MSA Quality & Why It Matters for Structure Prediction

Multiple Sequence Alignments (MSAs) are the evolutionary foundation upon which AlphaFold2 builds its remarkably accurate predictions. While the deep learning architecture of AlphaFold2 often captures headlines, the quality and depth of the MSA used as input is arguably the single most important determinant of prediction accuracy. Understanding what makes a good MSA, how to assess MSA quality, and what to do when MSA depth is insufficient can mean the difference between obtaining a highly accurate structural model and receiving a prediction that is essentially useless for downstream applications. This comprehensive guide will take you deep into the world of MSAs, explaining not just what they are, but why they matter so profoundly for structure prediction.

The fundamental insight that makes MSAs so powerful for structure prediction is that evolution is a powerful teacher. When a protein sequence evolves over millions or billions of years, mutations that disrupt the protein's structure or function are eliminated by natural selection, while mutations that maintain or improve function are retained. This creates patterns in the sequences of evolutionarily related proteins that encode information about structural and functional constraints. Positions that are absolutely conserved across all sequences are usually critical for structure or function—they cannot vary without breaking something essential. Conversely, positions that show correlated variation—where mutations in one position are consistently accompanied by compensatory mutations in another position—often indicate spatial proximity in the three-dimensional structure. These residue pairs need to covary to maintain favorable interactions like salt bridges, hydrogen bonds, or hydrophobic packing.

AlphaFold2 was explicitly designed to extract and utilize the evolutionary information embedded in MSAs. The Evoformer, the core neural network architecture at the heart of AlphaFold2, processes the MSA as one of its primary inputs and uses attention mechanisms to identify these patterns of conservation and coevolution. The network learns to recognize that certain patterns in the MSA correspond to specific structural features—alpha helices, beta sheets, turns, or specific types of inter-residue contacts. The more sequences present in the MSA, and the more diverse those sequences are in terms of evolutionary distance, the more information AlphaFold2 has to work with. This is why MSA depth and quality correlate so strongly with prediction accuracy, and why understanding MSAs is essential for anyone seriously using AlphaFold2 or other sequence-based structure prediction methods.

#What is a Multiple Sequence Alignment?

A multiple sequence alignment is, at its most basic level, exactly what the name suggests: an alignment of multiple protein sequences, arranged so that evolutionarily related positions are placed in the same column. The query sequence—the protein whose structure you want to predict—sits at the top of the alignment, and below it are sequences of homologous proteins from other organisms. These homologs are proteins that share a common evolutionary ancestor with your query and therefore have similar sequences and, usually, similar structures. The alignment process inserts gaps (represented by dashes) where insertions or deletions have occurred during evolution, ensuring that positions that are actually equivalent in three-dimensional space are aligned in the same column, even if the sequences have diverged significantly at the sequence level.

Creating an MSA is a computational process that typically involves several steps. First, you use your query sequence to search large sequence databases—typically comprehensive databases like UniRef90, UniRef50, or Pfam that contain millions of protein sequences from sequenced genomes across all domains of life. The search is performed using specialized algorithms like PSI-BLAST (Position-Specific Iterative BLAST), HHblits, or jackhmmer that can detect remote evolutionary relationships even when sequence similarity is low. These algorithms use position-specific scoring matrices that give different weights to different amino acid substitutions depending on what's been observed at each position in known homologs, making them much more sensitive than simple BLAST searches.

Once potential homologs have been identified, they need to be aligned to the query sequence. This alignment step is crucial because errors in alignment can mislead structure prediction algorithms by placing non-equivalent positions in the same column. The alignment algorithms use substitution matrices (like BLOSUM62) that encode the relative likelihood of different amino acid substitutions, gap penalties that discourage insertions and deletions (because these are evolutionarily rare and disruptive), and sometimes additional information like secondary structure predictions or existing domain annotations. For AlphaFold2, the default pipeline uses jackhmmer for database searching and MSA construction, running multiple iterations to build up the alignment incrementally, finding more remote homologs in each iteration based on the profile built from sequences found in previous iterations.

#Why MSA Quality Determines Prediction Accuracy

The correlation between MSA depth and AlphaFold2 prediction accuracy is one of the most robust relationships in computational structural biology. Study after study has shown that proteins with deep MSAs (hundreds to thousands of effective sequences) achieve average pLDDT scores above 90 and near-perfect structural accuracy, while proteins with shallow MSAs (dozens of effective sequences or fewer) frequently have large low-confidence regions with pLDDT scores below 70 and structural accuracy that is essentially unusable. This is not a minor effect—it's the difference between a prediction that's as good as an X-ray crystal structure and a prediction that's barely better than random. Understanding why this relationship exists requires diving into the coevolution signal that MSAs provide.

When two residues in a protein are in close spatial proximity and interact with each other—perhaps forming a salt bridge, a hydrogen bond, or hydrophobic contacts—mutations that occur at one position often need to be compensated by mutations at the other position to maintain the favorable interaction. For example, if a negatively charged glutamate at position A forms a salt bridge with a positively charged lysine at position B, and evolution replaces the glutamate with a positively charged arginine, there will be strong selective pressure to simultaneously replace the lysine with a negatively charged aspartate or glutamate to maintain the salt bridge. This correlated variation leaves a distinctive pattern in the MSA: whenever you see one type of amino acid at position A, you consistently see a particular type at position B, and when you see a different amino acid at position A, you see a corresponding different amino acid at position B.

AlphaFold2's Evoformer is specifically designed to detect these coevolution patterns. The attention mechanisms in the Evoformer analyze the MSA to identify which pairs of positions show correlated variation, and these covariation signals provide direct information about spatial proximity in the three-dimensional structure. With a deep MSA containing diverse sequences, these coevolution signals become statistically robust and reliable—you can clearly distinguish true coevolving pairs that are genuinely in contact from spurious correlations that arise by chance. With a shallow MSA, the statistical signal is weak and noisy, making it difficult to confidently identify which pairs are truly coevolving. This is why MSA depth matters so much: it's not just about having more sequences, it's about having enough data to reliably detect the evolutionary patterns that encode structural information.

The relationship between MSA depth and accuracy is not perfectly linear—there are diminishing returns as you add more and more sequences. The first 100 sequences provide enormous value, the next 900 sequences provide substantial additional benefit, and sequences beyond that provide incremental improvements. This is because the coevolution signal saturates once you have sufficient diversity to capture all the major evolutionary constraints. Additionally, adding very similar sequences (which differ by only a few mutations) provides little additional information compared to adding divergent sequences that explore different regions of sequence space. This is why "effective sequence depth" (which accounts for sequence diversity) is a better metric than raw sequence count for assessing MSA quality.

#How to Assess MSA Quality

Evaluating MSA quality before running a structure prediction can save you significant time and help you interpret the results appropriately. There are several key metrics that provide insight into MSA quality, each capturing a different aspect of what makes an MSA useful for structure prediction. Understanding these metrics and how to calculate them will help you develop intuition about which proteins are likely to have high-quality predictions and which might need additional work or alternative approaches.

Neff: Effective Number of Sequences

The most important single metric for MSA quality is Neff, the effective number of sequences. Unlike the raw sequence count, which simply counts how many sequences are in the MSA, Neff attempts to quantify the amount of independent evolutionary information present by downweighting highly similar sequences. If you have 1000 sequences but they're all 99% identical to each other, they don't provide much more information than a single sequence—they're essentially sampling the same point in evolutionary space. Neff corrects for this redundancy by clustering similar sequences and weighting each cluster by its size, effectively asking "how many truly diverse sequences do I have?"

The most common definition of Neff is based on sequence identity thresholds. For each sequence in the MSA, you count how many other sequences share greater than X% identity with it (typical thresholds are 62% or 80%), and then the weight for that sequence is 1 divided by the size of that group. The Neff is the sum of all sequence weights. This approach ensures that a cluster of 100 sequences at 99% identity contributes approximately 1 to Neff, while 100 sequences at 40% identity contribute closer to 100. A general rule of thumb is that Neff greater than 64 predicts good structure prediction accuracy, Neff between 16 and 64 predicts moderate accuracy with some uncertain regions, and Neff less than 16 predicts poor accuracy with large uncertain regions.

# Calculate Neff for an MSA
import numpy as np
from Bio import AlignIO

def calculate_neff(msa_file, threshold=0.62):
    """
    Calculate effective number of sequences (Neff) for an MSA

    Args:
        msa_file: Path to MSA in FASTA or A3M format
        threshold: Sequence identity threshold for clustering
    Returns:
        neff: Effective number of sequences
    """
    # Load MSA
    alignment = AlignIO.read(msa_file, "fasta")
    n_seqs = len(alignment)

    # Calculate pairwise sequence identities
    weights = np.ones(n_seqs)
    for i in range(n_seqs):
        seq_i = str(alignment[i].seq).replace('-', '')
        cluster_size = 0
        for j in range(n_seqs):
            seq_j = str(alignment[j].seq).replace('-', '')
            # Calculate sequence identity
            matches = sum(a == b for a, b in zip(seq_i, seq_j))
            identity = matches / min(len(seq_i), len(seq_j))
            if identity >= threshold:
                cluster_size += 1
        weights[i] = 1.0 / cluster_size

    neff = np.sum(weights)
    print(f"MSA has {n_seqs} sequences with Neff = {neff:.1f}")
    return neff

# Example usage
neff = calculate_neff("my_protein.a3m", threshold=0.62)
if neff > 64:
    print("✓ Deep MSA - expect high accuracy")
elif neff > 16:
    print("⚠ Medium MSA - expect moderate accuracy")
else:
    print("✗ Shallow MSA - expect low confidence regions")

Coverage: How Much of the Protein is Aligned

MSA coverage refers to what fraction of the query sequence is covered by the alignment—are there aligned sequences across the entire length of your protein, or are there regions where few or no homologous sequences could be found? Poor coverage can occur for several reasons: the region might be a recently evolved insertion unique to your protein (or your clade), it might be an intrinsically disordered region that evolves too quickly for homology to be detectable, or it might be a transmembrane region where the hydrophobic constraints are so strong that different sequence solutions converge on the same structure. Regardless of the reason, regions with poor MSA coverage will have poor structure predictions because AlphaFold2 has no evolutionary information to work with.

You can assess coverage by looking at the MSA visually (most MSA viewers color-code positions by conservation level) or by calculating per-position coverage scores. For each position in the query sequence, count what fraction of the MSA sequences have a non-gap character at that position. Positions with coverage above 70-80% typically predict well, while positions with coverage below 50% often have low pLDDT scores and high positional uncertainty. The relationship between coverage and confidence is particularly strong for terminal regions—N-terminal and C-terminal extensions often have poor coverage because different organisms have different lengths, and these terminal regions frequently have low confidence in predictions.

Sequence Diversity: Sampling Across Evolutionary Space

Having many sequences is good, but having evolutionarily diverse sequences is better. An MSA that contains sequences spanning a wide range of evolutionary distances—from very close homologs that differ by just a few percent to remote homologs that share only 20-30% sequence identity—provides richer information than an MSA that contains only closely related sequences. The close homologs provide high-resolution information about recent evolutionary constraints and which positions are absolutely conserved, while the remote homologs provide information about deeper structural constraints and long-range coevolution patterns that might not be visible in closely related sequences.

One simple way to assess diversity is to examine the distribution of sequence identities in your MSA. Plot a histogram of pairwise sequence identities between the query and all MSA sequences. A good MSA should have sequences distributed across a broad range, ideally from 30-40% identity (the "twilight zone" where homology becomes difficult to detect) up to 90-100% identity. An MSA that's dominated by very high-identity sequences (all above 80%) has redundancy that doesn't contribute much information, while an MSA composed entirely of very low-identity sequences (all below 40%) might have alignment errors that introduce noise. The sweet spot is a mixture of both close and remote homologs.

#Challenging Cases: When MSAs are Poor

Not all proteins have the luxury of deep, high-quality MSAs. Understanding which types of proteins tend to have MSA problems, and why, can help you anticipate difficulties and choose appropriate strategies. Some categories of proteins are inherently challenging for MSA-based structure prediction, not because there's anything wrong with the prediction method, but because the evolutionary information that the method relies upon is simply not available in current sequence databases.

Orphan Proteins and Taxonomically Restricted Genes

Orphan proteins, also called taxonomically restricted genes (TRGs), are proteins that appear to be unique to a single species or a narrow taxonomic group, with no detectable homologs in other organisms. These proteins might be recently evolved innovations, horizontal gene transfer events from viruses or other organisms, or they might be rapidly evolving proteins whose sequence similarity to ancestral forms has been erased by millions of years of mutation. Whatever their origin, orphan proteins present a fundamental challenge for MSA-based methods: if there are no homologous sequences in the databases, there is no MSA to build, and therefore no coevolution information to extract.

The frequency of orphan proteins varies dramatically across organisms. Bacteria and archaea tend to have fewer orphans because horizontal gene transfer and their large population sizes mean most genes have homologs somewhere. Eukaryotes, particularly plants and fungi, have higher rates of orphan proteins. Human proteins are generally well-covered by MSAs because we share most of our proteome with other vertebrates and many proteins with all eukaryotes, but even in humans, perhaps 5-10% of proteins are difficult to find good MSAs for. For organisms with small research communities or unusual biology, the situation can be much worse—some newly sequenced organisms have 30-50% orphan proteins. For these proteins, methods like ESMFold that don't rely on MSAs may perform better than AlphaFold2.

Rapidly Evolving Proteins

Some proteins evolve so rapidly that homologous sequences diverge beyond recognition within relatively short evolutionary timescales. This is particularly common for proteins involved in host-pathogen interactions (where there's an evolutionary arms race driving rapid change), reproductive proteins (which often evolve under sexual selection), and immune system proteins (which need to recognize and respond to rapidly evolving pathogens). These proteins may have homologs in related species, but the sequence similarity might be so low that standard search algorithms cannot detect the relationships, resulting in shallow MSAs even though homologous proteins actually exist in the databases.

One signature of rapidly evolving proteins is that you can find some homologs in closely related species (building a shallow MSA of very similar sequences) but no homologs beyond a certain evolutionary distance. For example, you might find good homologs across all mammals but nothing in birds or reptiles, even though you know from synteny or functional studies that the protein exists in those organisms. This creates an MSA with high sequence identity but low diversity, which provides limited coevolution information. For such proteins, you might try more sensitive search methods (like HHblits with lower e-value cutoffs) or metagenomic databases to find additional distant homologs.

Intrinsically Disordered Regions

Intrinsically disordered proteins or regions (IDPs/IDRs) are segments that lack stable three-dimensional structure under physiological conditions, instead existing as dynamic ensembles of conformations. These regions evolve much more rapidly than structured domains because there are fewer structural constraints—an amino acid substitution in a disordered region is less likely to break anything because there's no specific structure to break. The rapid evolution means that sequence similarity degrades quickly with evolutionary distance, resulting in poor MSA coverage for disordered regions even when the flanking structured domains have excellent MSAs.

AlphaFold2 correctly predicts disorder by assigning low pLDDT scores to these regions, which is actually the right answer—a low confidence prediction for a region that genuinely lacks structure is more useful than a high confidence prediction of an incorrect structure. However, users sometimes misinterpret low pLDDT in disordered regions as a failure of the prediction, when it's actually accurately reporting biological reality. You can distinguish true disorder (low pLDDT because the region is disordered) from poor MSA coverage (low pLDDT because of missing evolutionary information) by using disorder prediction tools like IUPred or by examining sequence composition—disordered regions are typically enriched in polar and charged residues and depleted in hydrophobic residues.

Membrane Proteins

Transmembrane proteins face unique evolutionary constraints. The transmembrane segments that span lipid bilayers must have predominantly hydrophobic residues to be compatible with the nonpolar membrane interior, and this strong compositional bias means that different sequence solutions can converge on similar structures. The result is that transmembrane segments often have low sequence similarity even between clearly homologous proteins, making MSA construction challenging. Additionally, membrane proteins are difficult to study experimentally (they comprise less than 5% of structures in the PDB despite being about 30% of the proteome), meaning there are fewer structures to train on and fewer sequences have been functionally characterized.

Despite these challenges, AlphaFold2 actually performs remarkably well on membrane proteins, much better than earlier structure prediction methods. The key seems to be that even with shallow MSAs, the strong physical constraints of membrane protein architecture (hydrophobic transmembrane helices, specific helix-helix packing geometries, conserved functional motifs) provide enough information for the network to make good predictions. However, membrane protein predictions should still be interpreted cautiously, particularly for complex multi-pass membrane proteins or large extracellular domains where MSA depth may be critical.

#Strategies for Improving Poor MSAs

When you encounter a protein with a poor MSA, you're not necessarily stuck with low-quality predictions. There are several strategies you can try to improve the MSA, each with its own tradeoffs in terms of computational cost, complexity, and likelihood of success. Understanding these strategies and when to apply them can help you salvage predictions for challenging targets and make better use of available computational resources.

Search Metagenomic Databases

One of the most effective strategies for improving shallow MSAs is to search metagenomic sequence databases. Traditional protein databases like UniProt consist of proteins from cultured organisms whose genomes have been sequenced and annotated. However, the vast majority of microbial diversity on Earth consists of uncultured organisms that cannot be grown in the laboratory. Metagenomic sequencing—extracting and sequencing DNA directly from environmental samples like soil, ocean water, or the human gut—has revealed an enormous reservoir of previously unknown protein sequences from these uncultured organisms.

For many proteins, particularly bacterial and archaeal proteins, searching metagenomic databases can dramatically increase MSA depth. Databases like MGnify contain hundreds of millions of protein sequences from metagenomic studies, many of which are not present in traditional databases. The environmental diversity sampled by metagenomics means you often find sequences spanning a wider range of evolutionary distances than in culture-based databases. AlphaFold2's default pipeline includes metagenomic database searches, but you can sometimes improve results by using more comprehensive or more recent metagenomic databases, or by tuning search parameters to be more sensitive (at the cost of longer search times and potentially more false positives).

Use More Sensitive Search Methods

The standard jackhmmer search used in AlphaFold2's default pipeline uses specific e-value cutoffs and iteration limits that balance sensitivity with specificity. For very difficult cases, you might gain additional sequences by using more sensitive search strategies. HHblits, which searches databases of profile Hidden Markov Models rather than individual sequences, can detect more remote homologs than sequence-based methods. Iterative searches like PSI-BLAST or jackhmmer with more iterations can progressively build up sensitivity by updating the search profile after each iteration. Lowering e-value thresholds allows more distant matches, though at the risk of including false positives that are not true homologs.

The tradeoff with more sensitive searches is twofold: computational time (more sensitive searches take longer) and potentially increased noise (false positive matches that aren't really homologs will add noise to the MSA rather than signal). For critical predictions where accuracy is paramount and you have the computational resources, trying multiple search strategies and comparing the results can be worthwhile. However, for routine predictions or when MSA depth is already good, the standard pipeline is usually sufficient.

Domain-Based MSA Construction

Many proteins are composed of multiple domains—independently folding structural units that have often been evolutionarily shuffled and recombined in different ways. A protein might have one domain that's well-conserved across all life (like an ATP-binding domain) and another domain that's specific to a narrow taxonomic group (like a species-specific regulatory domain). If you build an MSA for the whole protein, the poorly conserved domain will drag down the overall statistics, making the MSA appear shallow even though part of the protein has excellent coverage.

For such proteins, building separate MSAs for each domain and then concatenating them can improve prediction quality. Domain prediction tools like Pfam, SMART, or InterPro can help identify domain boundaries. You then build an MSA for each domain separately, potentially using different databases or search strategies optimized for each domain's evolutionary properties. The domain-specific MSAs are then combined into a single MSA for structure prediction. This approach is more labor-intensive but can substantially improve predictions for multi-domain proteins with heterogeneous evolutionary histories. Some prediction servers, including Protogen Bio, offer automated domain-based MSA construction as an advanced option.

Template-Based Approaches

When MSA-based approaches fail, you can fall back on template-based modeling. Even if you can't find many sequence homologs, you might be able to find a protein with known structure that has a similar fold. AlphaFold2 actually uses template information as one of its inputs (in addition to the MSA), searching the PDB for structures that might provide geometric hints about the fold. For proteins with very shallow MSAs, increasing the emphasis on template information or using dedicated homology modeling tools like MODELLER or SWISS-MODEL might yield better results than pure AlphaFold2 predictions.

The challenge with template-based approaches is that they're only useful if a suitable template exists—and for truly novel folds or orphan proteins, there might be no appropriate template. Additionally, template-based models are fundamentally limited by the template: if the template has errors or represents a different conformational state, those limitations will be inherited by your model. Still, for proteins with poor MSAs, a template-based model might be your best option, and it's worth trying multiple approaches and comparing the results.

#When MSAs Fail: ESMFold and Other Alternatives

The fundamental dependence of AlphaFold2 on MSAs is both its greatest strength (when MSAs are good) and its greatest weakness (when they're poor). This has motivated the development of alternative structure prediction methods that don't rely on MSAs, most notably Meta's ESMFold. Understanding when to use these alternatives, and what tradeoffs they involve, can help you choose the right tool for each prediction task.

ESMFold takes a radically different approach: instead of explicitly searching for and aligning homologous sequences, it uses a large language model (ESM-2) trained on hundreds of millions of protein sequences to internalize evolutionary patterns. The language model learns, in an unsupervised way, which amino acid sequences are likely and which are not, which sequence patterns tend to occur together, and which patterns are characteristic of different structural features. This learned representation captures much of the same evolutionary information that MSAs provide, but in a form that doesn't require explicit sequence searching at inference time. The result is that ESMFold can make predictions in seconds rather than minutes, and it performs comparably to AlphaFold2 on proteins with shallow MSAs.

The practical implication is clear: for proteins where AlphaFold2 MSA analysis indicates poor coverage (Neff < 16, low coverage, few diverse sequences), ESMFold is likely to give comparable or better results while being much faster. For proteins with good MSAs (Neff > 64, good coverage, diverse sequences), AlphaFold2 will generally give more accurate predictions. In the middle ground (Neff 16-64), either method might work well, and trying both and comparing confidence scores is a reasonable strategy. Protogen Bio makes this easy by offering both AlphaFold2 and ESMFold predictions, allowing you to choose the right tool for each protein.

#Practical Workflow: MSA-Guided Structure Prediction

Based on everything we've discussed, here's a practical workflow for structure prediction that takes MSA quality into account. This workflow will help you make informed decisions about which prediction method to use, how to interpret results, and when to invest time in improving MSAs versus moving to alternative approaches. Following this workflow can save you significant time and help you avoid common pitfalls.

#Conclusion: MSAs as the Foundation of Modern Structure Prediction

Multiple sequence alignments are not just an input to AlphaFold2—they are the evolutionary lens through which the algorithm views your protein. Every pattern of conservation and covariation in the MSA is a clue about structure, every position with deep coverage is a position where the prediction can be confident, and every gap in the MSA is a position where uncertainty is unavoidable. Understanding MSAs, knowing how to assess their quality, and learning when they're sufficient versus when alternative approaches are needed, is essential for anyone seriously using structure prediction in their research.

The field is evolving rapidly. New sequence databases are being created, better search algorithms are being developed, and methods that reduce or eliminate the dependence on MSAs (like ESMFold) are becoming more mature. However, the fundamental insight that evolution teaches us about structure—that correlated mutations reveal spatial proximity, that conservation indicates functional importance, and that diversity spans the space of allowed sequences—will remain relevant regardless of how the specific methods change. By understanding these principles and the role of MSAs in encoding evolutionary information, you'll be better equipped to use structure prediction effectively and to interpret its results appropriately.