What Variant Effect Predictors Do, and Which Ones to Run on Your Own Genome

A variant effect predictor is a program that takes a genomic coordinate and an allele change and returns a guess about consequence. There are two distinct things people mean by the phrase. The first is annotation. Given chr7:117559590 G>A, it answers which transcripts overlap it, which codon and what amino acid substitution. It also answers whether the variant hits a splice site. That is deterministic bookkeeping against a gene model, and Ensembl VEP or SnpEff does it. The second is pathogenicity or functional scoring: a model that outputs a number between 0 and 1 (or a scaled rank) estimating how damaging that change is. CADD, REVEL, AlphaMissense, SpliceAI, and AlphaGenome live here. You need both, and you should never confuse the second for a clinical answer. A high AlphaMissense score on a variant in your VCF means a model thinks the substitution is likely damaging to protein function. It does not mean you have a disease.

Your whole genome has roughly 4–5 million variants against GRCh38, of which about 20,000 are coding and 10,000–12,000 are missense. Perhaps 50–100 are rare protein-truncating variants in genes with known loss-of-function mechanisms. The job of a predictor stack is to take you from 5 million rows to a few dozen you can read.

Start with annotation, then layer scores

Run Ensembl VEP as the backbone. It is the most complete transcript annotator, and its plugin system lets you attach every score you care about in one pass rather than doing five rounds of bcftools annotate.

vep -i sample.vcf.gz -o sample.vep.vcf.gz \
  --cache --offline --dir_cache /refs/vep_cache --assembly GRCh38 \
  --vcf --compress_output bgzip --fork 8 \
  --everything --pick_allele_gene --canonical --mane_select \
  --plugin CADD,/refs/cadd/whole_genome_SNVs.tsv.gz,/refs/cadd/gnomad.genomes.indel.tsv.gz \
  --plugin AlphaMissense,file=/refs/AlphaMissense_hg38.tsv.gz \
  --plugin SpliceAI,snv=/refs/spliceai_scores.raw.snv.hg38.vcf.gz,indel=/refs/spliceai_scores.raw.indel.hg38.vcf.gz \
  --plugin dbNSFP,/refs/dbNSFP4.7a_grch38.gz,REVEL_score,ClinPred_score,gnomAD_exomes_AF \
  --custom /refs/clinvar_20250805.vcf.gz,ClinVar,vcf,exact,0,CLNSIG,CLNREVSTAT,CLNDN

Three flags matter more than the rest. --offline --cache avoids the REST API, which will rate-limit you at genome scale and take days. --fork 8 is roughly linear up to 8–16 cores. A WGS VCF takes 20–60 minutes on a laptop with a warm cache. --pick_allele_gene collapses the transcript explosion. Without it, a single variant in a gene with 15 transcripts gives you 15 output rows, and your downstream filtering silently counts things multiple times. We prefer --mane_select as the tiebreaker for human coding work because MANE transcripts are the ones ClinVar and most clinical labs use, so your coordinates line up with theirs.

Use --assembly GRCh38 and make sure your VCF, your cache, and every score file agree. Mixed builds are the single most common failure in a home variant pipeline, and they fail quietly: coordinates shift by a few hundred bases to a few megabases, VEP annotates the wrong gene, and nothing errors. Check with a sentinel variant whose position you know in both builds before trusting anything.

Which scores to trust, and for what

Predictors split by variant class, and a tool that is good at one class is usually useless on another. The literature on this is large and the practical summary is that method choice should follow the variant type and the question. It should not follow the leaderboard ¹.

Missense. This is where the field has the most tools and the most circularity. Many older scores (SIFT, PolyPhen-2, MutationTaster) were trained on ClinVar or HGMD, then benchmarked on ClinVar, which inflates apparent accuracy ². Prefer scores trained without direct pathogenicity labels. AlphaMissense uses structure and evolutionary constraint rather than clinical labels. ESM-1v and EVE are unsupervised on sequence alone. REVEL is an ensemble and performs well but does carry training-set overlap risk. We run AlphaMissense as primary, REVEL as a second opinion, and treat agreement as informative and disagreement as a flag to go read the gene.

Splicing. SpliceAI is the one to run. It predicts four deltas per variant over a window, default ±50 bp. The deltas are acceptor gain, acceptor loss, donor gain and donor loss. We widen to -d 500 when looking for deep intronic cryptic splice creation. A max delta above 0.5 is a strong signal, above 0.2 is worth a look. SpliceAI catches a class of variants that missense predictors have no vocabulary for, including synonymous changes that destroy an exonic splice enhancer.

Genome-wide and noncoding. CADD is the general-purpose fallback because it scores every possible SNV in the genome plus indels. Its output is PHRED-scaled: roughly, 20 = top 1% of all possible substitutions and 30 = top 0.1%. That gives a single comparable number across variant classes ³. It is a blunt instrument in regulatory regions. For regulatory variants specifically, the state of the field until recently was that detection outran prediction: we can find noncoding variants easily and explain them poorly ⁴. AlphaGenome is the current best attempt, predicting tracks like expression, chromatin accessibility, splicing, and contact maps from sequence at base resolution over a megabase context, which lets you ask what a variant does to a specific tissue’s regulatory output rather than just how conserved the base is ⁵. Deep-learning models of this kind have reshaped what is predictable from raw sequence ⁶.

Loss of function. Truncating variants need a different filter entirely. Run LOFTEE (a VEP plugin) to separate high-confidence LoF from the large fraction that are annotation artifacts: last-exon nonsense, variants in low-confidence transcripts, splice variants rescued by a nearby site. Cross-reference gnomAD pLI and LOEUF for the gene. A truncating variant in a gene with LOEUF above 1.0 carries far less weight than the same variant in a gene with LOEUF below 0.35.

A filtering order that works

Ordering matters because each step is cheap relative to the next and the reference data quality varies.

1. Quality. bcftools view -i 'FILTER="PASS" && FMT/DP>=10 && FMT/GQ>=20'. Low-depth calls dominate false positives, and no predictor can rescue a variant that is not real.
2. Frequency. Drop anything with gnomAD popmax AF above 0.001 unless you are specifically looking at common-variant pharmacogenomics. Use popmax, not global AF, or you will discard variants common in one ancestry and rare in yours, and keep the reverse.
3. Consequence. Keep missense, nonsense and frameshift, along with canonical splice, start/stop loss, and inframe indels. Keep everything with SpliceAI max delta above 0.2 regardless of consequence.
4. Score. Now apply AlphaMissense, REVEL, CADD. Do this last, because scores are the least reliable filter and you want to know what you discarded.
5. Known interpretation. Join against ClinVar with review status. A two-star or better pathogenic assertion outranks any predictor. A one-star submission from a single lab does not.

This ordering reflects a general point about genome interpretation: the constraint is rarely the calling, it is deciding which of the thousands of plausible variants matter, and that decision leans on population frequency and prior knowledge more than on any model score ⁷. Broader surveys of the tool ecosystem make the same point about pipeline design, that annotation and filtration are where most of the signal-to-noise gain happens ^{8 9}.

How to read a score without deceiving yourself

Predictor outputs are ranks. They are not probabilities of disease. Three habits keep you calibrated.

Know the prior. If you are healthy and screening a genome, the prior probability that any given high-scoring missense variant is causing you a problem is very low. Predictors are typically evaluated on datasets enriched for pathogenic variants, so their apparent precision does not transfer to a population screen ¹⁰.

Check whether the tool has seen the answer. If a score was trained on ClinVar and you are evaluating a ClinVar variant, the score tells you nothing new.

Look at the gene. The variant alone is not enough. A damaging prediction in a gene with no disease association and no constraint is noise. A moderate prediction in a highly constrained gene with a matching mechanism is worth attention.

If a variant survives all of this and sits in a gene with a real clinical association, that is the point to stop analyzing and talk to a genetic counselor or medical geneticist. Research-grade sequencing is not clinically validated, and confirmation by an accredited lab is a prerequisite before anyone should act on it.

Questions people also ask

What are variant effect predictors? Programs that take a variant and estimate its consequence. Some do deterministic annotation against transcript models (VEP, SnpEff). Others produce a learned score for how damaging the change is (CADD, AlphaMissense, SpliceAI and REVEL). Most workflows need one of each.

Is Ensembl VEP a pathogenicity predictor? Not by itself. Core VEP reports consequence terms and transcript context. Pathogenicity scores come from plugins and custom annotation files you attach at runtime.

Which predictor is most accurate? There is no single answer, and rankings depend heavily on the benchmark and on how much training-set overlap the evaluation allows ^{1 2}. For missense, AlphaMissense and REVEL are the current defaults. For splicing, SpliceAI. For regulatory variants, AlphaGenome ⁵. For everything else, CADD as a genome-wide baseline ³.

Why do two predictors disagree on the same variant? They use different evidence. Conservation-based tools ask whether the position has been preserved across species. Structure-based tools ask whether the substitution destabilizes a fold. Ensembles combine both plus training labels. Disagreement usually means the variant is genuinely ambiguous, which is a useful thing to learn.

Can I run this on 23andMe or AncestryDNA data? Partly. Array data covers roughly 600,000–900,000 sites, almost all common, so you can annotate what is there. Rare variants (the ones predictors are built for) are almost entirely absent. Imputation does not recover them. You need sequencing.

Footnotes

1. Cristian Riccio, Max L. Jansen, Linlin Guo, et al. Variant effect predictors: a systematic review and practical guide. Human Genetics, 2024. https://doi.org/10.1007/s00439-024-02670-5 ↩ ↩²

2. Abhishek Niroula, Mauno Vihinen. Variation Interpretation Predictors: Principles, Types, Performance, and Choice. Human Mutation, 2016. https://doi.org/10.1002/humu.22987 ↩ ↩²

3. Philipp Rentzsch, Daniela Witten, Gregory M Cooper, et al. CADD: predicting the deleteriousness of variants throughout the human genome. Nucleic Acids Research, 2018. https://doi.org/10.1093/nar/gky1016 ↩ ↩²

4. Elena Rojano, Pedro Seoane, Juan A G Ranea, et al. Regulatory variants: from detection to predicting impact. Briefings in Bioinformatics, 2018. https://doi.org/10.1093/bib/bby039 ↩

5. Žiga Avsec, Natasha Latysheva, Jun Cheng, et al. Advancing regulatory variant effect prediction with AlphaGenome. Nature, 2026. https://doi.org/10.1038/s41586-025-10014-0 ↩ ↩²

6. Wardah S. Alharbi, Mamoon Rashid. A review of deep learning applications in human genomics using next-generation sequencing data. Human Genomics, 2022. https://doi.org/10.1186/s40246-022-00396-x ↩

7. Tuuli Lappalainen, Alexandra J. Scott, Margot Brandt, et al. Genomic Analysis in the Age of Human Genome Sequencing. Cell, 2019. https://doi.org/10.1016/j.cell.2019.02.032 ↩

8. S. Pabinger, A. Dander, M. Fischer, et al. A survey of tools for variant analysis of next-generation genome sequencing data. Briefings in Bioinformatics, 2013. https://doi.org/10.1093/bib/bbs086 ↩

9. Jean-Pierre Desvignes, Marc Bartoli, Valérie Delague, et al. VarAFT: a variant annotation and filtration system for human next generation sequencing data. Nucleic Acids Research, 2018. https://doi.org/10.1093/nar/gky471 ↩

10. Ye Liu, William S. B. Yeung, Philip C. N. Chiu, et al. Computational approaches for predicting variant impact: An overview from resources, principles to applications. Frontiers in Genetics, 2022. https://doi.org/10.3389/fgene.2022.981005 ↩