The Annotation Is Not the Answer
Zetobit · Bioinformatics Insight Series
Pipeline ArchitectureThe Annotation Is Not the Answer
Why variant prioritization is a systems problem, not a lookup — and what separates a pipeline that surfaces the causal variant from one that buries it.
A modern germline exome yields on the order of twenty thousand variants per sample. A genome yields four to five million. Somewhere in that list may sit a single change that explains a patient's disease, drives resistance to a therapy, or invalidates a trial arm. The uncomfortable truth of clinical and translational genomics is that finding it is rarely limited by sequencing quality or annotation coverage. It is limited by prioritization — the logic that decides what a human ever looks at.
Annotation and prioritization are often discussed as one step. They are not. Annotation attaches facts to a variant: population frequency, predicted consequence, conservation, in-silico scores, prior classifications. Prioritization decides which of those annotated variants deserve attention, and in what order. The first is a data-joining problem that is largely solved. The second is where pipelines quietly succeed or fail, and it is almost never solved by adding another database.
01The failure mode is inclusion, not omission
When a prioritization pipeline underperforms, the instinct is to add sources — another frequency database, another predictor, another disease panel. This addresses the fear of missing the causal variant. But in practice the dominant failure mode in day-to-day interpretation is not that the true variant was dropped. It is that it was retained alongside two hundred others of superficially similar rank, and a reviewer with finite time never reached it.
Every annotation added to a variant is also a potential reason to keep it in view. More annotation without a corresponding filtering discipline does not sharpen a candidate list — it thickens it. The engineering question is therefore not "what else can we annotate?" but "what does each annotation actually change about the decision to look?"
02Frequency is a filter; consequence is a lens; scores are tie-breakers
Not all annotations play the same role, and conflating them is a common design error.1 Treating an in-silico pathogenicity score as a hard filter, for example, discards variants at a threshold that was never calibrated for that use. A cleaner mental model assigns each annotation class an explicit job.
| Annotation class | Proper role | Common misuse |
|---|---|---|
| Population frequency (gnomAD2) | Hard filter against a disease-appropriate threshold | Applying a single global cutoff regardless of inheritance model or ancestry |
| Predicted consequence | Lens that focuses attention by mechanism | Trusting a single transcript's annotation as definitive |
| In-silico scores (REVEL4, CADD5) | Ordering and tie-breaking within a filtered set | Using as a binary pathogenic/benign gate |
| Prior classifications (ClinVar3) | Strong prior, weighted by review status | Accepting one-star assertions at face value |
The distinction matters because filters are lossy and ordering is not. A filter you set wrong loses the answer permanently for that run; an ordering you set wrong only makes the answer harder to reach. Sound prioritization pushes as much logic as possible into ordering, and reserves hard filtering for annotations that are both reliable and mechanistically defensible — frequency being the cleanest example.
03Transcript choice quietly determines the answer
A single genomic position can be annotated as intronic against one transcript and as a missense change against another. If a pipeline reports consequence against an arbitrary or purely canonical transcript, it can systematically down-rank real variants in genes where the clinically relevant isoform is not the default. This is not an edge case; it is a routine source of discordance between labs analyzing identical data.
Design signal
Transcript selection should be explicit, versioned, and — where a disease-relevant isoform is known — panel-specific. "We used the canonical transcript" is a defensible sentence only if you can name which canonical definition, from which release.
04Prioritization is where phenotype enters
Purely variant-intrinsic ranking — frequency, consequence, scores — treats every case identically. But two patients with the same variant list and different phenotypes should not receive the same candidate ranking. Phenotype-driven approaches, whether HPO-term-based7 gene prioritization or explicit candidate-gene panels, are what convert a generic filtered list into a case-specific one.6
This is also where the strongest gains in reviewer efficiency come from, and where the most fragile assumptions hide. A phenotype model that is too narrow will down-rank the unexpected finding that turns out to be the answer. One that is too permissive contributes nothing. Tuning this balance is not a one-time configuration — it is an ongoing calibration against real case outcomes.
05What a defensible prioritization pipeline looks like
Annotate exhaustively and neutrally; make no ranking decisions during annotation. Ranking logic lives in one auditable place, not scattered across the join.
Reserve hard filters for frequency and clearly non-functional consequence classes. Everything softer becomes ordering, not exclusion.
Pin every reference — transcript set, frequency release, ClinVar snapshot — so a result is reproducible and a discordance is diagnosable.
Case-specific ranking should be driven by structured phenotype, not applied as an afterthought filter over a generic list.
Track where the causal variant actually landed in the ranked list across resolved cases. That distribution — not any single tool's benchmark — is your true measure of prioritization quality.
A caution on benchmarks
Published performance metrics for individual predictors are measured on curated variant sets, not on your case mix, your genes, or your ranking logic. A tool's standalone AUC tells you little about its marginal contribution once it sits inside a pipeline alongside frequency and phenotype. Evaluate the pipeline's output, not the parts in isolation.
06The takeaway
Variant annotation is a solved problem in the sense that matters: the facts are available, and joining them to a variant list is routine engineering. Prioritization is not solved, because it is not a data problem. It is a decision architecture — a set of explicit choices about what filters, what orders, what a reviewer sees first, and how those choices are calibrated against reality.
Pipelines that treat prioritization as an afterthought to annotation tend to produce candidate lists that are technically complete and practically unusable. The ones that treat it as the core design problem — assigning each signal a defined role, keeping filtering conservative, making every reference explicit, and measuring where the answer actually lands — are the ones that let a human reach the variant that matters. That distinction is invisible in a feature list and decisive in a clinic.
References & further reading
- Richards S, Aziz N, Bale S, et al. Standards and guidelines for the interpretation of sequence variants: a joint consensus recommendation of the ACMG and the AMP. Genet Med. 2015;17(5):405–424. Read the guideline →
- Karczewski KJ, Francioli LC, Tiao G, et al. The mutational constraint spectrum quantified from variation in 141,456 humans (gnomAD). Nature. 2020;581(7809):434–443. Read on Nature →
- Landrum MJ, Lee JM, Benson M, et al. ClinVar: improving access to variant interpretations and supporting evidence. Nucleic Acids Res. 2018;46(D1):D1062–D1067. Read on NAR →
- Ioannidis NM, Rothstein JH, Pejaver V, et al. REVEL: an ensemble method for predicting the pathogenicity of rare missense variants. Am J Hum Genet. 2016;99(4):877–885. Read on AJHG →
- Rentzsch P, Witten D, Cooper GM, Shendure J, Kircher M. CADD: predicting the deleteriousness of variants throughout the human genome. Nucleic Acids Res. 2019;47(D1):D886–D894. Read on NAR →
- Smedley D, Jacobsen JOB, Jäger M, et al. Next-generation diagnostics and disease-gene discovery with the Exomiser. Nat Protoc. 2015;10(12):2004–2015. Read on Nature Protocols →
- Köhler S, Gargano M, Matentzoglu N, et al. The Human Phenotype Ontology in 2021. Nucleic Acids Res. 2021;49(D1):D1207–D1217. Read on NAR →
Building or auditing a variant prioritization pipeline?
Zetobit designs and reviews annotation and prioritization workflows for clinical genomics, biotech, and pharma — from transcript strategy and filtering logic to phenotype-driven ranking and outcome calibration.
Start a conversation