Diretrizes para Submissão de Amostras
Why a High Variant Score Is Not a Research Conclusion
Variant prioritization is an evidence-integration problem, not a contest between prediction scores.
One genome may contain millions of differences from the reference sequence. Frequency and quality filters reduce that number, but the remaining candidates can represent very different biological questions: a de novo coding change, two variants acting under a recessive model, a cryptic splice event, a regulatory change in the relevant cell type, a structural event, or a mitochondrial variant whose abundance differs by tissue.
Genome-wide scores are valuable for ordering candidates. CADD, for example, integrates many annotations to estimate relative deleteriousness across coding and noncoding variants. Its developers also caution that no single cutoff is correct for every study and recommend using scores for ranking followed by further investigation. Review the CADD study.
The same boundary applies to newer AI predictions. A model may estimate the effect of a missense or splice change, but it does not automatically establish that the affected gene explains the observed phenotype, that the inheritance pattern is plausible, or that the event is active in the relevant tissue. We therefore keep each evidence family visible instead of collapsing the project into one unexplained score.
Questions a score cannot answer alone
- Was the variant called reliably in this sample?
- Is it compatible with the pedigree or cohort design?
- Does the gene match the specific phenotype?
- Is the transcript or regulatory element active here?
- Are supporting tools using overlapping information?
- What experiment would distinguish competing explanations?
Six Decisions We Help You Make
Every decision has a different evidence requirement and a different failure mode.
1. Is the candidate technically credible?
We review coverage, allelic balance, mapping context, strand and read-position patterns, call-set filters, sample identity, and the limits of the sequencing design. A biological ranking cannot repair an unreliable call. Zook and Salit describe why benchmark regions, variant representation, matching rules, and difficult genomic contexts must be stated when evaluating small-variant calls. Review the benchmarking guidance.
2. Does it fit the study phenotype?
We convert available features into structured terms, retain onset and absent findings when known, and compare candidate genes with curated human and model-organism phenotypes. Human Phenotype Ontology provides a computable vocabulary for this task, while Exomiser studies show how phenotype similarity can complement variant-level evidence. Review the ontology resource. Review the Exomiser evaluation.
3. Which inheritance model remains plausible?
We generate model-specific candidate lists rather than allowing dominant, recessive, X-linked, mitochondrial, mosaic, and somatic hypotheses to blur together. Family structure, sex, affected status, phase, penetrance, and cohort design define what can be inferred. A negative segregation result can lower a rank; missing relatives remain an uncertainty, not evidence against the candidate.
4. Is the proposed molecular effect credible?
We select predictors by variant class and avoid treating correlated tools as independent votes. ClinGen calibration work recommends calibrated score intervals and generally using one selected computational method for a given evidence criterion, rather than counting agreement among overlapping predictors. Review the calibration study.
5. Does the candidate survive family, cohort, and database checks?
We examine population frequency, ancestry context, internal recurrence, unrelated-case support, known gene–disease validity, literature dates, and database versions. A rare variant is not automatically important, and a database label is not timeless. The evidence matrix records where each claim came from and when it was retrieved.
6. What is the most informative next experiment?
We identify the uncertainty that controls the decision: call confirmation, phase, segregation, splice effect, transcript usage, regulatory activity, heteroplasmy, or cellular function. The next experiment is chosen because a positive or negative result would change the candidate rank, not because the assay is available.
An Evidence Model That Keeps Different Claims Separate
The final rank is traceable because each evidence layer has its own meaning, source, and uncertainty.
| Evidence layer | What we examine | What it can support | What it cannot prove alone |
|---|---|---|---|
| Technical validity | Read support, depth, alignment context, sample QC, caller output, orthogonal confirmation | The event is likely present in the tested material | Phenotype relevance or biological mechanism |
| Population evidence | Allele frequency, ancestry, homozygotes, constraint, internal controls | Compatibility with rarity and an assumed model | Causality simply because a variant is rare |
| Inheritance and segregation | Pedigree, affected status, parental origin, phase, mosaicism, penetrance | Compatibility with a defined genetic hypothesis | A complete conclusion when relatives are missing or phenotype is uncertain |
| Variant effect | Consequence, conservation, protein context, splice prediction, regulatory annotation | A plausible molecular impact that can guide testing | That the affected gene explains the study phenotype |
| Gene–phenotype evidence | Structured phenotypes, onset, known disease mechanisms, model-organism and pathway evidence | How well the candidate matches the observed biology | That the specific variant alters gene function |
| Experimental evidence | RNA, chromatin, segregation, targeted sequencing, cellular or perturbation assays | Direct support or contradiction for a defined mechanism | Broader generalization outside the tested tissue and system |
For each candidate, we provide the raw observations, transformed scores, source and version, supporting and conflicting evidence, missing evidence, and a scientist-written rationale. AI-assisted literature and annotation steps remain auditable; they do not replace review of the original record or paper.
Variant Classes and Research Scenarios Need Different Strategies
We build the analysis around the mechanism that the study can actually test.
Rare coding candidates in a family
For a trio or extended pedigree, we compare de novo, recessive, compound heterozygous, X-linked, and reduced-penetrance models. Missense tools such as REVEL or AlphaMissense can help rank protein-altering candidates, but the predictor is retained as one evidence type. REVEL was developed as an ensemble for rare missense variation, and AlphaMissense provides proteome-wide missense effect estimates. Review REVEL. Review AlphaMissense.
Suspected splice disruption
We assess canonical and nearby splice positions, deep intronic candidates, exon definition, transcript relevance, and predicted donor or acceptor changes. SpliceAI showed that sequence-based deep learning can identify cryptic splice effects, while the authors also noted the importance of transcript evidence in an appropriate tissue. Review the SpliceAI study.
Noncoding or regulatory candidates
We connect a variant to candidate promoters, enhancers, chromatin state, topological context, target genes, and phenotype. Genomiser combined regulatory scores, allele frequency, chromosomal domains, and phenotype relevance, illustrating why a noncoding score alone is incomplete. Review the Genomiser framework.
Structural, copy-number, or complex events
We examine breakpoint quality, gene dosage, orientation, repeat context, phase, affected regulatory regions, and consistency with the assay. Small-variant predictors are not reused for structural events. When the original experiment cannot resolve breakpoints or phase, the appropriate next step may be different sequencing rather than another ranking algorithm.
Mitochondrial variants
We consider maternal inheritance, haplogroup, heteroplasmy, tissue source, nuclear mitochondrial sequence interference, and variant-specific functional evidence. ClinGen mitochondrial specifications emphasize that heteroplasmy can differ between tissues and that no single threshold applies to all variants. Review the mitochondrial specifications.
Cohort, model-organism, and discovery studies
For unrelated cases, experimental lines, plants, animals, or microorganisms, we replace clinical-style assumptions with the available design: recurrence, burden, linkage, segregation, phenotype class, functional screen, or comparative conservation. The result is a research shortlist and evidence plan, not an individual medical report.
When Focused Experimental Evidence Should Change the Ranking
We recommend a new assay only after identifying the uncertainty it is designed to resolve.
| Evidence gap | Project-specific option | Why it can change the decision |
|---|---|---|
| The project needs a coding-variant discovery set, deeper coverage, or a matched model-organism comparison. | Human and mouse whole exome sequencing | Exome sequencing focuses the experiment on protein-coding regions and supports small-variant discovery. We define capture limits and do not treat uncovered or poorly covered regions as negative evidence. Review the variant benchmarking recommendations. |
| A shortlist needs focused confirmation, segregation, high-depth follow-up, or custom coverage of specific loci. | Targeted region sequencing | Targeted follow-up can test presence, allelic fraction, and family or cohort recurrence at predefined regions. It is most useful after candidates and decision thresholds are locked. The confirmation design preserves an independent test rather than reopening the initial ranking. Review the benchmarking principles. |
| A mitochondrial hypothesis depends on low-level heteroplasmy, tissue choice, or maternal segregation. | Mitochondrial DNA sequencing | Mitochondrial interpretation requires heteroplasmy and tissue context. The selected material and detection limit are stated because blood or one sampled tissue may not represent another tissue. Review the evidence basis. |
| A predicted splice or isoform effect lacks direct transcript evidence. | Full-length transcript sequencing | Full-length reads can connect exon usage and transcript structure within single molecules. The gene must be expressed in the sampled material; absence in an irrelevant tissue is not treated as refutation. Cummings et al. demonstrated how transcriptome sequencing can reveal aberrant expression and splicing that DNA-level analysis alone could not resolve. Review the transcriptome study. |
| A noncoding candidate is linked to a predicted regulatory element but its activity in the relevant cell state is unknown. | ATAC-seq | Chromatin accessibility can show whether a candidate region is open in the tested cells or condition. It supports regulatory context but does not by itself prove the variant changes target-gene expression. Buenrostro et al. established ATAC-seq as a genome-wide assay of open chromatin and regulatory structure. Review the ATAC-seq study. |
| Many candidate variants or genes require a scalable functional comparison in a defined cellular system. | CRISPR screen sequencing | Sequencing-based perturbation screens can rank functional effects under a prespecified phenotype. The cellular model, edit design, controls, and readout determine what the screen can support. Findlay et al. used saturation genome editing and sequencing to measure thousands of BRCA1 single-nucleotide variants, illustrating the power and model-specific limits of multiplexed functional evidence. Review the saturation editing study. |
These routes are alternatives, not a fixed bundle. A project may need none, one, or a staged combination. We define the candidate set, positive and negative controls, sample unit, replication, analysis threshold, and the result that would promote or demote each candidate before data generation.
A Traceable Workflow From Variant Set to Experimental Shortlist
One horizontal workflow keeps the evidence, decisions, and stopping points visible.
| Stage | What we do | Decision gate |
|---|---|---|
| 1. Define the research question | Lock the organism, phenotype, sample relationships, variant classes, inheritance or cohort hypotheses, and intended experimental decision. | What would count as a useful shortlist? |
| 2. Audit data and calls | Review reference build, sequencing design, sample QC, call quality, annotation compatibility, and known blind spots. | Which variant classes can the data support? |
| 3. Build model-specific candidate sets | Apply quality, frequency, region, inheritance, segregation, and cohort filters without merging incompatible hypotheses. | Which candidates survive each model? |
| 4. Integrate evidence | Add molecular-effect, phenotype, gene validity, pathway, tissue, literature, and database evidence with versions and links. | Which support is independent, conflicting, or missing? |
| 5. Review and rank | Compare candidates in an evidence matrix. Scientists inspect high-priority calls, model assumptions, and AI-assisted summaries. | Which candidates justify action now? |
| 6. Generate focused evidence | Confirm calls, test segregation, measure transcripts or chromatin, or run a functional experiment when the result can change rank. | Does the new evidence support, weaken, or leave the hypothesis unresolved? |
| 7. Report and version | Deliver ranked lists, evidence dossiers, exclusions, reproducible files, and a reanalysis plan tied to data and knowledge versions. | Advance, test, monitor, or stop |
What to Provide for a Defensible Prioritization Study
Good metadata can change the answer more than adding another prediction tool.
| Input category | Preferred materials | Why it matters |
|---|---|---|
| Research question | Organism, phenotype or assay endpoint, affected tissue, candidate mechanism, intended next decision, and known exclusions | Defines relevant evidence and prevents an unfocused ranking |
| Genomic data | FASTQ, aligned reads, VCF or gVCF, structural-variant files, reference build, caller versions, QC reports, and coverage summaries | Allows technical evidence and blind spots to be reviewed |
| Samples and relationships | Pedigree or cohort manifest, affected status, sex, parental or replicate relationships, tissue, time point, treatment, and batch | Supports inheritance, segregation, recurrence, and confounding checks |
| Phenotype and biology | Structured or narrative features, onset, absent findings, assay phenotypes, pathway hypotheses, and relevant model-organism observations | Connects gene and variant evidence to the actual study |
| Prior interpretation | Existing shortlist, filtering rules, reviewed variants, confirmation results, literature, database snapshots, and reasons for exclusion | Preserves previous knowledge and exposes differences in interpretation |
| Experimental materials | Available DNA, RNA, cells, tissues, family samples, model systems, and feasibility constraints | Determines which evidence gaps can be tested directly |
| Data-use limits | Permitted databases, secure-compute needs, reporting restrictions, and required output formats | Defines the analysis boundary and delivery plan |
Projects can begin from raw sequencing, a filtered call set, or an existing shortlist. When only processed results are available, we state which technical and inheritance checks cannot be repeated. Sample requirements for any new assay are provided after the mechanism and design are agreed.
Deliverables That Support the Next Research Decision
The deliverable is designed to be reviewed, challenged, updated, and used to plan experiments.
Ranked evidence package
- Model-specific candidate lists
- Variant and gene rankings
- Evidence matrix
- Supporting and conflicting findings
- Excluded-candidate log
Mechanism review
- Consequence and transcript context
- Phenotype and pathway fit
- Inheritance and segregation
- Population and cohort evidence
- Confidence and uncertainty
Action plan
- Call-confirmation priorities
- Focused experimental options
- Decision-changing outcomes
- Reproducible tables and figures
- Versioned reanalysis triggers
We can deliver variant-level dossiers, gene-level summaries, pedigree or cohort views, evidence heatmaps, pathway context, review-ready tables, and analysis files. Each candidate states why it is ranked, what would lower or raise its priority, and which conclusion is outside the available evidence.
Knowledge changes. New gene–phenotype relationships, database classifications, transcripts, population data, and functional studies can alter a shortlist without changing the original sequence. The report therefore records reference build, annotation sources, software and database versions, retrieval dates, and practical triggers for reanalysis.
References
- Rentzsch P, Witten D, Cooper GM, Shendure J, Kircher M. CADD: predicting the deleteriousness of variants throughout the human genome. Nucleic Acids Research. 2019.
- Pejaver V, Byrne AB, Feng BJ, et al. Calibration of computational tools for missense variant pathogenicity classification and ClinGen recommendations for PP3/BP4 criteria. The American Journal of Human Genetics. 2022.
- Ioannidis NM, Rothstein JH, Pejaver V, et al. REVEL: an ensemble method for predicting the pathogenicity of rare missense variants. The American Journal of Human Genetics. 2016.
- Cheng J, Novati G, Pan J, et al. Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science. 2023.
- Jaganathan K, Kyriazopoulou Panagiotopoulou S, McRae JF, et al. Predicting splicing from primary sequence with deep learning. Cell. 2019.
- Köhler S, Gargano M, Matentzoglu N, et al. The Human Phenotype Ontology in 2021. Nucleic Acids Research. 2021.
- Bone WP, Washington NL, Buske OJ, et al. Computational evaluation of exome sequence data using human and model organism phenotypes improves diagnostic efficiency. Genetics in Medicine. 2016.
- Smedley D, Schubach M, Jacobsen JOB, et al. A whole-genome analysis framework for effective identification of pathogenic regulatory variants in Mendelian disease. The American Journal of Human Genetics. 2016.
- McCormick EM, Lott MT, Dulik MC, et al. Specifications of the ACMG/AMP standards and guidelines for mitochondrial DNA variant interpretation. Human Mutation. 2020.
- Zook JM, Salit M. Best practices for benchmarking germline small-variant calls in human genomes. Nature Biotechnology. 2019.
- Vestito L, Jacobsen JOB, Walker S, et al. Efficient reinterpretation of rare disease cases using Exomiser. npj Genomic Medicine. 2024.
- Cummings BB, Marshall JL, Tukiainen T, et al. Improving genetic diagnosis in Mendelian disease with transcriptome sequencing. Science Translational Medicine. 2017.
- Buenrostro JD, Giresi PG, Zaba LC, Chang HY, Greenleaf WJ. Transposition of native chromatin for fast and sensitive epigenomic profiling of open chromatin, DNA-binding proteins and nucleosome position. Nature Methods. 2013.
- Findlay GM, Daza RM, Martin B, et al. Accurate classification of BRCA1 variants with saturation genome editing. Nature. 2018.
Example Genomic Variant Prioritization Report
The demo shows how candidate ranks change as technical, inheritance, phenotype, molecular-effect, and experimental evidence are added. It also shows why a candidate was excluded or retained.
A project-specific report may include call-quality review, pedigree-aware candidate lists, phenotype similarity, population evidence, transcript and protein context, splice or regulatory predictions, literature provenance, supporting and conflicting evidence, and an experimental decision table. Displayed values are tailored to the study and never presented as universal cutoffs.
AI-Assisted Genomic Variant Prioritization FAQs
1. Is this a clinical diagnostic service?
No. This Solution supports research prioritization and experimental planning. It does not provide an individual clinical diagnosis, treatment recommendation, or medical report.
2. Can you start from our existing VCF and phenotype notes?
Yes. We can begin with a VCF, pedigree or cohort manifest, phenotype notes, and prior shortlist. BAM or CRAM files and coverage information allow a stronger technical review. If raw or aligned reads are unavailable, that limitation is recorded.
3. Does AI decide which variant is causal?
No. AI can help organize annotations, compare phenotypes, retrieve literature, and rank evidence. Scientists review the source records, model assumptions, conflicts, and uncertainty. The output is a research shortlist, not an automated causal declaration.
4. Which variant types can be included?
Depending on the input data, we can examine single-nucleotide changes, small insertions and deletions, splice-region and noncoding candidates, mitochondrial variants, copy-number changes, and structural events. Each class uses different quality and effect evidence.
5. Do you use one fixed pathogenicity threshold?
No. Thresholds depend on the variant class, evidence calibration, study design, inheritance model, and intended action. We show the original score and its source, but do not treat one cutoff as sufficient across all projects.
6. Can you prioritize variants without a detailed phenotype?
Yes, but the result relies more heavily on frequency, inheritance, molecular effect, cohort recurrence, pathway, or experimental phenotype. We state how missing or broad phenotype information limits gene-level ranking.
7. How do you handle variants of uncertain significance?
We identify which evidence is missing or conflicting and rank possible next steps. A candidate is not promoted simply because several prediction tools agree. Transcript, segregation, regulatory, or functional evidence may be more informative.
8. Can you reanalyze a previously negative study?
Yes. We compare the original reference build, call set, phenotype terms, gene–disease knowledge, databases, transcripts, and filtering rules with current versions. Reanalysis cannot recover a variant that the original assay was unable to observe, so coverage and technology limits are reviewed first.
9. When is additional sequencing recommended?
Only when a specific evidence gap can change the decision. Examples include targeted confirmation, missing family members, insufficient mitochondrial depth, unresolved transcript structure, or a regulatory region not measured by the original design.
10. What files are delivered?
Typical outputs include ranked variant and gene tables, an evidence matrix, review notes, source and version records, visual summaries, excluded-candidate reasons, validation priorities, and a reanalysis plan. Exact formats are agreed at project intake.
Published Case Study
Independent Research Highlight
Reanalysis Turned a Large Unsolved Cohort Into a Focused Review Queue
This publication is an independent research example. It is not a CD Genomics customer project.
Background
Vestito et al. studied how phenotype-driven reanalysis could revisit previously unsolved whole-genome cases as disease–gene knowledge and variant classifications changed. The starting cohort contained 24,015 unresolved participants from the 100,000 Genomes Project who had already been assessed for variants in known disease genes.
Methods
The authors reran Exomiser with updated knowledge and evaluated stepwise filters designed to highlight newly relevant candidates. Their strategy used phenotype similarity, variant consequence, inheritance compatibility, updated gene–disease relationships, and an automated classification component. The analysis focused on reducing the number of candidates requiring manual reinterpretation while retaining known reanalysis findings.
Results
Reanalysis identified diagnoses in 463 participants, representing 2% of the unresolved cohort. The optimized strategy achieved 82% recall and 88% precision for highlighting new candidates and reduced the review burden to an average of one or two candidates per case. The study also showed that refreshed disease, gene, phenotype, and variant knowledge—not a new sequence experiment alone—can make an old dataset newly informative.
Why It Matters
A useful prioritization workflow must be versioned and repeatable. It should distinguish a candidate that was missed because knowledge changed from one that was never observable in the original assay. It should also produce a manageable, explainable review queue rather than a long list that transfers the entire burden to the researcher.
Conclusion
The study supports a practical design for research reanalysis: update knowledge systematically, preserve phenotype and inheritance context, use computation to reduce the search space, and retain expert review for the final interpretation.
Open-access note: The article is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License. This page summarizes the study without reproducing or adapting its figures.
Reference
- Vestito L, Jacobsen JOB, Walker S, et al. Efficient reinterpretation of rare disease cases using Exomiser. npj Genomic Medicine. 2024.
Selected Publications
These independent publications provide the methodological basis for this Solution. They are not presented as CD Genomics customer projects.
- Rentzsch P, Witten D, Cooper GM, Shendure J, Kircher M. CADD: predicting the deleteriousness of variants throughout the human genome. Nucleic Acids Research. 2019.
- Pejaver V, Byrne AB, Feng BJ, et al. Calibration of computational tools for missense variant pathogenicity classification and ClinGen recommendations for PP3/BP4 criteria. The American Journal of Human Genetics. 2022.
- Ioannidis NM, Rothstein JH, Pejaver V, et al. REVEL: an ensemble method for predicting the pathogenicity of rare missense variants. The American Journal of Human Genetics. 2016.
- Cheng J, Novati G, Pan J, et al. Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science. 2023.
- Jaganathan K, Kyriazopoulou Panagiotopoulou S, McRae JF, et al. Predicting splicing from primary sequence with deep learning. Cell. 2019.
- Köhler S, Gargano M, Matentzoglu N, et al. The Human Phenotype Ontology in 2021. Nucleic Acids Research. 2021.
- Bone WP, Washington NL, Buske OJ, et al. Computational evaluation of exome sequence data using human and model organism phenotypes improves diagnostic efficiency. Genetics in Medicine. 2016.
- Smedley D, Schubach M, Jacobsen JOB, et al. A whole-genome analysis framework for effective identification of pathogenic regulatory variants in Mendelian disease. The American Journal of Human Genetics. 2016.
- McCormick EM, Lott MT, Dulik MC, et al. Specifications of the ACMG/AMP standards and guidelines for mitochondrial DNA variant interpretation. Human Mutation. 2020.
- Zook JM, Salit M. Best practices for benchmarking germline small-variant calls in human genomes. Nature Biotechnology. 2019.
- Vestito L, Jacobsen JOB, Walker S, et al. Efficient reinterpretation of rare disease cases using Exomiser. npj Genomic Medicine. 2024.
- Cummings BB, Marshall JL, Tukiainen T, et al. Improving genetic diagnosis in Mendelian disease with transcriptome sequencing. Science Translational Medicine. 2017.
- Buenrostro JD, Giresi PG, Zaba LC, Chang HY, Greenleaf WJ. Transposition of native chromatin for fast and sensitive epigenomic profiling of open chromatin, DNA-binding proteins and nucleosome position. Nature Methods. 2013.
- Findlay GM, Daza RM, Martin B, et al. Accurate classification of BRCA1 variants with saturation genome editing. Nature. 2018.
