How Many Homologs Does AlphaFold Need? MSA Depth, Orphan Proteins, and Single-Sequence Prediction

Your model came back at pLDDT 45 across almost the entire chain. No confident core, no floating disordered tail you can point to, just a uniform wash of orange from one terminus to the other. You check the sequence again. It is real, it expresses, colleagues have a partial crystal structure. So why is AlphaFold guessing? You pull up the multiple sequence alignment and there it is: 11 sequences, most of them near-identical paralogs from the same organism. AlphaFold did not fail because your protein is disordered. It failed because it had nothing to read.

This is the single most common reason a legitimate, well-folded protein gets a bad AlphaFold model, and it is the one researchers most often misdiagnose as disorder. AlphaFold does not learn your fold from your sequence alone. It learns it from the evolutionary record of thousands of related sequences. When that record is thin, the model degrades in ways that look alarming but are actually predictable.

Key Takeaways

• AlphaFold2 accuracy depends heavily on multiple sequence alignment (MSA) depth. The model extracts coevolutionary signal from homologs, so accuracy drops sharply once the number of effective sequences (Neff) falls below roughly 30.
• Orphan proteins, de novo and designed proteins, fast-evolving viral proteins, and antibody CDR loops all get shallow MSAs and therefore low-confidence models, even when they are well-folded.
• A uniformly low pLDDT with no confident core is a signature of a shallow MSA, not disorder. Genuine disorder usually appears as low-pLDDT stretches against a confident background.
• Diagnose before you act: check how many sequences are in the MSA and estimate Neff before concluding a protein is disordered or mispredicted.
• When the MSA is thin, you have three moves: deepen it with metagenomic databases and sensitive search, use single-sequence predictors built on protein language models (like ESMFold), or fall back on templates and experimental data.
• Single-sequence predictors trade evolutionary constraints for language-model priors. They rescue orphan and designed proteins but do not match deep-MSA AlphaFold accuracy on ordinary targets.

Introduction

AlphaFold2 changed structural biology because it made accurate structure prediction routine. But "routine" hides a strong dependency that catches people off guard: the method is only as good as the evolutionary information you feed it. AlphaFold2 is, at its core, a machine for reading coevolution. Residues that touch in three dimensions tend to mutate in a correlated way over evolutionary time, and a deep, diverse alignment of homologs makes that correlation legible. Strip the alignment down and you strip out the signal.

For most human, model-organism, and well-studied bacterial proteins, this never surfaces as a problem because the databases are saturated with homologs. The trouble starts at the edges of sequence space: proteins with no known relatives, proteins humans designed from scratch, and proteins that evolve so fast the alignment scrambles. This article is about that edge. We will cover why AlphaFold needs homologs, what MSA depth and Neff actually mean, how to tell a shallow-MSA failure apart from real disorder, and what to do when the alignment is thin. We will not rehash general AlphaFold limitations or how to read pLDDT and PAE line by line, which we have covered elsewhere, and we will stay off the separate topic of sampling alternative conformational states.

Why AlphaFold Needs Homologs at All

AlphaFold2 takes two things as input: your query sequence and an MSA built by searching sequence databases for homologs. The MSA is not a formality. It is the primary source of the geometric information the network uses.

The intuition comes from decades of coevolution research. If two residues are in contact in the folded structure, a destabilizing mutation at one position is often compensated by a mutation at the other, so the two columns of the alignment vary together. Read enough homologs and these correlated columns reveal a contact map, which constrains the fold. This is the principle that powered contact prediction and the first deep-learning structure methods before AlphaFold2 (Senior et al., 2020), and AlphaFold2 inherited and refined it. Its Evoformer module operates directly on the MSA representation, passing information between the alignment and a pairwise residue map. Analyses of retrained AlphaFold2 models confirm that the network genuinely leans on MSA statistics rather than memorizing folds, which is exactly why its behavior tracks alignment quality so closely (Ahdritz et al., 2024).

The practical consequence is blunt: no homologs, no coevolution signal, degraded prediction. A single sequence, or a handful of near-identical ones, gives the Evoformer almost nothing to work with.

What "MSA Depth" and Neff Actually Mean

People use "MSA depth" loosely to mean "how many sequences are in the alignment." That raw count matters, but it overstates the useful information, because many database sequences are near-duplicates that carry no independent evolutionary evidence.

The better measure is the number of effective sequences (Neff): the count of homologs after down-weighting redundant, highly similar ones. Ten thousand nearly identical sequences from closely related strains might contribute an Neff of only a few dozen. Neff is what actually feeds coevolution analysis, and it is the quantity that correlates with prediction accuracy.

Neff is typically computed by clustering sequences at an identity threshold (commonly 62% or 80%) and weighting each by the inverse of its cluster size. Sensitive profile-based search tools like HHblits were built precisely to pull in distant homologs that naive BLAST-style search misses, raising Neff for hard targets (Remmert et al., 2011). If you only remember one number when a prediction looks bad, make it Neff, not the raw hit count.

The Accuracy-Versus-Depth Curve

The relationship between alignment depth and accuracy is not linear. It is a saturating curve with a steep drop at the low end.

The original AlphaFold2 work characterized this directly: prediction accuracy holds up well until the effective alignment depth becomes small, then falls off rapidly (Jumper et al., 2021). The practical breakpoint sits around a few dozen effective sequences. Above roughly Neff 100, adding more homologs buys you very little. Below about Neff 30, accuracy becomes unreliable and highly target-dependent. Below Neff 10, you are effectively in single-sequence territory and should treat the model as a hypothesis, not a structure.

The asymmetry is the useful part. Deepening an already-deep alignment is wasted effort. Deepening a shallow one, or switching methods, is where the gains live.

Who Gets Shallow MSAs

Shallow alignments are not random. Specific, recognizable classes of proteins are systematically underserved by sequence databases.

Orphan proteins. These are proteins with no detectable homologs in current databases, either because their family is genuinely small, because they come from an underexplored organism, or because they have diverged past the detection horizon of standard search. Taxonomically isolated organisms, extremophiles, and newly sequenced environmental genomes are rich sources of orphans. The fold may be perfectly ordinary; the evolutionary paper trail simply does not exist yet.

De novo and designed proteins. A protein designed computationally has, by construction, no natural evolutionary history. There are no homologs to align because the sequence never existed in nature. This is the purest case of a shallow MSA: depth of one. Notably, AlphaFold2 can still do surprisingly well on some designed proteins because their sequences are engineered to be highly foldable and idealized, but the coevolution machinery contributes nothing, and single-sequence methods often match or beat MSA-based ones here (Roney & Ovchinnikov, 2022).

Fast-evolving proteins. Some proteins mutate so quickly that even close relatives align poorly. Many viral surface proteins fall here, driven by immune-escape pressure. Rapid divergence both shrinks the pool of alignable homologs and fills the alignment with noise, so effective depth is low even when raw hit counts look adequate.

Antibody variable regions and CDR loops. Antibodies are a special and important case. The framework regions are conserved and align beautifully, but the complementarity-determining regions, especially CDR H3, are hypervariable by design. Standard MSAs give almost no useful coevolution signal for the very loops that determine binding, which is why AlphaFold2 struggles with CDR H3 geometry and why dedicated antibody-specific language-model methods emerged to handle it (Ruffolo et al., 2023).

Singletons and small families. Even outside these categories, any protein whose family has only a few members will land in the low-Neff regime. This includes many uncharacterized proteins from non-model organisms that dominate metagenomic datasets.

Reading the Symptoms: What a Shallow-MSA Failure Looks Like

The failure mode has a fingerprint. Once you have seen it, you stop misreading it as disorder.

• Uniformly mediocre pLDDT with no confident core. A shallow-MSA model often sits in the 40 to 70 band across most of the chain, with no sharply confident blue domain anywhere. The model found a plausible-looking fold but cannot commit to it.
• Plausible but soft topology. The secondary structure elements are roughly there, but they pack loosely and the overall shape looks generic, like the average of several folds rather than one specific one.
• High run-to-run variability. Small changes to the MSA (different databases, different search settings) swing the model noticeably. Deep-MSA predictions are stable; shallow ones are not.
• PAE that is diffusely uncertain rather than block-structured. Instead of clean confident domains separated by flexible linkers, you see a broad haze of moderate-to-high error.

Contrast this with genuine disorder, which typically shows low-pLDDT stretches set against a confident, well-structured background. Disorder is local and correctly identified; a shallow-MSA failure is global and driven by missing data. The pLDDT-as-disorder-predictor relationship that works so well for ordinary proteins (Ruff & Pappu, 2021) breaks down for low-Neff targets, because now low confidence can mean "no evolutionary data" rather than "no stable structure." Do not run a low-Neff protein through a pLDDT disorder heuristic and report the result as disorder. That is the classic mistake this whole article exists to prevent.

Fix One: Deepen and Improve the Alignment

Before you conclude the MSA is irrecoverably thin, make sure you actually searched hard for homologs. Standard pipelines miss distant relatives that better tools and bigger databases can find.

1. Use sensitive search, not just fast search. Profile-profile and HMM-based methods (HHblits and similar) detect homologs far below the identity threshold where pairwise search fails (Remmert et al., 2011). Iterative search that builds a profile and searches again pulls in a further shell of distant relatives.
2. Search metagenomic databases. Environmental sequencing has expanded known sequence space enormously. Adding metagenome-derived sequences more than tripled the number of protein families with enough homologs for accurate modeling in one landmark analysis (Ovchinnikov et al., 2017). Databases like BFD and MGnify exist for exactly this reason.
3. Use a fast, sensitive engine so bigger databases stay tractable. MMseqs2 made searching these massive databases practical at scale (Steinegger & Söding, 2017), and the ColabFold pipeline pairs MMseqs2 with environmental databases to build deeper alignments than default AlphaFold search often produces (Mirdita et al., 2022).
4. Add templates when homologous structures exist. Even a single related experimental structure gives AlphaFold2 a scaffold that partly compensates for a thin alignment. For a low-Neff query with a distant PDB relative, supplying that template is often the highest-leverage single move.

If deepening the search lifts Neff above 30, you are back in reliable territory and the prediction problem is solved. If it does not, the homologs genuinely are not there, and you move to a different class of method.

Fix Two: Single-Sequence Prediction With Protein Language Models

When there is no alignment to build, you need a method that does not require one. This is where protein language models come in.

A protein language model is trained on hundreds of millions of individual sequences to predict masked amino acids, and in doing so it internalizes statistical regularities about which residues can sit where. The key insight is that the model learns evolutionary and structural constraints implicitly, from the whole of sequence space at once, rather than explicitly from a per-query alignment (Rives et al., 2021). At inference time it needs only your single sequence.

ESMFold is the best-known structure predictor built this way. It replaces the MSA-and-Evoformer front end with the ESM-2 language model, folding directly from a single sequence at a fraction of the runtime, because it skips the search step entirely (Lin et al., 2023). Related work established that language models alone can drive competitive single-sequence prediction, and that on orphan and designed proteins these methods can outperform MSA-based ones precisely because the MSA those methods depend on is empty (Chowdhury et al., 2022).

The honest tradeoff:

Single-sequence prediction is not a free upgrade. On a normal protein with a deep alignment, MSA-based AlphaFold2 still wins. The language-model approach earns its place specifically at the low-Neff edge, where the alignment-based method has nothing to work with.

Fix Three: Know When to Stop Trusting the Model

Sometimes the right move is to lower your reliance on prediction altogether. If Neff is in the single digits, no template exists, and single-sequence methods return low confidence too, then the computational answer is genuinely uncertain and you should treat it that way.

• Do not over-interpret a low-confidence orphan model. A soft, low-pLDDT fold on a low-Neff protein is a hypothesis about topology, not a coordinate set you should design point mutations into.
• Use it for what it is still good for. Even a rough model can seed molecular replacement, suggest domain boundaries, or guide truncation choices, provided you remember its provenance.
• Let experiment lead. For a true orphan of real importance, SAXS for shape, cross-linking mass spectrometry for contacts, or a partial experimental structure will resolve what no predictor can currently guess from a family of one.

Case Study: The Orphan That Was Not Disordered

Consider a common scenario, drawn from the pattern we see repeatedly with proteins from underexplored organisms. A group is characterizing a 240-residue enzyme from a poorly sequenced marine bacterium. The AlphaFold2 model comes back with an average pLDDT of 52 and no confident domain. The reflex interpretation is intrinsic disorder, and the group nearly deprioritizes the target on that basis.

The tell is the shape of the confidence, not its magnitude. Genuine disorder would leave an ordered core somewhere. Here, the entire chain is uniformly soft, and the predicted topology looks like a generic alpha-beta fold with loose packing. That is the shallow-MSA fingerprint.

Checking the alignment confirms it: the default search returned 14 sequences, nearly all from close relatives, for an Neff in the single digits. Two moves change the picture. First, an iterative, sensitive search against metagenomic databases raises the effective depth into the dozens by pulling in distant environmental homologs that the default pipeline never saw. Second, a single-sequence language-model prediction, run in parallel, returns a compact, consistent fold that agrees with the newly deepened MSA model on the core topology. The agreement between two methods that fail in different ways is the confidence you were missing.

The enzyme was never disordered. It was under-sequenced. The fix was data and method choice, not abandoning the target. This is the everyday version of the orphan problem, and it is far more common than true disorder among proteins from the long tail of the tree of life.

Practical Tools: A Complete Decision Workflow

Here is the full procedure to run whenever an AlphaFold model comes back weak, before you conclude anything about the protein itself.

Step 1: Check the raw MSA depth

Open the alignment your pipeline produced. Count the sequences.

• Hundreds to thousands of diverse sequences: depth is not your problem. Interpret pLDDT and PAE normally and look elsewhere (real disorder, missing cofactor, wrong conformational state).
• Tens of sequences or fewer: proceed to Step 2. Depth is a live suspect.

Step 2: Estimate Neff, not just the raw count

Redundancy inflates raw counts. Estimate effective depth by clustering the alignment at 62% identity and counting clusters, or read Neff directly if your tool reports it.

• Neff > 100: alignment is deep enough. Failure is not depth-driven.
• Neff 30 to 100: probably adequate, but validate key regions experimentally or against templates.
• Neff < 30: depth is likely the cause. Continue.

Step 3: Read the confidence pattern

Distinguish the two failure fingerprints.

Low pLDDT, confident core present, low-pLDDT is localized
   -> Likely genuine disorder or a flexible region. Handle as disorder.

Low pLDDT, no confident core, uniformly soft, high run-to-run variance
   -> Likely a shallow-MSA failure. Do NOT call it disorder. Continue.

Step 4: Try to deepen the alignment

• Run sensitive, iterative, profile-based search (HHblits-style).
• Search metagenomic databases (BFD, MGnify) with a fast engine (MMseqs2 / ColabFold pipeline).
• Add any available homologous PDB structure as a template.
• Re-predict. If Neff now clears 30 and the model sharpens, you are done.

Step 5: If the alignment stays thin, switch methods

• Run a single-sequence, language-model predictor (ESMFold-class) that needs no MSA.
• Compare its fold to the deepened-MSA model. Agreement across the two raises confidence; disagreement flags the regions to treat as unresolved.

Step 6: If both stay low-confidence, escalate to experiment

• Use the model only for coarse purposes (molecular replacement, boundary guesses).
• Bring in SAXS, cross-linking mass spectrometry, or partial experimental structures for a truly important orphan.

Quick-reference decision table

Economic Analysis

The cost of misreading a shallow-MSA model is rarely the compute. AlphaFold inference is cheap. The cost is the decision you make next.

Misdiagnosing a low-Neff orphan as disordered has a specific price: you deprioritize a viable target. If that protein was a drug target, an industrial enzyme candidate, or a key player in a pathway you are studying, walking away on the basis of a data artifact can cost months of redirected effort or the loss of the program entirely. The inverse error is just as expensive. Trusting a soft, low-Neff model enough to design mutations or order constructs from it means paying for gene synthesis, expression, and assays against coordinates that were never reliable. A failed expression campaign for a mispredicted construct runs into weeks of bench time plus reagents, and the failure is uninformative because you cannot tell whether the biology or the model was wrong.

The workflow above is cheap insurance. Estimating Neff takes minutes. A deeper metagenomic search adds compute time measured in minutes to hours, not the days a wrong wet-lab decision costs. Running a single-sequence predictor in parallel is nearly free relative to a single round of synthesis. The economically rational default is to spend the small diagnostic cost every time a model looks weak, because the alternative is making a five-figure or six-figure decision on a number you misread.

There is also a portfolio effect. Teams working on non-model organisms, environmental samples, or designed proteins will hit the low-Neff regime constantly, not occasionally. For them, building the diagnose-then-deepen-then-switch workflow into standard operating procedure is not an edge case. It is the main case, and treating it as routine converts a recurring source of wrong calls into a controlled step.

The Bottom Line

The core habit is simple: when an AlphaFold model looks bad, check the alignment before you blame the protein. Low confidence has two very different causes, and only one of them is a property of your protein. The other is a property of the database, and it is fixable.

How Orbion Handles Thin Homology

Orbion runs structure prediction with AlphaFold2 for single chains and AlphaFold-Multimer for complexes, so the same physics applies inside the platform: a shallow MSA lowers structural confidence here exactly as it does anywhere else. We are direct about that, because pretending otherwise would set you up for the mistakes this article is about.

Where Orbion helps is in not leaving you stranded when the homologs are thin. Two things matter here.

First, the Characterization module does not depend on a deep alignment to tell you about your protein. The Astra model suite is built on ESM-2 (650M) single-sequence embeddings, the same class of protein language model that powers single-sequence structure prediction. That means AstraROLE (function), AstraSUIT (experimental suitability), AstraPTM (post-translational modifications), AstraUNFOLD (disorder, membrane topology, amyloidogenicity), and AstraBIND (binding sites) still produce predictions for orphan, designed, and fast-evolving proteins where MSA-dependent methods run out of signal. When AlphaFold2 gives you a low-confidence model because your protein has no relatives, you still get a functional, PTM, disorder, topology, and binding-site read from single-sequence embeddings. In the Stabilize module, AstraDDG and AstraDTM work from the same embedding foundation for variant analysis. This is the practical answer to the orphan problem: the parts of characterization that do not require coevolution keep working.

Second, Orbion's homology search runs over 700M+ proteins (UniProt plus PDB) inside our own infrastructure, which helps surface the distant relatives that a default search would miss and gives you an honest read on how populated your protein's neighborhood actually is. Seeing that your query has only a handful of near-identical hits is itself the diagnosis this article tells you to make.

To be clear about scope: Orbion's in-product structure prediction is MSA-based AlphaFold2, so it inherits the depth dependence described above. We do not run single-sequence structure prediction as a product feature. What we do is make sure that when the alignment is thin, you still get functional and biophysical characterization from single-sequence models, plus a clear view of your homology situation, so a low-confidence structure does not become a dead end.

Ready to see what your orphan protein looks like when function does not wait on homologs? Upload a sequence to Orbion and get characterization in minutes.

References

1. Jumper J, et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596:583-589. DOI: 10.1038/s41586-021-03819-2

2. Lin Z, et al. (2023). Evolutionary-scale prediction of atomic-level protein structure with a language model. Science, 379:1123-1130. DOI: 10.1126/science.ade2574

3. Rives A, et al. (2021). Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. PNAS, 118(15):e2016239118. DOI: 10.1073/pnas.2016239118

4. Chowdhury R, et al. (2022). Single-sequence protein structure prediction using a language model and deep learning. Nature Biotechnology, 40:1617-1623. DOI: 10.1038/s41587-022-01432-w

5. Roney JP & Ovchinnikov S. (2022). State-of-the-Art Estimation of Protein Model Accuracy Using AlphaFold. Physical Review Letters, 129:238101. DOI: 10.1103/PhysRevLett.129.238101

6. Ruffolo JA, et al. (2023). Fast, accurate antibody structure prediction from deep learning on massive set of natural antibodies. Nature Communications, 14:2389. DOI: 10.1038/s41467-023-38063-x

7. Ovchinnikov S, et al. (2017). Protein structure determination using metagenome sequence data. Science, 355(6322):294-298. DOI: 10.1126/science.aah4043

8. Mirdita M, et al. (2022). ColabFold: making protein folding accessible to all. Nature Methods, 19:679-682. DOI: 10.1038/s41592-022-01488-1

9. Steinegger M & Söding J. (2017). MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nature Biotechnology, 35:1026-1028. DOI: 10.1038/nbt.3988

10. Remmert M, et al. (2011). HHblits: lightning-fast iterative protein sequence searching by HMM-HMM alignment. Nature Methods, 9:173-175. DOI: 10.1038/nmeth.1818

11. Senior AW, et al. (2020). Improved protein structure prediction using potentials from deep learning. Nature, 577:706-710. DOI: 10.1038/s41586-019-1923-7

12. Ahdritz G, et al. (2024). OpenFold: retraining AlphaFold2 yields new insights into its learning mechanisms and capacity for generalization. Nature Methods, 21:1514-1524. DOI: 10.1038/s41592-024-02272-z

13. Ruff KM & Pappu RV. (2021). AlphaFold and Implications for Intrinsically Disordered Proteins. Journal of Molecular Biology, 433(20):167208. DOI: 10.1016/j.jmb.2021.167208

Related entities

Investors: [[../entities/insight|Insight]]