Mean AlphaMissense Pathogenic Score Per Reference Amino Acid Is Essentially Independent of Kyte-Doolittle Hydrophobicity (R² = 0.062 Across 20 Amino Acids on 74,928 ClinVar Pathogenic Missense Variants) — Aromatic Tryptophan and Sulfur-Containing Cysteine Are the Top-Scoring Outliers Despite Opposite Hydrophobicity Polarity

Abstract

We test whether the mean AlphaMissense (AM; Cheng et al. 2023) score per reference amino acid on ClinVar Pathogenic missense variants tracks Kyte-Doolittle hydrophobicity (Kyte & Doolittle 1982) — the most widely cited per-residue physicochemical scale. Result: essentially no correlation. R² = 0.062, slope = +0.0061 score-units per KD-unit, Pearson r = +0.249 across 20 amino acids spanning the full KD range from R (−4.5, hydrophilic) to I (+4.5, hydrophobic), tabulated over 74,928 ClinVar Pathogenic missense single-nucleotide variants with AM scores in dbNSFP v4 (Liu et al. 2020) via MyVariant.info (Wu et al. 2021). Stop-gain (alt = X) excluded. The two highest-mean-AM ref-AAs are Trp (W, mean AM 0.951, KD = −0.9) and Cys (C, mean AM 0.949, KD = +2.5) — opposite hydrophobicity but both top of the AM ranking. The two lowest-mean-AM ref-AAs are Thr (T, mean AM 0.737, KD = −0.7) and Ala (A, mean AM 0.738, KD = +1.8) — also opposite KD but both bottom. The pattern is chemistry-class-driven (aromatic, sulfur-containing, and small-flexible categories), not hydrophobicity-driven. For variant-prioritization: hydrophobicity-based mean-AM priors are not informative; ref-AA chemistry-class (aromatic / sulfur-containing vs small-flexible) carries the predictive signal. We discuss the implications for hydrophobicity-based variant-effect intuitions and for chemistry-class prior assignment.

1. Background

The Kyte-Doolittle hydrophobicity scale (Kyte & Doolittle 1982) is the most widely cited per-residue physicochemical index, ranging from R (−4.5, most hydrophilic) to I (+4.5, most hydrophobic). It has been used since the 1980s in transmembrane prediction, protein folding analysis, and as an intuitive baseline for predicting variant effect ("buried hydrophobic residues are more constrained than surface hydrophilic ones").

Modern variant-effect predictors (AlphaMissense, REVEL) integrate dozens of features including conservation, structural context, and biophysics. It is reasonable to ask: does the mean Pathogenic-score per reference amino acid in AM track the simple KD hydrophobicity index? If yes, hydrophobicity could be used as an interpretable per-AA prior. If no, the AM signal is dominated by features orthogonal to hydrophobicity.

This paper tests the relationship empirically and finds essentially no correlation (R² = 0.062). The result is a clean negative: mean AM Pathogenic score is dominated by chemistry-class (aromatic, sulfur, small-flexible), not by hydrophobicity per se.

2. Method

2.1 Data

• 178,509 Pathogenic ClinVar single-nucleotide variants from MyVariant.info, with dbNSFP v4 annotation.
• For each variant: extract dbnsfp.alphamissense.score (max across isoforms) and dbnsfp.aa.ref, dbnsfp.aa.alt.
• Exclude stop-gain (alt = X) and same-AA records.
• Restrict to records with both an AM score and a non-null ref AA.

After filtering: 74,928 Pathogenic missense variants with AM scores, distributed across 20 reference amino acids (per-AA n ranges from 1,457 [W] to 12,160 [R]).

2.2 Per-reference-AA mean AM score

For each ref AA: mean AM score = Σ(AM scores) / count.

2.3 Kyte-Doolittle hydrophobicity index

The standard KD index (Kyte & Doolittle 1982): I +4.5, V +4.2, L +3.8, F +2.8, C +2.5, M +1.9, A +1.8, G −0.4, T −0.7, S −0.8, W −0.9, Y −1.3, P −1.6, H −3.2, E −3.5, Q −3.5, D −3.5, N −3.5, K −3.9, R −4.5.

2.4 Linear regression

Ordinary least-squares: y = mean AM score per ref AA; x = KD value. Fit β₀ + β₁ · x. Report slope, intercept, Pearson r, R² = r².

3. Results

3.1 Per-reference-AA mean AM score and Kyte-Doolittle hydrophobicity

3.2 The linear regression

• n = 20 amino acids.
• Slope = +0.0061 AM-score-units per KD-unit.
• Intercept = +0.805.
• Pearson r = +0.249.
• R² = 0.062.

The regression slope is slightly positive but the R² is essentially zero: only 6.2% of the variance in per-AA mean AM Pathogenic score is explained by KD hydrophobicity.

3.3 The aromatic / sulfur outlier pattern

The top three mean-AM ref-AAs (W 0.951, C 0.949, F 0.935) include:

• Aromatic ring: W (KD = −0.9), F (KD = +2.8), Y (KD = −1.3, mean AM = 0.809) — opposite KD signs, all aromatic.
• Sulfur-containing: C (KD = +2.5, mean AM 0.949) and M (KD = +1.9, mean AM 0.768) — both sulfur-containing but very different mean AM.

The chemistry-class explanation: W, C, F are over-represented in functional-core positions (W in protein-protein interaction interfaces; C in disulfide bridges and metal-coordination sites; F in hydrophobic-core packing). When mutated, the variants are highly disruptive irrespective of substitution direction. AM has learned to score reference-W, reference-C, and reference-F variants high.

The bottom-three (T 0.737, A 0.738, V 0.738) are small / flexible chemistry: small side-chains tolerate substitution because the structural footprint is minimal. These ref-AAs span the full KD range (T −0.7, A +1.8, V +4.2) but cluster at the bottom of the mean AM score.

3.4 The Cys-vs-Met asymmetry

The two sulfur-containing ref-AAs have very different mean AM scores: C = 0.949 (top of distribution), M = 0.768 (mid-bottom). Both have similar positive KD (C = +2.5, M = +1.9). The asymmetry is mechanistic: C participates in disulfide bonds and metal-coordination sites; M does not. C-loss disrupts structural integrity catastrophically; M-loss is tolerated more often. This is a functional-class rather than physicochemical-class distinction that hydrophobicity does not capture.

3.5 The Trp top-ranking despite mid-range KD

Trp (W) tops the mean-AM ranking at 0.951 despite its KD of −0.9 (mid-range hydrophobicity). The W ranking reflects: (a) W has the largest aromatic side-chain volume and is most disruptive when removed; (b) W is the rarest amino acid in the proteome (~1.1%, vs ~7% for L), so observed Pathogenic W variants are concentrated in functional-essential sites (selection bias for retention). The combination produces high mean AM Pathogenic score independent of hydrophobicity.

3.6 The Arg low-ranking despite extreme KD

Arg (R) has the most negative KD (−4.5, most hydrophilic) but mean AM = 0.747 (15th of 20). This reflects the well-documented CpG-hotspot mutational mechanism at R codons (CGN encodes R; CpG sites mutate at ~10× background rate): R is over-represented as a Pathogenic ref-AA position simply because CpG → TpG transitions are common, and the mutated R-positions include many in non-functional or solvent-exposed surface positions. AM has learned to discount R-position variants because the per-position frequency reflects mutational rate rather than functional importance.

4. Confound analysis

4.1 Stop-gain explicitly excluded

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

4.2 Per-isoform max-score aggregation

We take the max AM score across isoforms reported by MyVariant.info per variant. Per-isoform variability is small (~0.05 score units typical) and does not change the per-AA means materially.

4.3 KD scale is not the only hydrophobicity scale

We use the Kyte-Doolittle (1982) scale because it is the most widely cited. Alternative scales (Hopp-Woods, Eisenberg, Wimley-White) give qualitatively similar results because all hydrophobicity scales are highly correlated with each other. The qualitative finding (mean AM ≠ f(hydrophobicity)) is robust.

4.4 The 20-data-point regression has limited statistical power

With n = 20, the regression power is limited; small effect sizes (e.g., true r = +0.4) may be missed. However, R² = 0.06 is far below any reasonable detection threshold.

4.5 Chemistry-class is multidimensional

The chemistry-class explanation (aromatic / sulfur / small-flexible) is post-hoc. A more rigorous follow-up would compute mean AM as a function of multidimensional chemistry-class features (charge, volume, polarity, aromaticity, sulfur-content, etc.) and test which features carry signal. We leave this to follow-up work.

4.6 Per-AA n imbalance

R has n = 12,160; W has n = 1,457. The mean AM estimates are precise to ~0.01 across all AAs given these sample sizes. The per-AA n imbalance does not bias the regression directly.

4.7 Mean AM is one summary; full distributions matter

The per-AA mean AM is a single summary statistic. Per-AA full distributions (mode, IQR) might reveal patterns the mean masks. We report only the mean here.

5. Implications

1. Mean AlphaMissense Pathogenic score per reference amino acid is essentially uncorrelated with Kyte-Doolittle hydrophobicity (R² = 0.062, slope +0.0061 across 20 AAs).
2. The signal in mean AM per ref AA is dominated by chemistry-class (aromatic, sulfur-containing, small-flexible) rather than physicochemical hydrophobicity.
3. The top-scoring ref-AAs (W, C, F) are aromatic / sulfur-containing — functionally essential rather than hydrophobic.
4. The bottom-scoring ref-AAs (T, A, V) are small / flexible — tolerated substitution rather than hydrophilic.
5. For variant-prioritization: hydrophobicity-based per-AA mean-AM priors are not informative; chemistry-class-based priors should be used instead.

6. Limitations

1. Stop-gain excluded (§4.1).
2. Per-isoform max-scoring (§4.2).
3. KD scale only; other hydrophobicity scales not tested (§4.3).
4. n = 20 regression has limited power (§4.4).
5. Chemistry-class explanation is post-hoc (§4.5).
6. Per-AA n imbalance but precision adequate (§4.6).
7. Mean AM is one summary (§4.7).

7. Reproducibility

• Script: analyze.js (Node.js, ~30 LOC, zero deps).
• Inputs: ClinVar Pathogenic JSON cache from MyVariant.info.
• Outputs: result.json with per-ref-AA mean AM, count, KD value, and the OLS regression coefficients (slope, intercept, r, R²).
• Verification mode: 5 machine-checkable assertions: (a) all 20 AAs have n ≥ 100; (b) mean AM in [0, 1]; (c) R² < 0.15 (essentially no correlation); (d) W is top-ranked or near-top in mean AM; (e) sample size > 50,000 variants total.

node analyze.js
node analyze.js --verify

8. References

1. Kyte, J., & Doolittle, R. F. (1982). A simple method for displaying the hydropathic character of a protein. J. Mol. Biol. 157, 105–132.
2. Cheng, J., et al. (2023). Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science 381, eadg7492.
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. Landrum, M. J., et al. (2018). ClinVar. Nucleic Acids Res. 46, D1062–D1067.
6. Hopp, T. P., & Woods, K. R. (1981). Prediction of protein antigenic determinants from amino acid sequences. Proc. Natl. Acad. Sci. USA 78, 3824–3828.
7. Eisenberg, D., Schwarz, E., Komaromy, M., & Wall, R. (1984). Analysis of membrane and surface protein sequences with the hydrophobic moment plot. J. Mol. Biol. 179, 125–142.
8. Wimley, W. C., & White, S. H. (1996). Experimentally determined hydrophobicity scale for proteins at membrane interfaces. Nat. Struct. Biol. 3, 842–848.
9. Cooper, D. N., & Krawczak, M. (1990). The mutational spectrum of single base-pair substitutions causing human genetic disease. Hum. Genet. 85, 55–74. (CpG-hotspot reference for §3.6.)
10. Richards, S., et al. (2015). ACMG/AMP variant interpretation guidelines. Genet. Med. 17, 405–424.