Protein Language Models vs MSA-Based Predictors: Which to Trust for Your Target

AlphaFold gives your orphan protein a soft, low-confidence fold after searching a shallow alignment. ESMFold returns a compact model in seconds with pLDDT above 80. The structures disagree in the first 120 residues—the exact region you need for construct design. Which model deserves the next month of bench work?

The tempting answer is to rank model names. The useful answer is to diagnose the information available for this target.

Multiple-sequence-alignment (MSA) methods retrieve evolutionary evidence explicitly from a target’s homologs. Protein language models (pLMs) compress statistical regularities from millions or billions of unaligned sequences into model parameters and can infer from one query sequence. Deep, diverse homolog families usually favor MSA-based structure prediction. Orphan, designed, rapidly evolving, or high-throughput targets often make single-sequence methods valuable. When the two agree, confidence rises. When they disagree, the disagreement identifies the experiment you should not skip.

Key Takeaways

• MSA and pLM methods use evolutionary information differently. An MSA retrieves target-family evidence at inference time; a pLM retrieves population-level sequence patterns implicitly from its weights.
• Use MSA-based AlphaFold as the default for homolog-rich natural proteins. AlphaFold’s own analysis found accuracy drops when effective alignment depth falls below roughly 30 sequences and gains become smaller above roughly 100.
• Use single-sequence pLM predictors for speed, orphan/design triage, and an independent model family. They are not automatically more accurate just because the MSA is shallow.
• Do not compare pLDDT values across methods as if they share one calibration. Compare topology, domain boundaries, contacts, seed agreement, and—where possible—benchmark behavior for the target class.
• For variant effect and function prediction, the decision differs from structure prediction. MSA conservation and pLM embeddings capture overlapping but non-identical signals; task-specific validation matters more than architecture labels.
• Trust consensus conditionally. Agreement between methods with different information paths is strong evidence for a fold; correlated training data and structural priors mean it is not experimental proof.

Two Ways to Read Evolution

MSA-based prediction: retrieve the family

An MSA arranges homologous sequences so residues occupying corresponding positions line up. Correlated substitutions reveal constraints: if a change at one position is repeatedly paired with a compensating change elsewhere, the pair may contact or participate in the same functional network.

AlphaFold2 processes the query sequence, an MSA, and optional templates through the Evoformer and structure modules. Its MSA is not a simple vote. The network learns which patterns matter and integrates them with geometric reasoning (Jumper et al., 2021).

The information is target-specific. A bacterial enzyme with thousands of diverse homologs offers many natural experiments. A de novo protein offers none.

Protein language modeling: compress the sequence universe

A pLM is trained on large sequence collections using self-supervised objectives such as masked-residue prediction. To predict a missing amino acid, the model must learn patterns related to secondary structure, motifs, residue compatibility, domain context, and evolutionary constraint.

ESM-2 representations feed ESMFold, which maps a single sequence to an atomic model without building an MSA. This eliminates the slow homology-search step and enabled structure prediction at metagenomic scale (Lin et al., 2023). RGN2, OmegaFold, and HelixFold-Single take different architectural routes toward the same broad goal: useful structure inference from one sequence.

The information is global and implicit. The model may recognize a fold pattern learned across remote families even when your target has no usable alignment. It may also reproduce a strong family prior when the query’s local biology is exceptional.

“MSA-Free” Does Not Mean “Evolution-Free”

pLMs learn from natural sequences shaped by evolution. The distinction is not evolutionary versus non-evolutionary prediction. It is:

• explicit, target-specific retrieval from homologs at inference time;
• versus implicit, population-scale retrieval from learned parameters.

This matters for interpretation. A pLM can perform well on a shallow-MSA target because it recognizes general sequence grammar. It can also fail on a genuinely novel topology because its strongest prior points toward a familiar fold.

Similarly, a deep MSA is only useful when it is diverse, correctly aligned, and relevant. Ten thousand near-identical sequences contribute less than their raw count suggests. Effective depth (Neff) down-weights redundancy and is more informative than the number of rows.

Our guide to MSA depth, orphan proteins, and single-sequence prediction explains the depth diagnostics in detail.

Structure Prediction: Where MSA Methods Usually Win

AlphaFold’s validation showed a pronounced relationship between alignment depth and backbone accuracy: performance fell when median effective depth was below about 30 and improved less beyond about 100 sequences (Jumper et al., 2021). On ordinary natural proteins with rich families, explicit coevolution remains difficult to beat.

A later benchmark replaced or bypassed MSA construction with ESMFold and OmegaFold on CASP14-derived targets. The MSA-guided AlphaFold variants were substantially more accurate overall, particularly when improved homology retrieval generated better alignments (Liu et al., 2024). DeepMSA2 likewise showed that better genomic and metagenomic search plus MSA selection improved both tertiary- and quaternary-structure prediction (Zhang et al., 2024).

Default recommendation

For a natural protein with a diverse Neff above roughly 30 and no obvious alignment pathology:

1. run an MSA-based predictor;
2. inspect pLDDT, PAE, templates, and seed agreement;
3. use a single-sequence model as a cross-check where the decision is expensive.

Do not switch to a pLM method solely because it is faster.

The Targets Where Single-Sequence Models Earn Their Place

Orphan proteins

If sensitive searches find few credible homologs, an MSA method lacks target-family coevolution. RGN2 was developed around this failure mode and outperformed AlphaFold2 and RoseTTAFold on the authors’ orphan and designed-protein sets while reducing compute substantially (Chowdhury et al., 2022).

That result does not make every single-sequence model superior on every orphan. It makes it rational to run one.

De novo designs and large libraries

Designed sequences may have no natural family. A pLM-based fold can provide a rapid self-consistency screen across thousands of candidates. ESMFold’s speed is particularly useful here.

The warning is circularity. If one model designs a sequence and a related model judges it, shared priors can inflate confidence. Use an architecturally independent predictor and experimental sampling for finalists.

Rapidly evolving proteins and hypervariable loops

Viral proteins, antibody CDRs, and fast-evolving regions can produce noisy or misleading alignments. Single-sequence methods avoid the alignment-construction failure, though flexible loops remain intrinsically difficult.

High-throughput annotation

When the goal is fold-level triage rather than atomic decisions, pLM inference can cover orders of magnitude more sequences. ESMFold’s metagenomic atlas demonstrated the value of this scale (Lin et al., 2023).

A Target-by-Target Decision Table

How to Compare Two Conflicting Models

1. Compare the same construct

Check signal peptides, tags, unresolved tails, chain breaks, mutations, and isoforms. Many apparent model disagreements are input disagreements.

2. Split the comparison into claims

Do the models disagree on:

• secondary structure?
• fold topology?
• domain boundary?
• domain orientation?
• oligomeric interface?
• loop or active-site geometry?

A global RMSD hides the distinction.

3. Inspect evidence quality

For the MSA model:

• What is raw depth and Neff?
• Is coverage uniform across the disputed region?
• Are sequences diverse or redundant?
• Is a strong template driving the result?
• Do multiple MSA subsamples agree?

For the pLM model:

• Is confidence locally high in the disputed region?
• Do recycles/seeds or pLM architectures agree?
• Does the region resemble a common fold or a low-complexity exception?
• Is the confidence score calibrated for this method and target class?

4. Look for external constraints

Use:

• transmembrane topology;
• disulfide connectivity;
• catalytic motifs;
• cross-links;
• SAXS envelope;
• CD secondary-structure content;
• limited proteolysis boundaries;
• known oligomeric state;
• mutational or binding data.

A small amount of orthogonal evidence can eliminate an entire fold hypothesis.

5. Prefer the model that survives falsification

Do not average incompatible structures into a meaningless compromise. Design a discriminating experiment: a cysteine cross-link, two activity-preserving mutations, a protease-protection assay, SAXS, or a domain construct.

Confidence Scores Are Not Interchangeable

ESMFold and AlphaFold both report pLDDT-like confidence, but the models are trained differently and encounter different error distributions. A score of 85 from one method is not guaranteed to mean the same empirical accuracy as 85 from another.

Compare:

1. within-model ranking across seeds or constructs;
2. structural agreement between methods;
3. known calibration on similar target classes;
4. experimental constraints.

The same warning applies to predicted aligned error, pTM, ipTM, and composite rankings. A single number becomes useful only after you know which failure it was trained to detect.

Function and Variant Effects: A Different Trust Decision

Structure prediction is only one use of evolutionary information.

MSA-based conservation methods

Explicit alignments are strong when a residue’s functional constraint is visible across a well-sampled family. They are interpretable: you can see whether a site is conserved, covaries, or changes within a subfamily.

They weaken when:

• the family is shallow;
• alignment is ambiguous;
• paralogs mix different functions;
• the query contains insertions or novel domains;
• the phenotype is target-specific rather than evolutionarily general.

pLM embeddings and zero-shot variant scores

pLMs can score any sequence and often capture distributed context that a per-column conservation score misses. They support rapid variant triage without hand-curated MSAs.

They weaken when:

• the target requires a specific assay calibration;
• the mutation affects a ligand, PTM, interface, or environmental condition absent from the sequence objective;
• the model learns family plausibility rather than the measured phenotype;
• multi-mutation epistasis lies outside the training distribution.

Practical rule

For variant effects, combine pLM plausibility, MSA conservation, structural environment, and the assay’s biological constraints. Rank candidates with a model validated on the relevant property, not whichever foundation model has the largest parameter count.

This is also why modern alternatives to FoldX and Rosetta should be compared by endpoint rather than hype; see the AI/ML alternatives guide.

Worked Example: The Orphan With Two Convincing Folds

Suppose a 230-residue bacterial protein has an MSA with eight effective sequences. AlphaFold returns a mixed α/β topology at mean pLDDT 58. ESMFold returns an all-helical bundle at mean pLDDT 82.

A weak decision picks ESMFold because 82 is larger than 58.

A stronger workflow finds:

1. the AlphaFold model changes across MSA subsamples;
2. two independent single-sequence predictors agree on the helical bundle;
3. CD shows predominantly α-helical secondary structure;
4. a limited-proteolysis fragment matches the pLM domain boundary;
5. a small helical-bundle construct expresses and remains monodisperse.

The pLM hypothesis earns trust through independent support. If CD instead showed substantial β-sheet, the high pLM score would become evidence of an overconfident family prior.

Practical Workflow

1. Run sensitive homology search and compute effective depth. Record local coverage, not only a global count.
2. Run the default MSA-based predictor. Keep all ranked models and PAE, not just the first PDB.
3. Run a single-sequence predictor when the target is shallow, unusual, high-value, or part of a large screen.
4. Compare by structural claim. Separate fold, boundaries, state, and interfaces.
5. Perturb the information source. Subsample the MSA, alter templates, and run multiple seeds.
6. Add orthogonal sequence tracks. Topology, disorder, signal peptide, cofactor motifs, PTMs, and binding sites can expose impossible models.
7. Choose a discriminating experiment. Spend wet-lab effort where models disagree, not where they already converge.

The Economics

The compute saved by skipping an MSA is valuable at a million sequences. It is negligible compared with a failed three-month construct campaign. Match the workflow to the decision cost.

Bottom Line

Trust an MSA-based predictor first when your target has a deep, diverse, well-covered family. Add a protein-language-model predictor when the alignment is shallow, the sequence is designed or unusual, throughput matters, or you need an independent information path. Do not declare a winner by comparing confidence numbers across models.

The most trustworthy result is not the most confident single model. It is the structural claim that remains stable across information sources and agrees with the biology you can measure.

How Orbion Uses Both Information Paths

Orbion does not force a false choice between language models and structure predictors. Astra models use ESM-2 embeddings augmented with physicochemical and structural context for function, experimental suitability, PTMs, binding, topology, disorder, amyloidogenicity, and stability tasks. AlphaFold2 and AlphaFold-Multimer provide atomic structure and confidence tracks. Homology search covers more than 700 million proteins, while the PAE Insight Engine surfaces domain and dynamics cues.

The result is a layered analysis: explicit homolog evidence, pLM representations, structural confidence, and target-specific experimental planning in one record.

References

1. Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021). doi:10.1038/s41586-021-03819-2
2. Lin, Z. et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 379, 1123–1130 (2023). doi:10.1126/science.ade2574
3. Chowdhury, R. et al. Single-sequence protein structure prediction using a language model and deep learning. Nature Biotechnology 40, 1617–1623 (2022). doi:10.1038/s41587-022-01432-w
4. Fang, X. et al. A method for multiple-sequence-alignment-free protein structure prediction using a protein language model. Nature Machine Intelligence 5, 1087–1096 (2023). doi:10.1038/s42256-023-00721-6
5. Liu, J. et al. Fast, sensitive detection of protein homologs using deep dense retrieval. Nature Biotechnology 42, 1759–1768 (2024). doi:10.1038/s41587-024-02353-6
6. Zhang, C. et al. Improving deep learning protein monomer and complex structure prediction using DeepMSA2 with huge metagenomics data. Nature Methods 21, 118–126 (2024). doi:10.1038/s41592-023-02130-4
7. Mirdita, M. et al. ColabFold: making protein folding accessible to all. Nature Methods 19, 679–682 (2022). doi:10.1038/s41592-022-01488-1