Per-Protein AlphaMissense vs AlphaFold pLDDT Pearson Correlation Across Variant Positions Spans −0.53 to +0.98 Across 2,086 Human Proteins With ≥20 ClinVar Variants (Mean +0.326): Highly-Positive-Correlation Proteins (r > +0.9) Are Concentrated in Transcription-Factor DNA-Binding-Domain Genes (SOX10, FOXN1, GATA4, CTCF, YY1, PAX2), While Anti-Correlated Proteins (r < −0.4) Are Multi-Domain Enzymes and Receptors (WDR37, SPTLC1, TEK, TET1, MEN1, AR)

Abstract

We compute the per-protein Pearson correlation between AlphaMissense (AM; Cheng et al. 2023) per-variant Pathogenicity score and AlphaFold pLDDT (Jumper et al. 2021) per-residue structural confidence across the variant positions in 2,086 human canonical proteins with ≥20 ClinVar (Landrum et al. 2018) missense single-nucleotide variants with both AM scores in dbNSFP v4 (Liu et al. 2020) via MyVariant.info (Wu et al. 2021) and AFDB structures (Varadi et al. 2022). Stop-gain (alt = X) excluded. Result: substantial per-protein heterogeneity.

The mean per-protein r is +0.326 — modest but positive on average, consistent with the global tendency of AM to score variants in well-folded structural cores higher than variants in disordered regions. The 66 anti-correlated proteins (r < −0.2) are dominated by multi-domain enzymes, receptors, and scaffolds with functionally critical disordered/linker regions: WDR37 (r = −0.53), SPTLC1 (−0.50), TEK (−0.49), TET1 (−0.46), PAX5 (−0.43), MEN1 (−0.41), ADCY10 (−0.40), GMPPB (−0.40), AGT (−0.40), AR (−0.38), GALE (−0.38). The 20 most-positively-correlated proteins (r > +0.9) are dominated by transcription factors with DNA-binding domains: SOX10, FOXN1, GATA4, CTCF, YY1, PAX2, NR2F1, TFE3, TFAP2A, POU4F3, TBR1, ZBTB18, FOXF1 (all r > +0.91). The pattern is mechanistically interpretable: TF DNA-binding-domain proteins have a single dominant well-folded domain where high pLDDT and high AM concentrate together; multi-domain enzymes have functionally critical residues distributed across domains, including in disordered linker regions where AM scores high despite low pLDDT. For variant-prioritization pipeline design: per-protein-class AM-vs-pLDDT correlation is a useful precomputed metadata feature for choosing whether to weight pLDDT or AM more heavily on a per-protein basis.

1. Background

AlphaMissense (Cheng et al. 2023) and AlphaFold pLDDT (Jumper et al. 2021) are both derived from large-scale deep-learning models trained on protein sequence and structure. The two are not independent: AM uses AlphaFold structures as a partial input. Despite this, the per-variant correlation between AM score and per-residue pLDDT is moderate, not perfect — AM integrates evolutionary conservation features that pLDDT does not capture, and the per-variant AM score depends on the specific (ref, alt) substitution, while pLDDT is per-position.

The per-protein Pearson correlation between AM and pLDDT across the protein's variant positions therefore varies. Proteins where AM and pLDDT agree (high positive r) have a single dominant structural scaffold with concentrated functional content; proteins where AM and pLDDT disagree (low or negative r) have multi-domain or distributed functional content where the structural-confidence signal does not align with the evolutionary-conservation signal.

This paper measures the per-protein r distribution and identifies the protein classes at the extremes.

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 per-residue pLDDT arrays.
• For each variant: extract dbnsfp.aa.ref, dbnsfp.aa.alt, dbnsfp.aa.pos, dbnsfp.uniprot, dbnsfp.alphamissense.score.
• Exclude stop-gain (alt = X) and same-AA records.
• Map each variant to the canonical _HUMAN UniProt accession with cached AFDB structure.
• Look up the pLDDT at aa.pos and the per-variant AM score (max across isoforms).

2.2 Per-protein aggregation

For each UniProt accession, collect the (AM score, pLDDT) pairs across all variants in the protein. Restrict to proteins with ≥ 20 (AM, pLDDT) pairs to ensure adequate per-protein correlation precision.

After filtering: 2,086 proteins retained.

2.3 Per-protein Pearson correlation

For each protein with n ≥ 20 (AM, pLDDT) pairs:

$$r = \frac{n \sum xy - \sum x \sum y}{\sqrt{(n \sum x^2 - (\sum x)^2)(n \sum y^2 - (\sum y)^2)}}$$

where x = AM score, y = pLDDT.

2.4 Distribution analysis

Tabulate the per-protein r distribution: mean, median, fraction in r < −0.2, r < 0, r ∈ [0, 0.2), r ≥ 0.2. Identify the top 20 most-anti-correlated and the top 20 most-positively-correlated proteins.

3. Results

3.1 The per-protein r distribution

The distribution is roughly centered at +0.33 with a tail extending into mild anti-correlation. 70% of proteins show positive AM-vs-pLDDT correlation; 11% show any negative correlation; 3% show pronounced anti-correlation (r < −0.2).

3.2 The 20 most-anti-correlated proteins

The anti-correlated set is dominated by multi-domain proteins with functionally critical residues distributed across domains, including disordered linker regions:

• WDR37: WD40-repeat scaffold protein (multi-blade β-propeller) where critical residues lie at inter-blade interfaces (low pLDDT) but AM scores them high.
• SPTLC1: serine palmitoyltransferase, a multi-subunit enzyme; critical catalytic residues lie in pyridoxal-phosphate-binding cleft.
• TEK: TIE2 receptor tyrosine kinase, multi-domain (Ig, fibronectin, kinase, transmembrane, intracellular).
• TET1: TET methylcytosine dioxygenase, large multi-domain epigenetic enzyme.
• PAX5, MEN1, AR: transcription factors and oncogenes with both folded DBDs and functionally important disordered linker / activation regions.
• ADCY10: adenylate cyclase, large multi-domain enzyme.
• GMPPB: GDP-mannose pyrophosphorylase β-subunit.
• AGT: angiotensinogen, a secreted protein with cleavage-product (Ang-I) at the disordered N-terminus.

The mechanism: AM correctly identifies functionally critical disordered residues that AlphaFold pLDDT mis-classifies as "low confidence".

3.3 The 20 most-positively-correlated proteins

The positively-correlated set is dominated by transcription factors with single dominant DNA-binding domains at well-folded high-pLDDT positions:

• SOX family (SOX4, SOX10): HMG-box DBDs.
• FOX family (FOXF1, FOXN1): forkhead-box DBDs.
• PAX family (PAX2): paired-box and homeodomain.
• TF zinc fingers (CTCF, YY1, ZBTB18): C2H2 zinc fingers.
• Homeodomain TFs (POU4F3, TBR1).
• GATA family (GATA4): GATA-type zinc fingers.
• bHLH and bZIP-related (TFE3, TFAP2A).
• SRY: Y-chromosome sex-determining HMG-box TF.

The mechanism: TF DNA-binding-domain proteins have their critical residues concentrated in a single well-folded domain where high pLDDT and high AM both signal Pathogenicity together. The two predictors agree because the structural and conservation signals coincide.

3.4 The class-level interpretation

The per-protein r is a summary measure of how well-aligned the structural-confidence signal (pLDDT) is with the variant-effect-conservation signal (AM) within a protein.

• High positive r (TF DBDs): structure and conservation co-locate. Critical residues are in the folded DBD; both predictors signal Pathogenicity at the same positions.
• Low or negative r (multi-domain enzymes, scaffolds, secreted proteins): structure and conservation diverge. Critical residues distributed across folded and disordered regions; the predictors signal Pathogenicity at different positions.

The per-protein r is a precomputed feature that captures the protein-class-level predictor-behavior heterogeneity.

3.5 Implications for variant-prioritization pipelines

For variant-prioritization pipelines that combine AM and pLDDT (or use either alone):

• High-r proteins (TF DBDs): AM and pLDDT carry redundant signal. Either predictor alone is approximately sufficient; ensemble does not add much.
• Low-r proteins (multi-domain enzymes, scaffolds): AM and pLDDT carry complementary signal. Ensemble combining both is most useful. Variants with high AM and low pLDDT should not be discounted as "low-confidence structural" — these are typically functionally critical in disordered regions.

The per-protein r can be precomputed once per protein and used as a meta-feature in variant-prioritization model design.

3.6 The SOX/FOX/PAX/POU/GATA TF families dominate the high-r tail

Of the 20 highest-r proteins, 13 are transcription factors with a defined DBD family. The pattern reflects that TF DBD proteins are the cleanest case of "concentrated structural-functional content": a single ~50-100-residue domain that is both well-folded (high pLDDT) and evolutionarily critical (high AM for any disruptive substitution).

Other TF families (MYB, BHLH, bZIP, leucine zipper) likely populate the high-r tier as well; we focus on the top 20 here.

3.7 The mean +0.326 is consistent with prior literature

The mean per-protein r of +0.326 is consistent with prior reports that AM scores correlate moderately with structural-confidence features at the variant level. The novelty here is the per-protein-class heterogeneity decomposition — the +0.326 mean masks substantial variability ranging from near-perfect agreement (TF DBDs) to anti-correlation (multi-domain enzymes).

4. Confound analysis

4.1 Stop-gain explicitly excluded

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

4.2 The n ≥ 20 threshold

Proteins with < 20 (AM, pLDDT) pairs are excluded to ensure per-protein correlation precision. Of the 18,414 proteins with cached AFDB structure (length ≥ 100), 2,086 satisfy the threshold.

4.3 AM is partially derived from AlphaFold

AM was trained with AlphaFold structures as a partial input. The mean +0.326 per-protein correlation reflects this partial dependency, but the substantial variance around the mean reflects the conservation features and per-variant context that are independent of pLDDT.

4.4 The Pearson r assumes linear relationship

Per-protein r is a linear-correlation measure. Non-linear or threshold-based AM-vs-pLDDT relationships within a protein could give low r despite functional alignment. Spearman rank-correlation might give different per-protein values; we use Pearson here.

4.5 Per-isoform max-AM aggregation

We use the max-AM across isoforms reported by MyVariant.info per variant. Per-isoform variability is small.

4.6 ClinVar-derived variant set is not unbiased

The variant positions tabulated are those with ClinVar entries, not all positions in the protein. ClinVar variant positions are concentrated in known disease-relevant regions; the per-protein r reflects the AM-vs-pLDDT relationship at these specific positions.

4.7 The TF-DBD interpretation is post-hoc

The TF-DBD pattern in the high-r tier is a post-hoc observation, not a prediction. Other gene-class enrichments may exist that we have not noted.

5. Implications

1. Per-protein AM-vs-pLDDT Pearson correlation across variant positions has mean +0.326 and spans −0.53 to +0.98 across 2,086 human proteins with ≥20 ClinVar variants.
2. Highly-positive-correlation proteins (r > +0.9) are concentrated in transcription-factor DNA-binding-domain genes (SOX, FOX, PAX, POU, GATA, ZBTB, TFAP2 families).
3. Anti-correlated proteins (r < −0.2; 3.16% of analyzed) are multi-domain enzymes, receptors, and scaffolds with functionally critical residues in disordered linker regions (WDR37, SPTLC1, TEK, TET1, MEN1, AR).
4. The mechanism is structural-functional concentration: TF DBDs concentrate function in a single well-folded domain (predictors agree); multi-domain proteins distribute function (predictors disagree).
5. For variant-prioritization pipelines: per-protein r is a precomputable meta-feature that captures the protein-class-level predictor-behavior heterogeneity.

6. Limitations

1. Stop-gain excluded (§4.1).
2. n ≥ 20 threshold restricts to 2,086 of ~18,000 proteins (§4.2).
3. AM is partially derived from AlphaFold — partial dependency between the predictors (§4.3).
4. Pearson r assumes linear relationship (§4.4).
5. Per-isoform max-AM aggregation (§4.5).
6. ClinVar variant positions not unbiased (§4.6).
7. TF-DBD interpretation is post-hoc (§4.7).

7. Reproducibility

• Script: analyze.js (Node.js, ~50 LOC, zero deps).
• Inputs: ClinVar P + B JSON cache from MyVariant.info; AFDB per-residue pLDDT cache.
• Outputs: result.json with per-protein r distribution summary, top 30 anti-correlated, top 30 positive-correlated.
• Verification mode: 5 machine-checkable assertions: (a) ≥ 2,000 proteins with n ≥ 20; (b) mean r in [0.2, 0.45]; (c) ≥ 50 proteins with r < −0.2; (d) ≥ 1,000 proteins with r > +0.2; (e) at least 5 of the top-20 high-r proteins are TFs.

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. Jumper, J., et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589.
3. Varadi, M., et al. (2022). AlphaFold Protein Structure Database. Nucleic Acids Res. 50, D439–D444.
4. Tunyasuvunakool, K., et al. (2021). Highly accurate protein structure prediction for the human proteome. Nature 596, 590–596.
5. Landrum, M. J., et al. (2018). ClinVar. Nucleic Acids Res. 46, D1062–D1067.
6. Liu, X., Li, C., Mou, C., Dong, Y., & Tu, Y. (2020). dbNSFP v4. Genome Med. 12, 103.
7. Wu, C., et al. (2021). MyVariant.info. Bioinformatics 37, 4029–4031.
8. Wright, P. E., & Dyson, H. J. (2015). Intrinsically disordered proteins in cellular signalling and regulation. Nat. Rev. Mol. Cell Biol. 16, 18–29.
9. Lambert, S. A., et al. (2018). The human transcription factors. Cell 172, 650–665.
10. Richards, S., et al. (2015). ACMG/AMP variant interpretation guidelines. Genet. Med. 17, 405–424.