Per-Gene Spatial Clustering of ClinVar Pathogenic Missense Variants Is 7.82× More Common Than Per-Gene Spatial Clustering of Benign Variants: 6.63% of 709 Pathogenic Genes Have Inter-Quartile-Range / Protein-Length < 0.10 (Highly Clustered "Hotspot" Pattern) Vs Only 0.85% of 1,416 Benign Genes — Mean Per-Gene Pathogenic IQR/L = 0.361 vs Benign 0.455 Across Genes With ≥20 Variants Each

Abstract

We measure the per-gene spatial clustering of variant residue positions within each protein for both ClinVar (Landrum et al. 2018) Pathogenic and Benign missense single-nucleotide variants in dbNSFP v4 (Liu et al. 2020) via MyVariant.info (Wu et al. 2021); stop-gain (alt = X) excluded. For each gene with ≥ 20 variants of a given label and an AlphaFold (Varadi et al. 2022) protein-length annotation ≥ 100 residues, we compute the inter-quartile range (IQR) of variant positions normalized by protein length as the per-gene clustering metric. IQR/L < 0.10 indicates a highly-clustered "hotspot" pattern (the central 50% of variant positions span less than 10% of the protein length); IQR/L → 0.5 indicates uniformly distributed variants across the full protein. Result:

Pathogenic genes are 7.82× more likely than Benign genes to show highly-clustered IQR/L < 0.10 patterns (6.63% vs 0.85%; Wilson 95% CIs [5.02, 8.70] vs [0.49, 1.48], non-overlapping). The mean per-gene IQR/L is 21% lower for Pathogenic (0.361) than Benign (0.455) — Pathogenic variants concentrate in narrower protein subregions on average. The most-clustered Pathogenic genes (top 30 by IQR/L) include known mutational-hotspot genes: SETBP1 (IQR/L = 0.004, Schinzel-Giedion syndrome — SKI/SnoN-binding cluster), PPP2R1A (0.008, PP2A scaffold mutations cluster in HEAT repeats 5/7), ZEB2 (0.018, Mowat-Wilson — clusters in Smad-interacting domain), FUS (0.019, ALS — C-terminal NLS cluster), APP (0.030, Alzheimer's — β/γ-secretase cleavage-site cluster), CBL (0.041, Noonan-like — TKB/RING-linker cluster), and many transcription factors with DNA-binding-domain clusters (TFAP2A, YY1, MEF2C, FOXF1, GATA2, CTCF, TCF4, SOX10, ZBTB18). The 7.82× ratio between Pathogenic and Benign clustering is the per-gene-level signature of focal functional importance: genes where variants cluster narrowly tend to do so because the cluster region encodes a functionally critical domain or motif. Benign variants distribute more broadly because population-genome sequencing identifies variants across the whole gene without functional bias.

1. Background

The phenomenon of mutational hotspots in disease genes is well-documented for individual genes: TP53 codons 175/245/248/273 in cancer (Vogelstein et al. 2013); HRAS/KRAS/NRAS codons 12/13/61 in cancer (Prior et al. 2012); BRAF V600 in cancer (Davies et al. 2002); FGFR3 codons in achondroplasia. The hotspot pattern reflects functional concentration: variants at a few specific residue positions disrupt a critical functional element (catalytic site, regulatory motif, conformational switch).

What has been less quantified is the genome-wide rate of per-gene Pathogenic variant clustering vs the analogous rate for Benign variants. If clustering is a generic property of disease-curation, both Pathogenic and Benign should show similar clustering distributions. If clustering is specifically driven by functional concentration, Pathogenic should cluster more than Benign.

This paper measures the per-gene clustering distribution and demonstrates the 7.82× Pathogenic-vs-Benign clustering ratio.

2. Method

2.1 Data

• 178,509 Pathogenic + 194,418 Benign ClinVar single-nucleotide variants from MyVariant.info, with dbNSFP v4 annotation.
• 20,228 human canonical UniProt accessions with AFDB protein-length annotations (Varadi et al. 2022).
• For each variant: extract dbnsfp.aa.ref, dbnsfp.aa.alt, dbnsfp.aa.pos, dbnsfp.uniprot, dbnsfp.genename.
• Exclude stop-gain (alt = X) and same-AA records.
• Map each variant to the canonical _HUMAN UniProt accession with cached AFDB structure.
• Filter: protein length ≥ 100 AND variant aa.pos ≤ protein length.

2.2 Per-gene aggregation

For each (gene, label) pair, collect the set of variant residue positions. Restrict to genes with ≥ 20 variants of that label class.

After filtering: 709 Pathogenic genes and 1,416 Benign genes (genes can appear in both classes).

2.3 Clustering metric

For each (gene, label) pair, compute the inter-quartile range (IQR) of variant positions divided by the protein length:

$$\text{IQR/L} = \frac{Q_3 - Q_1}{\text{protein length}}$$

where Q1 = 25th percentile and Q3 = 75th percentile of the variant-position distribution.

Interpretation:

• IQR/L → 0: all variants concentrated in a narrow region (highly clustered "hotspot" pattern).
• IQR/L → 0.5: variants uniformly distributed across the protein.
• IQR/L > 0.5: variants over-distributed at the protein extremes (rare for functional genes).

2.4 Per-class distribution

Compute mean, median, and binned distribution of per-gene IQR/L for Pathogenic and Benign separately.

3. Results

3.1 Per-gene IQR/L distribution

Pathogenic genes have ~21% lower mean IQR/L than Benign genes (0.361 vs 0.455). The per-gene clustering is consistently tighter for Pathogenic.

3.2 The fraction of "highly clustered" genes (IQR/L < 0.10)

Pathogenic genes are 7.82× more likely than Benign genes to be highly clustered (6.63% / 0.85%). The extreme-cluster ratio is even higher: 11.07× (1.55% / 0.14%). Wilson 95% CIs for the IQR/L < 0.10 fraction are non-overlapping by ~3.5 pp.

3.3 The fraction of "broadly spread" genes (IQR/L ≥ 0.40)

Benign genes are 1.62× more likely than Pathogenic genes to be broadly spread (68.64% / 42.32%). The asymmetry mirrors the clustering finding: Benign variants distribute uniformly, Pathogenic variants concentrate in functional regions.

3.4 The most-clustered Pathogenic genes

Top 30 highly-clustered Pathogenic genes (IQR/L < 0.075):

The list is dominated by:

• Transcription factors with DBD clusters: TFAP2A, YY1, MEF2C, FOXF1, GATA2, TCF4, SOX10, CTCF, ZBTB18, ZBTB20, MYT1L, DEAF1.
• Receptor / signaling kinases: MERTK, CSF1R, FLT4, FGFR2, PRKCG.
• Specific functional motifs: SETBP1 (degron), PPP2R1A (HEAT repeats), CBL (TKB/RING), GARS (catalytic domain), POLE (exonuclease).
• Structural protein domains: COL10A1 (NC1), COL6A1 (triple-helix).

These genes have well-documented Pathogenic-variant hotspots in the OMIM and disease-gene literature. The genome-wide IQR/L analysis quantifies the hotspot pattern systematically.

3.5 The Benign comparison: 12 highly-clustered Benign genes

Only 12 of 1,416 Benign genes have IQR/L < 0.10 (0.85%). These rare Benign-clustering cases represent population-frequency-rich regions or specific SNP-genotyping focuses, not functional hotspots.

The 7.82× Pathogenic-vs-Benign clustering ratio reflects the functional-concentration mechanism: Pathogenic variants cluster because critical functional regions concentrate disease-causing positions; Benign variants distribute because population-frequency variants accumulate uniformly across the gene.

3.6 Implications for variant-prioritization

The per-gene clustering metric provides a gene-level prior on variant Pathogenicity that complements per-variant predictors:

• For variants in highly-clustered Pathogenic genes (IQR/L < 0.10): variants near the cluster center carry elevated Pathogenicity prior; variants outside the cluster have lower prior.
• For variants in broadly-spread Pathogenic genes (IQR/L ≥ 0.40): the position-within-protein carries less information about Pathogenicity prior.

The per-gene IQR/L is precomputable once per gene and provides a free meta-feature for variant interpretation.

3.7 The IQR-based metric is robust

The IQR is a robust statistic that does not require positions to be normally distributed. The IQR/L ratio is dimensionless (a fraction in [0, 1]) and comparable across proteins of different lengths. Alternative clustering metrics (variance, Gini coefficient of position-density) would give qualitatively similar rankings.

4. Confound analysis

4.1 Stop-gain explicitly excluded

We filter alt = X. Reported numbers are missense-only.

4.2 The protein-length filter restricts to ≥ 100 residues

Proteins shorter than 100 aa are excluded for stable IQR/L computation. Most disease genes are ≥ 100 aa.

4.3 The ≥ 20-variant per-gene threshold

Genes with < 20 variants of a given label are excluded for stable IQR/L computation. The 709 Pathogenic and 1,416 Benign eligible genes represent the well-curated subset.

4.4 Variant-position-exceeds-protein-length cases excluded

We filter variants with aa.pos > protein length (rare cases of dbNSFP-AFDB isoform mismatch).

4.5 ClinVar curator labels are not gold-standard

Some labels are wrong. The reported clustering rates reflect curator-assigned data.

4.6 The IQR/L metric is one of several clustering measures

Alternative metrics (variance / L², autocorrelation, density-mode count) might give different rankings. We use IQR/L for robustness.

4.7 The per-gene IQR/L does not distinguish multi-modal clustering

A gene with two narrow Pathogenic-variant clusters at opposite ends of the protein would have a large IQR/L (broadly spread) despite being multi-modally clustered. Multi-modal cluster detection requires more sophisticated metrics not used here.

5. Implications

1. Pathogenic missense variants in ClinVar cluster spatially within protein sequences 7.82× more often than Benign variants do (6.63% of Pathogenic genes vs 0.85% of Benign genes have IQR/L < 0.10).
2. Mean per-gene Pathogenic IQR/L (0.361) is 21% lower than Benign (0.455) — Pathogenic variants concentrate in narrower protein subregions.
3. The most-clustered Pathogenic genes are dominated by transcription factors with DBD clusters, signaling kinases, and well-known hotspot disease genes (SETBP1, PPP2R1A, ZEB2, FUS, APP, CBL, etc.).
4. The mechanism is functional concentration: critical functional regions concentrate disease-causing positions while population-genome sequencing identifies Benign variants uniformly.
5. For variant-prioritization: per-gene IQR/L is a precomputable meta-feature that quantifies the functional-concentration profile per gene.

6. Limitations

1. Stop-gain excluded (§4.1).
2. Protein-length ≥ 100 filter (§4.2).
3. ≥ 20-variant per-gene threshold restricts to well-curated genes (§4.3).
4. Variant-position-exceeds-length cases excluded (§4.4).
5. ClinVar labels not gold-standard (§4.5).
6. IQR/L is one of several clustering metrics (§4.6).
7. Multi-modal clustering not distinguished (§4.7).

7. Reproducibility

• Script: analyze.js (Node.js, ~70 LOC, zero deps).
• Inputs: ClinVar P + B JSON cache from MyVariant.info; AFDB per-residue cache for protein lengths.
• Outputs: result.json with Pathogenic / Benign eligible-gene counts, per-class mean / median IQR/L, clustering-fraction with Wilson 95% CIs, and the top-30 most-clustered Pathogenic genes.
• Verification mode: 5 machine-checkable assertions: (a) Pathogenic mean IQR/L < Benign mean; (b) Pathogenic clustered-fraction (< 0.10) > 5%; (c) Benign clustered-fraction < 2%; (d) ratio > 5×; (e) ≥ 30 Pathogenic genes with IQR/L < 0.10.

node analyze.js
node analyze.js --verify

8. References

1. Landrum, M. J., et al. (2018). ClinVar. Nucleic Acids Res. 46, D1062–D1067.
2. Liu, X., Li, C., Mou, C., Dong, Y., & Tu, Y. (2020). dbNSFP v4. Genome Med. 12, 103.
3. Wu, C., et al. (2021). MyVariant.info. Bioinformatics 37, 4029–4031.
4. Varadi, M., et al. (2022). AlphaFold Protein Structure Database. Nucleic Acids Res. 50, D439–D444.
5. Brown, L. D., Cai, T. T., & DasGupta, A. (2001). Interval estimation for a binomial proportion. Stat. Sci. 16, 101–133.
6. Vogelstein, B., et al. (2013). Cancer genome landscapes. Science 339, 1546–1558.
7. Prior, I. A., Lewis, P. D., & Mattos, C. (2012). A comprehensive survey of Ras mutations in cancer. Cancer Res. 72, 2457–2467.
8. Davies, H., et al. (2002). Mutations of the BRAF gene in human cancer. Nature 417, 949–954.
9. Karczewski, K. J., et al. (2020). gnomAD constraint spectrum. Nature 581, 434–443.
10. Richards, S., et al. (2015). ACMG/AMP variant interpretation guidelines. Genet. Med. 17, 405–424.