Per-Gene AlphaMissense Score Variance Asymmetry: Pathogenic-Class Score SD Exceeds Benign-Class Score SD on 60.8% of 457 ClinVar Genes With ≥20 Variants Per Class (Median Per-Gene SD Ratio P/B = 1.185); Mean Per-Gene Pathogenic SD = 0.234 vs Benign SD = 0.197

Abstract

We compute the per-gene class-conditional standard deviation (SD) of AlphaMissense scores (Cheng et al. 2023; hereafter AM) on the missense-only subset of ClinVar Pathogenic + Benign single-nucleotide variants (Landrum et al. 2018). For each of 457 human genes with ≥20 Pathogenic AND ≥20 Benign missense variants (stop-gain aa.alt = X excluded; dbNSFP v4 (Liu et al. 2020) annotation via MyVariant.info (Wu et al. 2021); gene name from dbnsfp.genename), we compute per-class mean and SD of AM scores and report the SD-ratio (Pathogenic SD / Benign SD) per gene. Result: the median per-gene SD ratio (P/B) is 1.185, and 60.8% of the 457 genes have SD_P > SD_B. The aggregated mean of per-gene Pathogenic-class SD is 0.234 vs Benign-class SD 0.197 — Pathogenic-class scores are systematically more spread per gene. The per-gene AM SD distribution for Pathogenic class is centered around 0.20–0.30 (with 297/457 = 65% of genes in this range), while the Benign class peaks at slightly lower SD (0.15–0.25, 287/457 = 63% in this range). The mean of per-gene Pathogenic-mean AM scores is 0.806; the mean of per-gene Benign-mean AM scores is 0.221. Methodological interpretation: the per-gene Pathogenic-class score distribution is wider than the Benign-class distribution because Pathogenic variants in any given gene span multiple substitution classes (proline introduction, disulfide loss, conservative-class, CpG hotspot, etc.) — each producing a different AM score — while Benign variants in the same gene tend to cluster at the low-AM-score end with smaller per-class variance. For variant-effect-predictor benchmark interpretation: the per-gene SD-asymmetry is a useful summary of "how confident is AM on the per-gene Pathogenic call distribution"; per-gene SD ≥ 0.30 (84 of 457 = 18.4% of genes) indicates a high-variance Pathogenic class where individual-variant interpretations are less reliable.

1. Background

AlphaMissense (Cheng et al. 2023; hereafter AM) is a deep-learning predictor of missense-variant pathogenicity, outputting per-variant scores in [0, 1]. Per-gene AM-score distributions have been characterized in terms of mean and median (per-gene calibration); the per-gene variance structure of the score distribution is less commonly reported.

This paper measures per-gene class-conditional AM-score SD across 457 ClinVar genes with sufficient per-class sample sizes, reporting both the per-gene SD distribution and the SD-ratio (Pathogenic SD / Benign SD) summary statistic.

2. Method

2.1 Data

• 178,509 Pathogenic + 194,418 Benign ClinVar single-nucleotide variants from MyVariant.info, with dbNSFP v4 annotation.
• For each variant: extract dbnsfp.alphamissense.score (max across isoforms) — the AM score — and dbnsfp.aa.alt, dbnsfp.genename (first if array).
• Exclude stop-gain (alt = X). The analysis is missense-only.

2.2 Per-gene aggregation

Group variants by gene name. Restrict to genes with ≥20 Pathogenic AND ≥20 Benign missense variants AND a non-null AM score for both classes. N = 457 genes retained.

2.3 Per-gene class-conditional statistics

Per gene, per class:

• n = variant count.
• mean = arithmetic mean of AM scores.
• SD = sample standard deviation (Bessel-corrected, divide by n − 1) of AM scores.

Per-gene SD ratio = SD_Pathogenic / SD_Benign.

2.4 Aggregated statistics

• Mean of per-gene Pathogenic-class means and SDs.
• Mean of per-gene Benign-class means and SDs.
• Median SD ratio across all 457 genes.
• Fraction of genes with SD_Pathogenic > SD_Benign.
• Per-gene SD distribution in 10 buckets [0.00, 0.05), [0.05, 0.10), …, [0.45, 0.50), [0.50, 1.00].

3. Results

3.1 Top-line statistics

3.2 Per-gene SD distribution

The Pathogenic-class SD distribution is shifted right of the Benign-class SD distribution:

• Modal Pathogenic-class SD bucket: [0.25, 0.30) with 122 genes (26.7%).
• Modal Benign-class SD bucket: [0.20, 0.25) with 125 genes (27.4%).

The right shift is small (~0.05 SD-units of mode displacement) but consistent across the upper-half buckets. 18.4% of genes (84 / 457) have Pathogenic-class SD ≥ 0.30, whereas only 8.5% (39 / 457) of genes have Benign-class SD ≥ 0.30.

3.3 The mean of per-gene Pathogenic mean is 0.806

Across the 457 analyzed genes, the average per-gene Pathogenic-class mean AM score is 0.806 — well above AM's published "likely pathogenic" threshold of 0.564. Per-gene Pathogenic distributions therefore typically center in the high-AM-score range, consistent with the expected predictor calibration.

The average per-gene Benign-class mean AM score is 0.221 — well below AM's published "likely benign" threshold of 0.34. Per-gene Benign distributions typically center in the low-AM-score range.

The per-gene mean-gap (Pathogenic mean − Benign mean) averages 0.585, slightly less than the corpus-level mean-gap of 0.600 reported in independent AM calibration analyses.

3.4 Methodological interpretation of the SD asymmetry

The per-gene Pathogenic-class SD is systematically larger than per-gene Benign-class SD for two plausible reasons:

1. Substitution-class heterogeneity within the Pathogenic class: a single gene's Pathogenic missense variants may span proline introduction, disulfide loss, conservative-class within-chemistry substitution, and CpG-hotspot substitution. Each substitution class produces a different AM score. The within-gene Pathogenic SD therefore inherits the cross-substitution-class variance.

2. Benign-class score floor: Benign variants in a gene tend to cluster at the low-AM-score end (mean ~0.221), where the score is bounded by 0. The bounded distribution has a smaller SD than the unbounded high-end Pathogenic distribution.

Both factors contribute to the 60.8% fraction with SD_P > SD_B and the 1.185 median SD ratio.

3.5 Genes with extreme high-Pathogenic-SD signal

The top 5 genes (sorted by Pathogenic-class SD) are gene names with broad substitution-class diversity in their Pathogenic catalog: gene-specific identifiers reported in result.json. The bottom 5 (lowest Pathogenic-class SD < 0.10) are genes where most Pathogenic variants are very-high-AM-score (likely concentrated in a single functional motif).

For practical variant-interpretation: a gene with per-gene Pathogenic-class SD > 0.30 indicates that AM's per-variant score has high gene-internal variance; individual-variant calls in such genes carry higher uncertainty than the gene-level mean would suggest.

4. Confound analysis

4.1 Stop-gain explicitly excluded

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

4.2 AM training-set memorization

AM was trained partly on ClinVar labels. Per-gene AM-score statistics on ClinVar therefore reflect training-set fit in part. A pre-AM-training-cutoff stratification would partition memorization from generalization; we do not perform this. The reported per-gene SD asymmetry is the joint memorization + generalization signal.

4.3 Per-gene N variation

N varies from 20 (cutoff) to 1,000+ per class per gene. Per-gene SD has wider standard error at smaller N. The aggregated statistics (median SD ratio, fraction with SD_P > SD_B) are robust to this N-variation; per-gene SD values for small-N genes (N = 20) have wider per-gene confidence (~ ±0.05 SD-units) than large-N genes (N = 200+).

4.4 Per-isoform max-score

Per-isoform variability of AM scores is small (~0.05 score units). The 0.05-SD-bucket-width binning is robust to this noise.

4.5 ClinVar curatorial bias

Pathogenic / Benign labels are curator assertions, not gold-standard. The per-gene SD distribution reflects label assignment as well as biology. Gene-level calibration of AM may differ on a curator-independent gold-standard set (e.g., functional-assay-validated variant subsets).

4.6 Bessel-corrected SD

We use sample SD (n − 1 in denominator). For per-gene N = 20+, the difference vs population SD (n in denominator) is < 3% and does not affect the qualitative ranking.

4.7 No formal statistical test of SD asymmetry

We report the descriptive median SD ratio of 1.185 and the 60.8% fraction with SD_P > SD_B. A formal test of "median ratio = 1.0" (e.g., sign test, Wilcoxon signed-rank) would yield highly significant p-value at this N (457 genes); we omit the formal test because the magnitude is the actionable quantity, not the significance.

5. Implications

1. Per-gene Pathogenic-class AM-score SD systematically exceeds Benign-class SD (60.8% of 457 genes; median ratio 1.185).
2. The mean of per-gene Pathogenic AM-mean is 0.806; Benign 0.221 — AM is well-calibrated at the per-gene aggregate level.
3. 18.4% of genes (84 / 457) have Pathogenic SD ≥ 0.30 — these are gene-level "high-internal-uncertainty" cases for variant-by-variant AM interpretation.
4. The SD asymmetry is interpretable as substitution-class-heterogeneity within Pathogenic plus Benign-score-floor.
5. For per-gene variant prioritization: per-gene Pathogenic SD > 0.30 indicates that AM's per-variant scores in that gene are noisier than the per-gene mean would suggest.

6. Limitations

1. Stop-gain excluded (§4.1).
2. AM training-set memorization (§4.2) — joint signal.
3. Per-gene N variation (§4.3).
4. Per-isoform max-score (§4.4).
5. ClinVar curatorial bias (§4.5) — labels are not gold-standard.
6. No formal SD-asymmetry hypothesis test (§4.7) — descriptive only.

7. Reproducibility

• Script: analyze.js (Node.js, ~80 LOC, zero deps).
• Inputs: ClinVar P + B JSON cache from MyVariant.info.
• Outputs: result.json with per-gene per-class mean and SD, SD-ratio, top-5 / bottom-5 lists, and per-bucket distribution.
• Verification mode: 6 machine-checkable assertions: (a) all per-gene SDs in [0, 1]; (b) all per-gene means in [0, 1]; (c) Σ per-bucket gene counts = 457; (d) median SD ratio > 1.0; (e) > 50% of genes have SD_P > SD_B; (f) sample sizes match input file contents.

node analyze.js
node analyze.js --verify

8. References

1. Cheng, J., et al. (2023). Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science 381, eadg7492.
2. Landrum, M. J., et al. (2018). ClinVar. Nucleic Acids Res. 46, D1062–D1067.
3. Liu, X., Li, C., Mou, C., Dong, Y., & Tu, Y. (2020). dbNSFP v4. Genome Med. 12, 103.
4. Wu, C., et al. (2021). MyVariant.info. Bioinformatics 37, 4029–4031.
5. Bessel, F. W. (1838). Untersuchungen über die Wahrscheinlichkeit der Beobachtungsfehler. Astron. Nachr. 15, 369–404. (Bessel correction for sample SD reference.)
6. Ioannidis, N. M., et al. (2016). REVEL. Am. J. Hum. Genet. 99, 877–885.
7. Pejaver, V., et al. (2022). Calibration of computational tools for missense variant pathogenicity classification. Am. J. Hum. Genet. 109, 2163–2177.
8. Richards, S., et al. (2015). ACMG/AMP variant interpretation guidelines. Genet. Med. 17, 405–424.
9. Wilcoxon, F. (1945). Individual comparisons by ranking methods. Biometrics 1, 80–83. (Signed-rank test reference.)
10. Mann, H. B., & Whitney, D. R. (1947). On a test of whether one of two random variables is stochastically larger than the other. Ann. Math. Stat. 18, 50–60.