Why Conservation and Constraint Answer Different Questions
A locus can have evidence from evolutionary comparisons as well as from observed human variation. These evidence types are related, but they describe different units and ask different questions. A conservation score concerns a base or short sequence: how strongly has that sequence been preserved across the species and evolutionary history represented by a comparison? A gene-level constraint metric summarizes whether a gene has fewer variants of a specified class than expected in a human population data set.
This distinction matters when interpreting candidate variants from a fine-mapped locus. A conserved base may make a variant at that position more compelling for follow-up, but it does not establish that the variant affects a gene or causes the trait. A constrained gene may be sensitive to some classes of genetic change, but that does not establish that a particular locus variant affects that gene. Neither kind of score identifies a causal mechanism.
Base-Level Conservation Scores
Conservation scores are calculated from a multiple-sequence alignment, which places corresponding genomic positions from different species alongside one another. The result depends on the species included, the quality of their assemblies and alignment, and the evolutionary model used. A score at a coordinate is therefore not an intrinsic property independent of the comparison that produced it. Interpretation also requires compatible genome coordinates and allele representation: an annotation at the wrong build or position may describe a different base.
Several widely used scores summarize conservation in different ways. phastCons uses a phylogenetic hidden Markov model to identify conserved elements against a background model; its per-base output is commonly interpreted as support, on a zero-to-one scale, for membership in a conserved element. phyloP tests conservation or acceleration at individual sites using a phylogenetic model. Depending on the output convention, positive values support slower-than-neutral evolution (conservation) and negative values support faster-than-neutral evolution (acceleration); values near zero provide little evidence for either relative to the model. GERP++ estimates rejected substitutions: substitutions expected under neutrality minus those inferred to have occurred. A larger positive rejected-substitution score indicates stronger inferred constraint at that position.
These scores are not interchangeable. A phastCons value of 0.8 is not directly comparable to a phyloP score of 4 or a GERP++ score of 4. Even scores from the same method are interpreted in relation to their model, alignment, and local genomic context. The scores describe evolutionary patterns, not the biochemical effect of a particular alternate allele. A highly conserved position could be affected differently by different substitutions, and an alignment may be uninformative when few suitable species are available.
Worked Example: Comparing Two Candidate Bases
Suppose a hypothetical locus has two candidate substitutions, A and B. Both coordinates have been checked against the same reference genome build, and both are covered by the same conservation tracks. At A, phastCons is 0.96, phyloP is +4.2, and GERP++ is +3.8. At B, the corresponding values are 0.18, −1.1, and 0.2. What can these values support?
For each score separately, A shows stronger evidence of conservation than B: its phastCons score is higher, its positive phyloP score supports slower-than-neutral evolution, and its GERP++ score indicates more rejected substitutions. At B, phastCons provides less support for a conserved element, phyloP is negative in this scoring convention, and GERP++ is close to zero.
The conclusion is a comparison of evolutionary evidence at the two positions. It is not valid to add 0.96, 4.2, and 3.8 and treat the sum as a calibrated effect score: the methods use different scales and models. Nor does A’s higher conservation establish that substitution A changes a functional element or influences the phenotype. Such claims require evidence beyond these base-level scores.
Gene-Level Constraint Metrics
Constraint analyses compare observed counts of variants in a gene with expected counts. Expectations are typically generated using a mutation model that accounts for factors such as sequence context and the probability of different nucleotide changes. Analyses often separate variant classes, such as predicted loss-of-function (LoF) or missense variants, because they represent different possible consequences. The counts and estimates depend on the population sample, variant calling and filtering, transcript definitions, and the method used to predict variant consequences.
For a given class, let \(O\) be the observed number of variants and \(E\) the expected number under the model. The observed-to-expected ratio, \(O/E\), describes the observed count relative to expectation. A value below one indicates fewer observed variants than expected; the smaller it is, the greater the estimated depletion. The simple quantity \(1-O/E\) expresses the proportional shortfall when \(O/E\) is between zero and one. This is a descriptive summary, not a direct measurement of a selection coefficient or the probability that a specific variant is harmful.
Some constraint resources report uncertainty around \(O/E\), since observed counts are finite and the expected count is model-based. The upper bound of an interval is useful when ranking genes by how confidently their counts are depleted: a lower upper bound indicates that even the less constrained end of the reported range remains relatively depleted. This still depends on the underlying data and model.
Two commonly encountered metrics need particularly careful interpretation. pLI is a model-based probability that a gene belongs to a class of genes intolerant of heterozygous LoF variation, under the model used to estimate it. It is not the probability that a particular LoF variant is pathogenic, nor the probability that a gene causes a particular trait. LOEUF is the upper bound of a 90% confidence interval for a gene’s LoF observed-to-expected ratio in the gnomAD framework. Lower LOEUF values indicate stronger evidence of LoF constraint under that framework; higher values indicate less evidence for strong LoF depletion. pLI and LOEUF summarize related but not identical outputs and should not be treated as the same scale.
Worked Example: Calculating and Interpreting LoF Depletion
In a hypothetical gene, a specified population data set contains \(O=6\) high-confidence LoF variants, while the mutation model predicts \(E=24\). Assume for this example that the gene definition, variant filters, and expectation model have been checked and that the data set is large enough for the reported estimate to be informative.
The observed-to-expected ratio is
The observed count is one quarter of the expected count. The corresponding proportional depletion is
Thus, the observed count is 75% below the model expectation. This supports a description of LoF depletion for this gene in this data set. It does not mean that every LoF change has the same effect, that any particular candidate variant is causal, or that the gene is involved in the phenotype under study. Uncertainty in the count, expectation model, and gene definition should be considered before making a strong comparison with another gene.
A Short Justification for the Depletion Summary
The ratio \(O/E\) is useful because it puts a count on a scale relative to what the specified model predicts. If the observed count equals expectation, \(O/E=1\) and \(1-O/E=0\), indicating no proportional depletion. If \(O\) is half of \(E\), then \(O/E=0.5\) and \(1-O/E=0.5\), a 50% shortfall. If no variants are observed and \(E>0\), the ratio is zero and the shortfall is 100%. This arithmetic justifies calling \(1-O/E\) a proportional depletion summary when \(O/E\leq1\); it does not establish why depletion occurred. The interpretation as selection-related constraint relies on the adequacy of the mutation expectation, variant calling, and other assumptions of the analysis.
A ratio above one gives \(1-O/E<0\), which indicates an observed count above expectation, not “negative constraint.” It may reflect sampling variation, model misspecification, or other features of the data. Constraint estimates should therefore be interpreted with their uncertainty and analytic context, rather than as exact biological constants.
Worked Example: Reading a Gene-Constraint Profile
Consider a hypothetical gene with a LoF \(O/E\) estimate of 0.18, a 90% interval of 0.10 to 0.31, LOEUF of 0.31, and pLI of 0.98. Assume the metrics were obtained from a suitable population resource, and that the gene and transcript definitions are appropriate for the question.
The point estimate indicates that the observed LoF count is 18% of expectation, or \(1-0.18=0.82\), an 82% shortfall. The upper bound of 0.31 is also below one, so the reported range supports depletion rather than merely a point estimate below expectation. The LOEUF value summarizes that upper bound in this framework, and its low value is consistent with strong LoF constraint. The high pLI value is also model-based evidence that the gene is likely to fall in a LoF-intolerant class under the pLI model.
Taken together, these values support the qualified statement that the gene shows strong evidence of depletion for LoF variation in the analyzed data. They do not show that a candidate locus variant is LoF, that it affects this gene, or that the gene’s constraint explains the trait association. Those are separate questions requiring variant- and locus-specific evidence.
Combining the Evidence Without Confusing Its Scale
Conservation and constraint can agree, disagree, or be unavailable. A conserved base within a constrained gene is consistent with evolutionary importance at both a particular position and the broader gene level. It is not two independent proofs of a particular variant’s effect: both measures reflect evolutionary pressures and can be related. Conversely, weak conservation at one base does not negate gene-level constraint, because a gene can contain positions and variant classes with different consequences.
Worked Example: A Conserved Base in a Gene With Modest Constraint
A hypothetical candidate substitution lies at a base with phastCons 0.91 and positive phyloP, while the gene has a LoF \(O/E\) of 0.88 and a LOEUF of 1.20. Assume the base-level annotation and the gene-level estimates are correctly mapped, but no functional assay or molecular association is available.
The base-level values support evolutionary conservation at that position. The gene-level values provide little evidence for strong depletion of LoF variants in the analyzed data, although that does not establish that the gene is unimportant or that every variant class is tolerated. The appropriate summary keeps these statements separate: the nucleotide is conserved in the chosen comparison, while the gene does not show strong LoF constraint by these estimates.
Neither result settles whether the candidate substitution affects the phenotype. In particular, a constrained gene is not required for every functional variant, and conservation at a base does not establish the affected gene or process. The scores help prioritize interpretation; they do not supply a causal mechanism.
Common Pitfalls and Practical Checks
- Equating different score scales. A large number from one conservation method cannot be compared directly with a number from another. Interpret each score using its own definition and reference model.
- Calling a conservation score a variant-effect prediction. Conservation indicates evolutionary preservation at a position, not the consequence of a particular alternate allele. Check the variant’s coordinate, reference allele, and consequence separately.
- Reading pLI as a variant probability. A high pLI is a gene-level, model-based summary about LoF intolerance. It is not the probability that a candidate variant is pathogenic or causal.
- Ignoring uncertainty and data quality. A low \(O/E\) estimate is harder to interpret if the expected count is small, the interval is wide, or variant calling and gene definitions are unsuitable.
- Turning an annotation into a mechanism. Neither conservation nor constraint alone establishes which gene mediates an association, what molecular process is altered, or whether the candidate variant is causal.
Check Your Understanding
Answer each question by distinguishing what the score supports from what it does not establish.
- What is the main difference between the genomic unit summarized by a conservation score and the unit summarized by a gene constraint metric?
- A phyloP score is negative under a convention where positive values support conservation. What does that sign indicate, and what does it not prove about a variant?
- A gene has \(O=10\) and \(E=40\) for a variant class. Calculate \(O/E\) and \(1-O/E\), then state the qualified interpretation.
- Why is a high pLI not the probability that a candidate variant is pathogenic?
- How should an analyst describe a conserved candidate base in a gene whose LoF constraint estimates show little depletion?