Multiple sequence alignment depth shows up in almost every discussion of computational protein design as a variable that matters. "Deep MSA" and "shallow MSA" get used casually, but the practical implications are rarely spelled out in a way that helps a protein engineer decide what to do when their protein family has borderline coverage.
This piece is an attempt to be concrete about what MSA depth actually affects, how to measure it usefully, and what the downstream consequences are for a generative design campaign.
What MSA Depth Is Actually Measuring
An MSA is a collection of homologous sequences aligned by position. Depth in this context means the number of sequences in the alignment, but raw count is a misleading metric for most practical purposes. What you care about is effective depth: the number of sequences that contribute independent information to the statistical analysis.
Effective sequence number (Neff) is typically calculated by weighting each sequence by the inverse of the number of sequences within a defined similarity threshold of it. Two sequences that are 95% identical contribute roughly one effective sequence between them, not two. For most protein families, the raw MSA count from a jackhmmer search of UniRef90 can look impressively large while Neff is much smaller because the search returns many closely related paralogs or variants from single sequencing projects that cluster together tightly.
The statistical method most relevant to co-evolutionary design, direct coupling analysis (DCA), has well-characterized dependence on Neff. Below roughly Neff of 100-200, the inverse covariance matrices used to identify coupled positions become unstable. The method can still run and produce output, but the identified couplings include a higher fraction of statistical artifacts compared to genuine co-evolutionary signals. Predictions built on those couplings are correspondingly less reliable.
How MSA Depth Affects Specific Types of Predictions
The consequences of shallow MSAs are not uniform across all types of predictions. Different outputs have different sensitivity to depth degradation.
Positional conservation scores are relatively robust to shallow MSAs. If a position is invariant across all 200 sequences in your MSA, you know it is conserved regardless of whether those 200 sequences represent Neff of 60 or Neff of 180. Conservation is a single-column statistic and tolerates low depth reasonably well.
Pairwise co-evolutionary couplings are much more sensitive. Identifying that position 34 and position 97 co-vary in a way that suggests structural or functional interaction requires enough sequence diversity that the correlation between them is statistically distinguishable from noise. This is fundamentally a sample size problem. Short sequences and thin MSAs both reduce the precision of coupling estimates.
Generative sampling from learned distributions amplifies the effect of shallow MSAs in a particular way. When a model trained on a sparse distribution generates new sequences, those sequences tend to cluster closer to the training data than a well-sampled model would produce. The generative diversity is suppressed. For design purposes, this means the candidate list may not actually explore the breadth of viable sequence space that the fitness landscape would permit. You get a narrow set of candidates that looks locally consistent but may miss structurally distinct variants that are equally viable.
Practical Thresholds and What to Do at Each Level
Rather than a single threshold, it is more useful to think in terms of operational zones.
Above Neff of approximately 500 with reasonable phylogenetic diversity, co-evolutionary signals are generally reliable for standard design applications. This is where most well-studied enzyme families and antibody frameworks sit. You can run a generative design campaign with standard confidence and interpret ranked outputs at face value with appropriate uncertainty bands.
Between Neff of 100 and 500, the signal is meaningful but degraded. Coupling predictions for positions deep in the alignment are still useful; couplings involving highly variable positions should be treated with more skepticism. In this zone we generate a larger candidate pool than we would for a deep-MSA family and screen a higher fraction of the top-ranked candidates before committing to a refined second round. The retrospective validation step is particularly important here, because the rank-correlation on existing data tells you empirically whether the model's predictions are informative for your specific protein family.
Below Neff of 100, co-evolutionary design methods are in territory where the false discovery rate on predicted couplings becomes high. This does not mean structural guidance is unavailable. If a high-quality structure exists, residue-level structural analysis can substitute for co-evolutionary coupling in identifying which positions to target. The candidate generation shifts from a co-evolution-driven approach to a structure-guided approach, and the expected hit rate per screened variant may be lower, requiring a broader screening campaign per round.
The Orphan Protein Problem
A specific case worth addressing: what happens when your target protein has essentially no homologs in any public database? This is not common for the types of proteins most engineering groups work with, but it does arise for newly characterized proteins from environmental metagenomics, some viral proteins, and certain classes of synthetic biology constructs.
For these truly orphan proteins, MSA-based approaches are not viable. The honest answer is that the computational toolkit is much weaker. Structure-guided rational design, assisted by recent structure prediction tools, is typically the best available approach. You can identify the active site, assess the structural consequences of specific mutations using predicted free energy calculations, and generate a small focused set of candidates based on physical chemistry reasoning rather than evolutionary statistics.
We are not claiming structure-guided approaches are always inferior to MSA-based generative design. For well-characterized structural problems with clear mechanistic hypotheses, structure-guided design can outperform evolutionary approaches even when MSA depth is adequate. The issue is that orphan proteins combine a hard problem (unknown sequence-function relationship) with a weak signal (no evolutionary context), which makes confident computational prediction genuinely difficult regardless of method.
Augmenting Thin MSAs Before Design
When a protein family has a thin MSA but is not a true orphan, there are practical strategies to improve effective depth before running a design campaign.
The first strategy is iterative database search. A single jackhmmer search against UniRef90 with default parameters often misses distantly related sequences that carry informative variation. Running iterative rounds with progressively looser inclusion thresholds and filtering for structural plausibility (rather than sequence similarity alone) can substantially increase coverage for families that are modestly underrepresented in curated databases but have structural relatives in the broader sequence space.
The second strategy is structural homolog inclusion. If the target protein has a solved or predicted structure, searching for structurally similar proteins using tools that compare geometry rather than sequence can identify evolutionary relatives that fall below sequence similarity thresholds. These structural homologs, aligned structurally rather than by sequence, can supplement the MSA with additional co-evolutionary information provided the structural alignment is handled carefully to preserve meaningful positional correspondences.
Neither strategy is a free lunch. Iterative search risks introducing sequences that are not genuinely homologous, which adds noise rather than signal. Structural homolog inclusion can introduce alignment artifacts at variable loop regions. Both adjustments require careful quality control on the augmented MSA before using it for design. When we run MSA augmentation for a thin-coverage family, we always compare the coupling predictions before and after augmentation to assess whether the additional sequences are improving or degrading the signal quality in the positions most relevant to the design objective.