Two AI systems now dominate the protein structure prediction landscape: DeepMind's AlphaFold 3 and the University of Washington's RoseTTAFold All-Atom. Both represent major scientific achievements. Both have real limitations. And — as most leading computational drug discovery teams have concluded — both are worth using, for different reasons.
Architectural Differences That Produce Different Results
AlphaFold 3 uses a diffusion-based architecture, generating atomic coordinates by iteratively refining a probabilistic distribution. This approach, borrowed from the image generation field, is well-suited to predicting the binding poses of standard drug-like molecules — the core use case in pharmaceutical discovery.
RoseTTAFold All-Atom, published in Science in September 2024, extends the original RoseTTAFold transformer architecture to include explicit all-atom representation of small molecules, DNA, RNA, metal ions, and cofactors. The explicit atomic representation means the model can more directly reason about the chemistry of unusual binding interactions — particularly covalent bonds and metal coordination.
These architectural differences are not just academic. They produce meaningfully different performance profiles on different types of problems.
- AlphaFold 3: Diffusion-based; stronger on standard drug-like small molecule docking
- RoseTTAFold All-Atom: Explicit atomic representation; stronger on covalent inhibitors and metalloprotein targets
- RoseTTAFold is fully open-source with permissive academic licence
- AlphaFold 3 commercial pharmaceutical use requires Isomorphic Labs agreement
- Published in Nature (AlphaFold 3, May 2024) and Science (RoseTTAFold All-Atom, Sep 2024)
Where AlphaFold 3 Leads
For the most common drug discovery use case — predicting how a standard drug-like small molecule (roughly Lipinski Rule of Five compliant) binds to a protein target — AlphaFold 3 has demonstrated stronger performance in the benchmarks presented in DeepMind's Nature publication. The diffusion-based approach produces highly plausible binding geometries for this class of molecules, and the results have been sufficient for many pharmaceutical teams to use predicted structures in place of experimentally determined ones during early discovery stages.
Where RoseTTAFold All-Atom Takes the Lead
Covalent inhibitors — drugs that form permanent chemical bonds with their targets — are a growing drug discovery strategy for previously undruggable targets. Predicting covalent binding correctly requires explicit representation of the bond-forming chemistry, which is where RoseTTAFold All-Atom's explicit atomic representation provides an architectural advantage over AlphaFold 3's diffusion approach.
Metalloprotein targets — proteins that require metal ions for their function, including zinc-containing proteins targeted by many antibiotic and antifungal agents — similarly benefit from RoseTTAFold's explicit metal coordinate representation.
The Open-Source Dimension
RoseTTAFold All-Atom is fully open-source, with model weights freely available under a permissive academic licence from the Baker Lab at the University of Washington. AlphaFold 3's research access is available via the AlphaFold Server, but commercial pharmaceutical applications require licensing from Isomorphic Labs.
For large pharmaceutical companies, the licensing cost is not a meaningful barrier. For academic laboratories, smaller biotechs, and drug discovery efforts in lower-resource settings, RoseTTAFold's open availability is a significant practical advantage — and is driving continued investment in RoseTTAFold development as an open alternative.
The Practical Verdict
Most computational drug discovery teams that have access to both systems use both — AlphaFold 3 as the primary tool for standard small-molecule docking campaigns, RoseTTAFold All-Atom for covalent chemistry, metalloprotein targets, and cases where open access is operationally necessary. Running both on key targets and cross-validating results adds confidence that cannot be obtained from either system alone. The computational cost of doing so is modest relative to downstream experimental validation costs.