Beyond AlphaFold2: Choosing Between AlphaFold3, ESMFold, RoseTTAFold2, and Boltz

AlphaFold2 solved single-chain structure prediction well enough that most labs stopped treating it as a research problem and started treating it as infrastructure. But the field did not stop there. In the years since, four predictors have reshaped what "run a structure prediction" even means: AlphaFold3, ESMFold, RoseTTAFold2, and the fully open Boltz. Each makes a different trade-off between accuracy, speed, and what kinds of molecules it can model, and picking the wrong one wastes either your afternoon or your GPU budget.

Why AlphaFold2 Was Not the End of the Story

AlphaFold2 was built and trained almost entirely around single protein chains, using a multiple sequence alignment (MSA) as its primary source of evolutionary signal. That architecture has two real limits. First, building an MSA means searching enormous sequence databases, which can take minutes to hours per target and fails badly for proteins with few known relatives — orphan genes, viral proteins, or anything from an underrepresented lineage. Second, AlphaFold2's native output is a single protein chain; predicting a protein bound to DNA, RNA, a small-molecule ligand, or a metal ion required bolt-on extensions like AlphaFold-Multimer, which were never as accurate as the core model.

The four tools below each attack one of those limits directly: AlphaFold3 generalizes the architecture to any biomolecular complex, ESMFold removes the MSA step entirely, RoseTTAFold2 offers an open three-track alternative architecture, and Boltz reproduces AF3-class capability under a fully open license.

AlphaFold3: Diffusion-Based Prediction for Any Complex

AlphaFold3, from Google DeepMind and Isomorphic Labs, replaces AlphaFold2's structure module with a diffusion model. Instead of directly regressing atomic coordinates, AF3 starts from random noise and iteratively denoises it into a 3D structure, conditioned on the same kind of Evoformer-style pairwise and MSA representations that powered AF2. This generative approach is what lets AF3 model essentially any combination of standard and modified amino acids, nucleic acids, ligands, and ions in a single unified framework, rather than needing a separate specialized model for each molecule type.

In practice this means you can hand AF3 a protein sequence plus a ligand SMILES string plus an RNA sequence and get back a jointly folded and docked complex, something that used to require chaining together a structure predictor and a separate docking tool. Reported accuracy improvements are largest for protein-ligand and protein-nucleic acid interfaces, where AlphaFold-Multimer and generic docking pipelines struggled.

The catch is access. AF3's full training code and weights are not distributed as freely as AlphaFold2's; DeepMind provides the AlphaFold Server for online, size- and rate-limited, largely non-commercial predictions, with source code and weights available separately under a restrictive research license for those who want to run it locally. Budget accordingly if your lab needs high-throughput or commercial use.

ESMFold: MSA-Free Prediction from a Protein Language Model

ESMFold, developed by Meta FAIR and distributed in the facebookresearch/esm repository, takes the opposite approach to speed. Instead of an MSA, it feeds a single sequence through ESM-2, a large protein language model trained on hundreds of millions of sequences, and reads structural information directly out of the model's internal representations. There is no database search step at all.

That single design choice changes the economics of structure prediction completely. A prediction that takes AlphaFold2 several minutes of MSA search plus inference can finish on ESMFold in a few seconds, because you skip the slowest part of the pipeline entirely. That speed is exactly why ESMFold was used to fold and publicly release predicted structures for the ESM Metagenomic Atlas, covering hundreds of millions of largely uncharacterized metagenomic protein sequences — a scale that would be impractical with MSA-based methods.

The trade-off is accuracy, particularly for sequences with few evolutionary relatives. Co-evolutionary signal from an MSA is genuinely informative, and a language model trained on single sequences is approximating what that signal would tell you rather than measuring it directly. For well-characterized protein families ESMFold is competitive; for orphan sequences or proteins in sparsely sampled lineages, expect its confidence (pLDDT) to drop and its accuracy to trail AlphaFold2 and AlphaFold3 more noticeably.

RoseTTAFold2: The Open Three-Track Architecture

RoseTTAFold2, from the Baker Lab at the University of Washington's Institute for Protein Design, is built around a three-track architecture that processes 1D sequence information, 2D residue-pair (distance and orientation) information, and 3D structural coordinates simultaneously, with information flowing between all three tracks at every layer. That simultaneous exchange, rather than a strictly sequential pipeline, is the central architectural idea: the model can use partial 3D structural hypotheses to refine its 2D and 1D representations mid-inference, not just at the end.

RoseTTAFold2 extends the original RoseTTAFold to handle multi-chain complexes more capably and was one of the first openly released tools to approach AlphaFold-Multimer-level performance on protein-protein interfaces. It is fully open source, with code and trained weights available for local installation, which made it foundational infrastructure for the broader Baker Lab ecosystem, including RFdiffusion for de novo protein design. If your work sits closer to protein engineering than pure structure lookup, RoseTTAFold2's tight integration with those downstream design tools is a real advantage over the other three predictors here.

Boltz: A Fully Open Alternative to AlphaFold3

Boltz, developed at MIT and distributed at jwohlwend/boltz, was built explicitly to close the openness gap left by AlphaFold3. It targets the same problem space — joint structure prediction of proteins, nucleic acids, and small-molecule ligands in a single complex — using a broadly similar diffusion-based approach, but ships its full training code and model weights under a permissive open-source license from day one. Boltz-2 extended the original release with binding affinity prediction, letting you estimate not just the pose of a bound ligand but roughly how tightly it binds, in the same run that generates the structure.

For labs that need to run large batches of complex predictions, retrain or fine-tune on custom data, or simply cannot accept AlphaFold3's licensing terms, Boltz is currently the most credible fully open substitute. Accuracy on standard complex benchmarks is close to, though generally still slightly behind, AlphaFold3, and the gap has been narrowing release over release.

Choosing the Right Tool

There is no single best predictor; the right choice depends on what you are modeling and what you can afford to wait for.

A few practical rules of thumb follow from that table:

• Predicting one well-studied protein chain, and you need it fast: start with ESMFold. If the pLDDT confidence looks poor, escalate to AlphaFold2 or AlphaFold3.
• Predicting a protein-ligand or protein-nucleic acid complex for structure-based drug design: use AlphaFold3 if licensing allows, otherwise Boltz. Either output is a reasonable starting pose for downstream molecular docking with AutoDock Vina refinement.
• Screening thousands of sequences, or working with metagenomic or environmental samples with no close homologs: ESMFold's MSA-free design is the only option here that scales economically.
• Building a local, fully open pipeline you can modify, retrain, or run offline without licensing concerns: RoseTTAFold2 or Boltz, depending on whether protein design (RFdiffusion) or complex/affinity prediction is the priority.
• Doing protein engineering or de novo design work: RoseTTAFold2's ecosystem integration with RFdiffusion and ProteinMPNN makes it the natural open-source hub.

Practical Takeaways

None of these four tools makes AlphaFold2 obsolete for the case it was built for — single, well-characterized protein chains with deep MSAs — where AF2 remains fast, accurate, and thoroughly validated. What has changed is that structure prediction is no longer one tool for one job. AlphaFold3 and Boltz have turned "predict this complex" into a routine single-model query instead of a multi-tool pipeline. ESMFold has made structure prediction cheap enough to run at database scale. And RoseTTAFold2 keeps a fully open, actively developed architecture at the center of the protein design ecosystem.

Whichever predictor you choose, treat the output the same way you would treat AlphaFold2's: check the confidence metrics before trusting the model, and remember that a predicted structure is a hypothesis, not a measurement. If you are new to the underlying architecture these tools all build on, our earlier piece on the AlphaFold revolution in protein structure prediction is a good primer on the Evoformer and pLDDT confidence scores that carry over into every model discussed here.