specifically, HelixFold-Single occupies a narrower niche. Its stated advantage is that it needs only the primary sequence, which means no alignment database to host or query. A full AlphaFold-style pipeline requires an MSA, and building one is often the slowest part of the workflow. If your bottleneck is database infrastructure rather than GPU time, HelixFold-Single changes the shape of the problem rather than just the speed.
For molecular property prediction, HelixGEM-2 competes with graph neural network libraries and with pretrained chemistry models from other frameworks. The README's claim of a first-place OGB PCQM4Mv2 result is a point-in-time benchmark result, and leaderboard positions move. Treat it as evidence that the approach was competitive when published, not as a current ranking. The real differentiator for HelixGEM-2 is that it is packaged with a training pipeline inside this repository, so you can fine-tune rather than only call an inference endpoint.
Maintenance, Releases and What the Timeline Says
The release history is thin. The listed releases are v1.0 (2021-07-09), v1.1.0 (2021-12-15) and v1.2.2 (2023-08-01). The last push to the repository is dated 2026-03-31, so the project is active even though tagged releases stopped in 2023. The practical consequence is that new work lands on dev and is announced in the README's news section rather than through a version bump. Anyone who pins to v1.2.2 will not have HelixFold3, HelixDock or HelixFold3.2, because those arrived after that tag. Anyone who tracks dev accepts that the branch can change under them.
That is a real maintenance cost, and it is not hidden: the README is organised around dated news entries, which is the release mechanism in practice. For a team, this means budgeting for periodic rebases against dev, or forking a known-good commit and treating it as your own release. There is no changelog file described in the material, so the news entries are the closest thing to one.
On licensing, the metadata's NOASSERTION value is not a licence. It means GitHub could not classify the terms. The README's non-commercial statement about HelixFold3 is the clearest signal available, and it applies to that model specifically. Read the LICENSE file in the repository root and any per-app licence before you ship anything. This is a factual observation about the repository, not legal advice.
A Note on What the Repository Does Not Tell You
The README is a news feed with links, not a reference manual. It does not state hardware requirements for inference, expected runtime per prediction, memory ceilings for the 6600-amino-acid monomer case, or the size of the model parameter downloads. It does not list a compatibility matrix between apps and PaddlePaddle versions. It does not describe a deprecation policy for the older models still present in the tree, such as the original GEM alongside GEM-2, or the original HelixFold alongside HelixFold3 and HelixFold3.2.
That absence is informative. PaddleHelix is maintained as a research publication vehicle: a paper is accepted, code and weights are pushed, a news entry is added. The repository accumulates models rather than consolidating them. If you need an operational service with an SLA, the README points to the hosted PaddleHelix website and the paid HelixFold3 API, which is where the operational guarantees live. The GitHub repository is the research surface. Planning around that distinction is the difference between a smooth integration and a surprise.
Editorial conclusion
PaddleHelix suits teams already running PaddlePaddle, or groups that want HelixFold-Single's MSA-free inference path and are willing to work from the repository's per-app instructions. It is the wrong choice if you need a single stable API across all models, or if you cannot accept that the newest HelixFold3 weights are released for non-commercial academic research only. Before committing, read the LICENSE file at the repository root, confirm the exact PaddlePaddle version each app directory expects, and check whether the model you need is open code or only reachable through the paid HelixFold3 API.
Community notes