Variant Prioritization and Interpretation

Estimated time: 45 minutes

Combine the paper, GTEx, HuBMAP, and Pharos results while preserving what each source measures. Then use the combined table to choose a follow-up question.

Combine the results

The paper reports 54 variants across 25 genes. The analysis adds information from GTEx, HuBMAP, and Pharos to each variant based on its gene.

This information helps prioritize variants for experimental follow-up. The combined table records the classification reported for each variant.

Keep evidence types separate

Each source measures a different level of biology, even when the results appear in the same table:

Source What it provides How it informs follow-up
Paper fields The exact variant, phenotype, score, and class reported by the study Which reported variant and phenotype are being evaluated? The score records how the study selected variants for manual review.
GTEx Median gene expression reported in the selected reference heart tissues Which heart tissue is appropriate for follow-up?
HuBMAP Whether indexed values are available in up to the first 500 ventricular cardiac-myocyte records Are these cells suitable for follow-up, or is another atlas or experiment needed?
Pharos The current protein annotations and target development level Should follow-up begin with a known drug relationship, chemical probe, biological mechanism, or basic protein characterization?

Each column represents the measurement or annotation defined by its source. Together, the columns form a qualitative framework for experimental follow-up.

Candidate selection depends on interpreting the evidence together. The goal is a justified next step with each result tied to its source.

Build the combined table

Load the source data and API helpers

Load the 54 published variant rows and import the GTEx, HuBMAP, and Pharos wrappers.

One row represents one published observation with its gene, HGVS descriptions, study class, phenotype, and source. Each join must preserve all 54 rows.

Request gene-level data

Query all three APIs for the 25 genes. HuBMAP takes the longest because it checks each gene separately.

The dated teaching responses return 50 GTEx rows, one for each of 25 genes in two tissues, along with 25 HuBMAP rows and 25 Pharos rows. At this stage, every table describes gene-level information. The paper table supplies the individual variant rows.

Build one row per gene

Create one row per gene with two GTEx tissues, one HuBMAP cell type, and selected Pharos fields. This structure prevents extra variant rows during the final join.

The 25-row gene_results table contains GTEx median TPM, HuBMAP expression and availability, and Pharos tdl and drug_count. Column prefixes identify each source.

Join gene-level data to every variant

Match the 25 gene rows back to all 54 variant rows. Complete the join type that keeps every row from the published variant table.

Keep every row from the variant table, including variants with unavailable gene-level data.

# Join gene-level data by gene symbol.
combined_variants = variants.merge(
    gene_results,
    on="gene_symbol",
    how="left",
    validate="many_to_one",
)

# Confirm that all 54 variant rows remain.
assert len(combined_variants) == len(variants) == 54

# Inspect variant and gene-level fields together.
combined_variants.loc[
    :,
    [
        "subject_id",
        "gene_symbol",
        "hgvs_c",
        "study_class",
        "phenotype",
        "gtex_ventricle_tpm",
        "hubmap_availability",
        "tdl",
    ],
].head(10)

The combined table places each variant beside the corresponding GTEx, HuBMAP, and Pharos fields. A many-to-one join allows several variants to share gene data while preserving all 54 published rows.

Prioritize a focused DCM follow-up set

We focus on dilated cardiomyopathy, abbreviated DCM, and ventricular cardiac myocytes.

Define the follow-up question

Which DCM variant rows have HuBMAP values for ventricular cardiac myocytes, and how do they rank by GTEx left-ventricle expression?

How many rows meet these criteria? Which genes contribute P or LP findings? Which VUS occur in those genes?

The filter returns 15 of the 28 DCM rows. Ten are P or LP in the paper, and five are VUS.

The focused subset contains the 15 rows with a returned HuBMAP value for the selected cell type. The complete table includes the other 13 rows for clinical and biological interpretation.

Examine one missense variant with ProtVar

The paper classified TNNT2 c.776A>C (p.Asp259Ala) as a VUS. TNNT2 encodes cardiac troponin T, a component of the troponin complex that helps regulate calcium-dependent contraction in heart muscle.

Two other participants with DCM carried pathogenic or likely pathogenic TNNT2 variants. These findings add cohort-level information about TNNT2 and DCM. Additional variant-specific evidence can evaluate the effect of p.Asp259Ala.

The source study’s score incorporated PolyPhen, SIFT, REVEL, and CADD. In this follow-up analysis, ProtVar provides AlphaMissense and EVE results from newer missense-effect models and a FoldX estimate of stability change.

In the dated teaching response, AlphaMissense returns Pathogenic (0.8296), EVE returns Uncertain (0.575), and FoldX reports a ΔΔG of -0.369 kcal/mol. The pLDDT of 92.88 reports confidence in the local AlphaFold structure around residue 259.

Learn more about the returned protein scores in the ProtVar API documentation.

Request protein predictions

ProtVar maps p.Asp259Ala to UniProt accession P45379 at residue 259. Query the score endpoint for AlphaMissense and EVE, then query the FoldX stability endpoint.

GTEx shows high TNNT2 expression in heart tissue, and HuBMAP reports expression in ventricular cardiac myocytes.

Pharos reports 441 linked publications and 167 protein-interaction records for cardiac troponin T. This protein interacts with troponin I, troponin C, and tropomyosin to regulate thin-filament contraction.

One follow-up experiment would compare wild-type and D259A troponin complexes across calcium concentrations to test whether the substitution changes calcium-regulated thin-filament activity. The paper classified this variant as a VUS.

Interpret the integrated evidence

Interpret each result according to what its source measures.

Interpret the DCM follow-up set

The combined table records the paper’s classification for each variant. The API results help select genes and experimental systems for follow-up.

Birch and colleagues use a similar approach by combining variant classification, phenotype match, and molecular results while reporting the strength of each finding separately.

The label P/LP with cardiac expression measurements means that the paper classified a variant as pathogenic or likely pathogenic and both GTEx and HuBMAP returned relevant cardiac gene-expression values. These measurements help select an experimental model. Variant-specific experiments test the proposed mechanism.

Our interpretation Variants in dcm_follow_up What the evidence supports
P/LP with cardiac expression measurements DES c.735G>A; ACTC1 c.301G>A; TNNT2 c.547C>T; TNNT2 c.547C>G; MYBPC3 c.2490dup in two subjects; MYBPC3 c.442G>A; PLN c.25C>T; MYLK3 c.618dup; MYLK3 c.1569-2A>C The paper classified these variants as P/LP. The GTEx and HuBMAP measurements help identify a cardiac experimental model for their genes.
VUS with cardiac expression measurements TNNI3 c.337G>A; ACTC1 c.1132T>C; TNNT2 c.776A>C; FLNC c.4181A>G; TNNI3K c.827+1G>T The paper classified these variants as VUS. The expression measurements help formulate a cardiac follow-up hypothesis, and variant-specific experiments can test the effect of each allele.
NoteWhat the integrated results added

Variant analysis often leaves several credible candidate genetic variants. The next decision is which candidate genetic variant to investigate first and which experimental system fits the question.

GTEx provided cardiac-tissue measurements for all 25 genes. HuBMAP coverage provided the more selective filter.

Within this analysis, the results for DES, TNNT2, MYBPC3, ACTC1, PLN, and MYLK3 identify cardiac myocytes as a possible experimental model. The results also identify possible experimental tools and follow-up questions.

Compare variants within the same gene

All three TNNT2 rows receive the same GTEx, HuBMAP, and Pharos values. Yet the paper classified c.547C>T as P, c.547C>G as LP, and c.776A>C as VUS.

TNNT2 had a GTEx left-ventricle median of 2,896.66 TPM and a value above zero in 38.2% of retrieved HuBMAP ventricular cardiac-myocyte records. These results support a cardiac-cell model. Variant-specific testing can distinguish the alleles’ effects.

Identify possible experimental starting points

The paper describes MYLK3 as an emerging DCM gene and classified both variants as LP. MYLK3 had a GTEx left-ventricle median of 39.99 TPM and a value above zero in 13.4% of retrieved HuBMAP ventricular cardiac-myocyte records.

Pharos classifies the protein as Tchem and reports five ligands and one drug relationship. These records identify compounds to inspect before designing functional studies of MYLK3.

Pharos classifies TNNI3K as Tchem, and its chemical records may support experimental planning. In the retrieved HuBMAP records, TNNI3K had a value above zero in 0.4% of ventricular cardiac myocytes. Researchers could compare additional cardiac cell types or measure expression directly when planning TNNI3K follow-up.

Account for coverage and repeated observations

Coverage gaps and repeated variants affect which findings are prioritized.

Keep coverage gaps visible

Create a unique gene list of HuBMAP coverage gaps. These genes may require another atlas, another cell type, or direct measurement.

The full table includes the 13 DCM rows outside dcm_follow_up. These rows include P/LP findings in MYL3, TTN, GYG1, LMNA, and DMD, plus one LMNA VUS.

HuBMAP records these genes as unavailable for the selected cell type. Another atlas, cell type, or direct measurement could provide cell-level expression evidence.

Optional: Identify recurrent variants

Find exact variant observations that occur in more than one participant. The grouping uses both coding and protein HGVS fields to preserve distinct variants.

Three variants appear in more than one participant. MYBPC3 c.2490dup appears in three participants, while LMNA c.1304_1307dup and TTR c.323A>G each appear in two.

Recurrence describes this study cohort. Pathogenicity and population frequency come from their respective evidence sources.

Check your understanding

✔

How many rows should the table contain after gene-level data are joined back to the variants?

Correct. The join returns the 25 gene-level data rows to the full variant table while preserving all published variant records.

This is the number of gene-level data rows before they are joined back to the variants.

This is the number of represented subjects. The expected join size is 54 variant rows.

✔

Why is this a many-to-one join?

Correct. The table has 54 variant rows across 25 genes, so one gene-level data row may match multiple variants.

The relationship refers to multiple variant rows matching one gene-level data row. The number of GTEx tissues is a separate dimension.

Capstone activity: Complete an evidence record

Apply the workflow to a new candidate from the worked dataset. Keep variant-level and gene-level evidence separate.

Choose one row from dcm_follow_up other than TNNT2 c.776A>C. Complete a one-page evidence record with:

  1. The exact variant.
  2. The phenotype and study classification.
  3. One relevant GTEx result.
  4. One relevant HuBMAP result.
  5. One relevant Pharos result, plus a ProtVar result if the selected variant is missense and data are available.
  6. Additional evidence that would inform interpretation.
  7. A proposed experimental model or assay.
  8. One sentence explaining your prioritization decision.

Use this self-check before completing the activity:

  • Each claim names its source.
  • Gene-level and variant-level evidence are identified separately.
  • Additional evidence is stated explicitly.
  • The proposed experiment tests a specific hypothesis.

The paper supplies the study classification. The APIs supply gene, protein, and variant evidence that guides experimental follow-up while preserving that classification.

Key points

  • The many-to-one join preserves all 54 published variant observations.
  • In the dated teaching data, 15 DCM rows have indexed ventricular cardiac-myocyte values; 10 are P or LP in the paper and 5 are VUS.
  • Tissue expression, cell-type measurements, and protein information guide selection of candidate genetic variants and models while recording study classifications.
  • Comparing AlphaMissense, EVE, and FoldX for one VUS shows how variant-level predictions can refine the next experiment alongside the study class.

Next: Review the workflow and apply it to another expert-reviewed variant table.

View Jupyter notebook