The Post-AlphaFold Model Zoo: MSA-Free Predictors and When to Trust Them

“MSA-free” now covers everything from fast single-chain language-model folders to diffusion systems that co-model proteins, nucleic acids, and ligands. They do not solve the same problem, use the same evidence, or deserve the same kind of trust.

The useful way to navigate the model zoo is not to memorize a leaderboard. Classify the target, the claim, the information available, and the consequence of being wrong.

Key Takeaways

• MSA-free means no target-specific alignment at inference time. It does not mean the model learned without evolutionary data.
• Single-sequence predictors are strongest for speed, scale, orphan targets, and designed sequences. MSA-based methods often remain more accurate on homolog-rich natural proteins.
• Single-chain and complex predictors are different categories. A correct monomer fold does not validate an interface or ligand pose.
• Public availability is multidimensional. A paper, server, weights, code, license, and reproducible environment are not equivalent.
• Confidence scores are model-specific. Calibrate on the target class and compare structural claims rather than raw numbers.
• Trust grows through perturbation and orthogonal evidence. Multiple models help most when their errors are genuinely different.

First Define “MSA-Free”

An MSA-based predictor retrieves homologous sequences for each target and extracts family-specific covariation. An MSA-free predictor accepts the query sequence without building that alignment.

Most MSA-free systems still learn from large collections of natural protein sequences and structures. Evolutionary information is embedded in model parameters rather than supplied as a per-target table. “Single-sequence” is usually the clearer term.

The distinction matters operationally:

• no homology-search bottleneck;
• faster inference across large libraries;
• no alignment-construction failure;
• less target-specific covariation;
• dependence on learned priors from training data.

Category 1: Language-Model Single-Chain Folders

ESMFold

ESMFold uses ESM-2 representations to predict atomic structures directly from a single sequence. Its speed enabled a metagenomic structure atlas at enormous scale (Lin et al., 2023).

Best fit:

• atlas-scale annotation;
• orphan and designed proteins;
• rapid independent fold check;
• triage before deeper analysis.

Main caution:

• overall accuracy often trails strong MSA-based workflows on ordinary homolog-rich targets;
• high local confidence does not solve state, assembly, or expression behavior.

RGN2

RGN2 combines a protein language model with a geometric network. It was explicitly evaluated on orphan and designed proteins and performed strongly in those regimes (Chowdhury et al., 2022).

Best fit:

• true singletons;
• designed proteins;
• an architecture independent from AlphaFold-style MSA processing.

Main caution:

• benchmark wins on specialized sets do not imply universal superiority.

OmegaFold and HelixFold-Single

These systems take different routes to single-sequence structure prediction, using pretrained sequence representations and geometry-aware networks. They expand useful diversity in the model family.

Best fit:

• cross-model comparison on shallow-MSA targets;
• high-throughput structure generation;
• identifying regions where single-sequence architectures disagree.

Main caution:

• availability, maintained implementations, supported sequence lengths, and confidence calibration may differ by release.

Category 2: MSA-Free Complex and Biomolecular Models

Newer systems increasingly predict complexes containing proteins, nucleic acids, ions, and small molecules. Some use diffusion or generative structure modules and can operate without a conventional target MSA.

This is not simply “ESMFold with more chains.” Complex prediction adds:

• stoichiometry;
• interface orientation;
• ligand identity and chemistry;
• covalent modifications;
• competition among binding sites;
• conformational state;
• ranking across alternative assemblies.

AlphaFold 3 demonstrated major gains across biomolecular interactions using a diffusion-based architecture (Abramson et al., 2024). Other publicly described systems, including Chai-1 and Boltz releases, broaden access to complex prediction. Evaluate the exact public release you can run; names can hide material differences in weights, inference options, licenses, and training data.

Category 3: Structure-Conditioned Sequence Models

ProteinMPNN and related inverse-folding systems are often mentioned beside structure predictors, but they answer the reverse question: given a backbone, which sequence fits it? They belong in the post-AlphaFold ecosystem while serving a different decision.

Do not compare their sequence scores with pLDDT. Use them for design, then validate whether the designed sequence recovers the intended structure. See ProteinMPNN and Inverse Folding, Explained for Bench Scientists.

A Practical Model Map

When MSA-Free Predictors Deserve Priority

True orphan proteins

If sensitive searches find few diverse homologs, explicit covariation is weak. Run several single-sequence predictors and compare them with an MSA-based attempt.

De novo designs

Natural MSAs do not exist. Fast single-sequence folding supports library-scale triage. Guard against circularity when related learned priors generate and validate the design.

Massive sequence collections

At millions of sequences, MSA search becomes a practical bottleneck. A rapid model can identify fold families and prioritize unusual targets for slower analysis.

Alignment pathologies

Domain shuffling, paralog mixing, low complexity, and uneven coverage can corrupt an MSA. Single-sequence models provide a useful independent path.

Fast design cycles

When sequence changes daily, rapid structure feedback may be more valuable than maximal accuracy on every candidate.

When the MSA Still Wins

A deep, diverse, correctly aligned family contains target-specific evolutionary experiments. For natural proteins with strong coverage, MSA-based AlphaFold remains a strong default.

Prefer or retain MSA-based analysis when:

• Neff is substantial across the relevant domain;
• long-range contacts are supported by covariation;
• remote homologs add meaningful diversity;
• complex partners can be paired credibly;
• the decision needs the strongest available single-chain accuracy;
• single-sequence models disagree in the core topology.

The important qualifier is effective depth. Thousands of nearly identical sequences do not provide thousands of independent constraints.

How to Compare Outputs Without Fooling Yourself

Compare identical inputs

Use the same isoform, residue range, mutations, chain definitions, and ligand state. Many “model disagreements” are actually prompt disagreements.

Compare claims, not global RMSD

Break the structure into:

• secondary structure;
• fold topology;
• domain boundaries;
• domain orientation;
• loops and active site;
• oligomeric interface;
• ligand pose.

Two models can agree on the fold and disagree completely on the question that matters.

Perturb each information source

For MSA methods, vary MSA subsampling, templates, and seeds. For single-sequence models, compare architectures and repeated samples. Stable conclusions deserve more weight.

Check confidence calibration

A pLDDT-like number is meaningful within the context of the model that produced it. Do not create a cross-model leaderboard from raw scores without calibration on similar targets.

Add external constraints

Topology, disulfides, motifs, cross-links, SAXS, CD, oligomeric mass, ligand binding, and mutational data can eliminate otherwise plausible predictions.

Public Model Does Not Mean Reproducible Model

Before choosing a tool for a long-lived workflow, record:

• Are weights downloadable?
• Is inference code available?
• Is the license compatible with the use case?
• Is a hosted server the only route?
• Are databases and model versions pinned?
• Can raw confidence outputs be exported?
• Are sequence or ligand inputs retained by the service?
• Can the run be reproduced six months later?

For regulated or proprietary work, provenance and deployment may matter as much as benchmark accuracy.

Worked Example: A Designed Enzyme Family

You have 20,000 designed 180-residue sequences built around one catalytic motif. None has natural homologs.

A rational funnel:

1. remove sequences that violate catalytic or synthesis constraints;
2. run a fast single-sequence model across all 20,000;
3. cluster by predicted fold and confidence pattern;
4. retain candidates with the intended topology and diverse sequences;
5. run two independent predictors on 500 finalists;
6. examine active-site geometry and alternative oligomers for 100;
7. apply physics and developability filters;
8. synthesize a diverse plate, not only the top near-clones.

Running a maximal MSA workflow on every candidate adds little because no natural family exists. Running one language model and selecting its top 96 also creates a single-model monoculture. The combined funnel matches compute to uncertainty.

Worked Example: A Conserved Metabolic Enzyme

A 320-residue bacterial enzyme has 15,000 diverse homologs. ESMFold and AlphaFold agree on the core but disagree on a lid domain.

Here the MSA carries direct family evidence. Inspect local alignment coverage, templates, and PAE; compare known ligand-bound homologs; generate alternative states. The single-sequence disagreement is valuable because it identifies a mobile region, not because the MSA should be discarded.

A Trust Ladder

Treat each step as stronger than the one before:

1. one model, one run;
2. repeated samples from one model;
3. agreement across model families;
4. stability to MSA/template/input perturbation;
5. agreement with sequence and physical constraints;
6. agreement with low-resolution experiments;
7. direct biochemical or structural validation.

No model name lets you skip the ladder. The level required depends on the cost of the downstream decision.

Using Public Models in Orbion

Orbion’s structure-prediction workflow can organize outputs from public model families alongside confidence, target annotations, and downstream construct decisions. The useful product is not a gallery of structures. It is a comparison that records which evidence each model used, where outputs agree, and what experiment resolves the remaining ambiguity.

For the direct information-path comparison, read Protein Language Models vs MSA-Based Predictors. For a focused ESM-2 discussion, see ESM-2 vs AlphaFold.

Bottom Line

The post-AlphaFold model zoo is useful because it offers different tradeoffs, not because every new system replaces the last. MSA-free predictors bring speed, scale, and credible alternatives for orphan and designed proteins. MSA-based systems retain powerful target-family evidence. Complex models extend the task but introduce new failure modes.

Classify the problem first. Then choose the smallest set of genuinely different models that can change the decision, compare specific structural claims, and demand experimental evidence in proportion to the cost of being wrong.

References

1. Lin Z, et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science. 2023. doi:10.1126/science.ade2574
2. Chowdhury R, et al. Single-sequence protein structure prediction using a language model and deep learning. Nature Biotechnology. 2022. doi:10.1038/s41587-022-01432-w
3. Wu R, et al. High-resolution de novo structure prediction from primary sequence. bioRxiv. 2022. doi:10.1101/2022.07.21.500999
4. Fang X, et al. HelixFold-Single: MSA-free protein structure prediction by using protein language model as an alternative. Bioinformatics. 2023. doi:10.1093/bioinformatics/btad333
5. Abramson J, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature. 2024. doi:10.1038/s41586-024-07487-w