# Do Neanderthal genomes carry a DNA-repair defect?

**Revision status, 12 September 2026:** The variant calls below are archived computational observations. The [revised article](../) and [revision record](astra-revision-2026-09-12.md) supersede the earlier functional and rate conclusions. No L580P functional experiment was performed. The Denisovan discussion below remains exploratory; missing independent coverage does not refute a variant, and mutation spectra do not authenticate or reject every individual call.

**Task.**  Eric asked for a computational check of the proposed repair defect.  Marshall Jordan's 2024 *Answers Research Journal* paper proposes that Neanderthals "inherited, early in their descent from Ham, a defect in one of their DNA repair mechanisms," and names Werner syndrome (the *WRN* helicase) as his model.  Robert Carter's 2009 version names the mitochondrial DNA polymerase directly and says the right thing about it: "we can measure the fidelity of DNA polymerases including the human mitochondrial DNA polymerase, and we know that mutations in DNA polymerases can elevate error rates in human mitochondria." The search below addresses specified public exome data; it does not establish that no one else has investigated the question.

**Date:** 2026-08-25.  Companion to `neanderthal-hypermutation-evidence.md`, which measures the Neanderthal mtDNA branch itself.

---

## 1.  The answer in three sentences

The scanned catalogues contain no annotated stop-gained, stop-lost, or essential-splice single-nucleotide changes in the 206-gene repair/maintenance panel; exome capture reached 195 genes. This does not exclude indels, regulatory changes, incomplete coverage, or altered function from missense variants.  Across the same panel they carry amino-acid changes at exactly the rate chance predicts: 15 of their 924 lineage-specific coding changes fall in a panel that holds 1.32 % of the exome, against 11 of our 769.  Neither lineage is enriched (P = 0.25 and P = 0.44).

The scan also identified the POLG L580P candidate. Its functional effect remains unmeasured.

---

## 2.  POLG p.Leu580Pro — the one real find

**What it is.**  A derived amino-acid change in *POLG*, the catalytic subunit of DNA polymerase gamma, the only polymerase that replicates mitochondrial DNA.  GRCh37 chr15:89,868,891 A→G.

**Who carries it**, from the MPI-EVA combined exome VCF (three Neanderthals, the Denisovan and nine present-day human exomes in one file):

| individual | genotype | depth |
|---|---|---|
| Altai Neanderthal | **1/1** homozygous derived | 39 reads, all G, GQ 99 |
| Vindija Neanderthal | **0/1** heterozygous | 18 reads, 8 A / 10 G, GQ 99 |
| El Sidrón Neanderthal | no call | 0 |
| Denisovan | 0/0 ancestral | 21 |
| nine present-day humans | 0/0 ancestral | 13–32 each |

**How good the evidence is.**

- The human–chimp, human–gorilla and human–orangutan ancestral alleles are all A, so G is the derived state and the Neanderthals carry it.
- The site is conserved: **mammal conservation score 0.992, GERP 4.99.**
- The archived annotation reports gnomAD exome allele frequency **6.9 × 10⁻⁷** (rs1463396518). This is a chromosome frequency, not a carrier frequency. Recover the database version and allele-count denominator before quoting it as a current estimate.
- **It is outside the usual cytosine-deamination classes.** A→G is not C→T or G→A. The archived Altai call has 39-fold coverage and strand support, but other errors still require read-level and independent-call validation.
- Ensembl VEP calls it **possibly damaging (PolyPhen 0.758)**; SIFT calls it **tolerated (0.48)**.  A split prediction.

**Where residue 580 sits, which is the interesting part.**  UniProt P54098 annotates residue **579 as an RNA-primer binding residue** — Leu580 is its immediate neighbour.  Every natural variant UniProt records between residues 570 and 600 is a disease allele:

| residue | change | disease |
|---|---|---|
| 574 | R→W | autosomal-recessive progressive external ophthalmoplegia |
| **579** | R→W | autosomal-recessive progressive external ophthalmoplegia |
| **580** | **L→P (the Neanderthal allele)** | — |
| 587 | P→L | PEO, and mitochondrial DNA depletion syndrome 4A/4B (Alpers–Huttenlocher) |

And the substituted residue is a proline, which is the residue most likely to break a backbone conformation.

**What remains unresolved.**

1. Branch counts do not isolate this variant’s effect. Their annual-rate interpretation depends on specimen dates; the former unconditional 1.36× ceiling was incorrect.
2. Residue 580 lies in the spacer region. Being outside the exonuclease active site does not exclude effects on binding, movement, or the balance between synthesis and proofreading.
3. Nearby disease variants and structural predictions do not establish whether L580P changes point-mutation fidelity, mtDNA copy number, deletions, or none of these.
4. POLG variants are not uniformly recessive. Vindija’s heterozygous call cannot be treated as proof of no effect.

Report L580P as an untested candidate. Computational checks can verify calls and prioritize experiments; a quantitative fidelity claim needs functional evidence. The main article describes existing-data searches and purified-enzyme assays.

---

## 3.  The systematic search

### 3a.  Method

Two independent passes over the Max Planck Institute for Evolutionary Anthropology archaic-genome resources, against a hand-built **206-gene panel** covering mitochondrial DNA replication and maintenance, base-excision repair (the pathway that actually operates on mitochondrial DNA), nucleotide-excision repair, mismatch repair, double-strand-break repair, the progeroid helicases including *WRN*, Fanconi anaemia, replicative polymerases and proofreading, translesion synthesis, and general genome stability.  Gene list in `code/repair-genes/genes.py`.

**Pass 1 — the fixed-difference catalogs.**  Castellano et al. 2014 published, with the three-Neanderthal exome paper, VEP-annotated catalogs of every amino-acid change fixed on the Neanderthal lineage, on the shared archaic lineage, and on the modern human lineage, each with PolyPhen-2, SIFT, GERP and 1000 Genomes frequencies.  Scanned in `code/repair-genes/scan_catalog.py` and `scan_all.py`.

**Pass 2 — the proteins themselves.**  The same FTP site carries the translated protein sequence of every captured transcript in all three Neanderthal exomes.  Each was aligned residue-by-residue against the Ensembl GRCh37 reference protein for the same transcript (`code/repair-genes/protein_diff.py`), which catches changes the catalog's fixed-difference filter drops — heterozygotes, single-individual variants, and anything not in a CCDS transcript.  This is how POLG L580P was found.

Candidates were then verified against the actual genotype data: the Altai high-coverage genome VCF by remote tabix, and the combined exome VCF that carries all three Neanderthals, the Denisovan and nine present-day humans in one file with the great-ape ancestral alleles in its INFO field (`code/repair-genes/pull_exome_vcf.sh`, `verify2.py`).

### 3b.  Result 1 — no matching annotated loss-of-function SNVs

**Zero** stop-gained, stop-lost or essential-splice changes in the 206-gene panel, on the Neanderthal lineage, the shared archaic lineage, or ours.

That zero means something, because those classes are not absent from the catalogs: genome-wide they contain 47 Neanderthal-specific, 19 archaic-specific and 29 modern-human-specific loss-of-function changes.  The panel simply has none.

One honest limit: these catalogs are single-nucleotide only.  A frameshift mutator allele would not appear.

### 3c.  Result 2 — no enrichment, measured against a real null

The 195 panel genes that the exome capture reached hold **1.32 %** of its coding residues (145,049 of 10,966,163, computed from the capture's own protein set).  So under neutrality that is the fraction of lineage-specific amino-acid changes expected to land in the panel.

| lineage | changes in panel | total | share | expected | P(≥ observed) |
|---|---|---|---|---|---|
| Neanderthal-specific | 15 | 924 | 1.62 % | 1.32 % | 0.247 |
| modern-human-specific | 11 | 769 | 1.43 % | 1.32 % | 0.439 |
| archaic-specific (Nea + Den) | 3 | 410 | 0.73 % | 1.32 % | — |

Neanderthal against modern human directly: Fisher exact **p = 0.844**.  The repair machinery took its fair share of change on both lineages and no more.

### 3d.  Result 3 — the core mitochondrial machinery is untouched, and it was well covered

The protein-level diff found **no amino-acid difference at all** between any of the three Neanderthals and the human reference in:

*TWNK* (Twinkle helicase), *MGME1*, *SSBP1*, *TFAM*, *TOP3A*, *LIG3*, *RRM2B*, *MPV17*, *SLC25A4*, *TK2*, *PIF1*, *TFB1M*, *MFN2*, *ELAC2*, *FBXL4*, *MTPAP*, *YME1L1*, *SUCLG1*, *AFG3L2* — and, in base-excision repair, ***OGG1***, ***UNG***, *NEIL1*, *APTX*, *EXOG*, *FEN1*, *MPG*, *SMUG1*, *TDG*, *POLB*, *PNKP*, *PARP1*, *PARP2*, *NUDT1*.

Coverage in those genes is good, so the zero is informative — for example *TWNK* 97 %/94 %/86 % of residues called across Altai/Vindija/El Sidrón, *MGME1* 94 %/94 %/81 %, *OGG1* 94 %/90 %/68 %, *LIG3* 96 %/93 %/81 %, *POLG* 95 %/79 %/61 %.  Spot checks in the Altai genome VCF confirm it independently: *OGG1* homozygous reference at 35-fold, *TWNK* at 37-fold, *MGME1* at 60-fold.

**Where nothing can be concluded**, because coverage is too thin: *RNASEH1* (12 %), *POLRMT* (28 %), *SPG7* (24 %), *SLC25A33* (19 %), *RAD17* (18 %), *ERCC3* (28 %), *XRCC6* (32 %), *TDG* (40 %), and *DDX11*, *PMS2*, *POLD2*, *GTF2H2* (0 %).  *POLRMT* is the one that matters — it is the mitochondrial RNA polymerase, and it also lays down the replication primer.

Three genes gave frame or alignment mismatches between the 2014 transcript build and the current reference and were discarded rather than reported: *BRCA1*, *TIMELESS*, *RAD23A*.

### 3e.  Result 4 — every other candidate dissolves on inspection

Each was checked for its genotype in all three Neanderthals, its ancestral state from the great-ape alignment, its frequency in living humans, and its conservation.

| change | Neanderthals | humans | conservation | prediction | verdict |
|---|---|---|---|---|---|
| ***WRN* p.Leu1074Phe** | all with calls 1/1 | **rs1801195, 49.6 % of human chromosomes**; the Denisovan and 5 of 9 human exomes carry it too | mammal 0.000 | benign | **Not a Neanderthal change at all.**  The ancestral allele is T; the human *reference genome* carries the derived allele and half of humanity does not.  Jordan's named gene, and the "difference" is one of the commonest polymorphisms in the human genome. |
| *WRN* p.His760Pro | Altai and Vindija 1/1 | gnomAD 6.8 × 10⁻⁷ | mammal 0.583, **GERP −1.31** | **benign (0), tolerated** | Archaic-restricted but at an unconserved site with no predicted effect. |
| *SLX4* p.Arg204Cys | all three 1/1 | **6.0–6.8 % of humans**; 1 of the 9 human exomes is a heterozygote | mammal 0.959 | probably damaging (0.999) | The highest PolyPhen score in the whole panel, and one person in eight carries it.  A useful calibration on what these predictions are worth. |
| *POLG2* p.Ala169Thr | Altai 1/1 | **rs1427463, 29 %**; 2 of the 9 human exomes are homozygous | mammal 1.000 | benign | Ordinary human polymorphism. |
| *DNA2* p.Phe60Ser | Altai 1/1 | **rs1535539, 25 %** | — | benign | Ordinary human polymorphism. |
| *PRIMPOL* p.Gln168Arg | all three | **rs2463447 — the derived allele is 96.7 % of humans**; the reference carries the minor allele | — | benign | Neanderthals carry what nearly everyone carries. |
| *PRIMPOL* p.Thr505Lys | all three | rs14969, 14.6 % | — | benign | Ordinary human polymorphism. |
| *MUTYH* p.Arg123Trp | **Altai 1/1 only**; Vindija 0/0 at 66-fold, El Sidrón 0/0 at 20-fold | gnomAD 2.7 × 10⁻⁵ | mammal 0.999 | **probably damaging (0.985)** | Real and rare, but private to the Altai individual — the one whose parents were half-siblings.  Individual mutational load in an inbred person, not a lineage allele.  Also a G→A change, the deamination class, so it carries the standard ancient-DNA caveat even at 32 of 33 reads. |
| *PALB2* p.Gln1023Arg | all three 1/1 | absent from 1000 Genomes and the 9 exomes | mammal 0.996, GERP 4.19 | possibly damaging (0.53–0.86) | Genuinely Neanderthal-specific, on a conserved site — but *PALB2* is BRCA2's partner in nuclear homologous recombination and has nothing to do with mitochondrial DNA. |

---

## 4.  TEFM p.Asp320His — the mitochondrial candidate, run to ground

The catalog's headline hit was in a mitochondrial gene: *TEFM*, the mitochondrial transcription elongation factor, at chr17:29,226,312 C→G, PolyPhen **probably damaging (0.983)**, SIFT deleterious (0), GERP 5.1, absent from 1000 Genomes.  Altai and Vindija are homozygous derived (47- and 21-fold), El Sidrón has no call, the Denisovan and all nine human exomes are ancestral.

It does not survive.

**The residue is not conserved.**  The catalog's own mammal-conservation score at the site is **0.013**, and a direct check bears it out.  Aligning 114 mammalian *TEFM* orthologues from UniProt against the human protein (`code/repair-genes/tefm_conserv.py`):

| residue at position 320 | mammals (n = 114) | tetrapods (n = 146) |
|---|---|---|
| D (human) | 53.5 % | 48.6 % |
| E | 42.1 % | 35.6 % |
| N | — | 10.3 % |
| **H (the Neanderthal residue)** | 0.9 % | **2.1 %** |
| G, A, S | 2.7 % | 2.8 % |

Histidine at position 320 occurs in living vertebrates — the Cape golden mole (*Chrysochloris asiatica*) and the bearded dragon (*Pogona vitticeps*) both carry it.  The site tolerates at least six different residues.

**The prediction has since been downgraded.**  Current Ensembl VEP scores the canonical transcript at **possibly damaging (0.65)**, not 0.983; the 2014 catalog itself carries 0.691 on a second transcript.  PolyPhen has been retrained in the intervening decade.

**And a TEFM defect is the wrong kind of defect.**  TEFM makes mitochondrial transcription processive and governs the switch between transcription and replication-primer formation.  Human *TEFM* disease variants (Rius et al., *Sci Rep* 2023, PMC9950373) cause respiratory-chain deficiency and childhood-onset neurological disease, with defective transcription elongation, unprocessed transcripts and *elevated* mitochondrial DNA copy number.  Nothing in that literature touches the mutation rate.

---

## 5.  What the published literature says, and does not say

- **Castellano et al. 2014** built these catalogs and ran the obvious enrichment test on them.  What came out was skeletal morphology on the Neanderthal lineage and pigmentation and behaviour on ours.  The words "repair" and "replication" do not occur anywhere in the paper.
- **Aneli et al. 2024** (*Hum Mol Genet* 33:1152–1163) did the functional version of this search.  They looked for non-synonymous, fixed, archaic-derived variants in mitonuclear genes that are rare or absent in humans, then expressed the archaic and human alleles in *Saccharomyces cerevisiae* and measured what happened.  One variant had a real effect — in *YARS2*, the mitochondrial tyrosyl-tRNA synthetase — giving reduced respiratory activity and lower Cox2.  That is mitochondrial **translation**, not replication or repair, and the phenotype is respiratory deficiency, not hypermutation.
- **No paper anywhere proposes an archaic DNA-repair defect.**  Jordan reaches it purely by elimination: "If sequencing errors did not cause the excessive number of Neanderthal Y chromosome variant alleles, the only other option is hypermutation."

---

## 6. Current article treatment

The September revision reports the candidate, archived genotypes, and scan limits. It removes the domain-based dismissal, uniformly recessive claim, and false annual-rate ceiling. Mouse mutation-burden measurements are examples of possible effects, not a universal ceiling for human POLG variants.

---

## 7.  Open items

- **POLG L580P deserves a functional answer, and the assay exists.**  Aneli et al.'s yeast complementation system is the obvious route: yeast *MIP1* is the POLG orthologue, and the Parma group has already used it to test archaic mitonuclear variants.  Until somebody does that, the direction of effect is a prediction, not a measurement.
- **Chagyrskaya 8 could not be included.**  Its VCFs on the EVA server are plain gzip, not BGZF, so the shipped tabix indexes do not work; the same is true of the Vindija 33.19 and Denisova files in the Vindija-project directory, which failed with `Invalid BGZF header` on every attempt.  Only the Altai high-coverage VCF and the 2014 exome VCFs were queryable.  Adding Chagyrskaya and Vindija 33.19 would say whether POLG L580P is common in Neanderthals or a two-individual accident.
- **Indels were not examined.**  The catalogs are single-nucleotide only.
- **Regulatory change was not examined.**  A mutator could in principle be a promoter or enhancer variant.  The catalogs do carry upstream and downstream rows for panel genes (84 upstream and 86 downstream on the Neanderthal lineage, 83 and 47 on ours) but they are unannotated for effect.

---

## 8.  Files

- Code: `code/repair-genes/` — `genes.py` (the 206-gene panel), `scan_catalog.py`, `scan_all.py`, `protein_diff.py`, `tefm_conserv.py`, `verify2.py`, `pull_exome_vcf.sh`.
- Data: `data/repair-genes/` — `catalog-scan.txt`, `catalog-scan-all-classes.txt`, `protein-diff.json`, `tefm-residue-320-conservation.txt`, `candidate-sites-exome-vcf.txt`.
- Library: `sources/data/mpi-eva-2014_archaic-aa-change-catalog/` (the three VEP catalogs plus their column definitions), `sources/papers/castellano-2014_neandertal-exomes-coding-variation.pdf`.

---

## The Denisovan panel: scanned 2026-08-25 16:32, interpreted 2026-08-25 21:10

The Denisovan mitochondrial distance is a separate constraint. L580P is ancestral in the stored Denisovan exome call, so that variant alone would not explain the Denisovan result. The old 11–22× requirement used an invalid branch allocation and is withdrawn.

`denisovan_scan.py` returned **222 rows, 178 distinct variants** across the 206-gene panel from the Denisova 3 high-coverage genome.  Raw, that looks like a much richer harvest than the Neanderthals gave — including **four loss-of-function calls** where the Neanderthal lineage had **zero**.  It is not.  Three separate signatures identify almost all of it as artifact, and the residue is in the wrong pathways.

### 1. The rare calls carry the deamination signature

Split the 178 by modern-human allele frequency.  Variants common in living people (AF ≥ 1%) are shared polymorphism by construction and are the null; variants rare or absent (AF < 0.1%) are the Denisovan-specific candidates.

| change class | rare (n=89) | common (n=78) |
| --- | --- | --- |
| C→T | 21 | 10 |
| G→A | 23 | 12 |
| **C→T + G→A share** | **50.0%** | **28.2%** |

Post-mortem cytosine deamination produces C→T and G→A and **nothing else** — it cannot make a transversion.  The candidate-specific class is enriched 1.8-fold in exactly the two changes damage produces, against a null drawn from the same file, the same gene panel and the same caller.

### 2. And they are heterozygous

| set | het | hom | het share | median DP |
| --- | --- | --- | --- | --- |
| rare | 57 | 32 | **64%** | 27 |
| common | 11 | 67 | **14%** | 30 |

Deamination damages one strand in one read family, so it surfaces as a heterozygous call.  Real fixed archaic differences surface homozygous.  The rare set is 4.5 times more heterozygous than the common set at the same depth.

**The two signals compound.  Of the fifteen "damaging and rare" hits, ten are heterozygous C→T or G→A** — the damage profile exactly.

### 3. Gene clustering betrays paralogue mismapping

| gene | rare variants |
| --- | --- |
| **PRIM2** | **14** |
| SLX4, LIG1, POLE | 4 each |
| XRCC3, ERCC6, TOP1MT, BRCA2 | 3 each |

Fourteen rare variants in one gene in one individual is not a genotype, it is mismapping.  *PRIM2* lies in a segmental duplication with a known paralogue, and it supplies **three of the four loss-of-function calls** (a frameshift, a stop-gain and a stop-loss, all heterozygous, all within 69 kb).  The fourth, *MTPAP* R71*, is called at GQ 32.7 — the lowest-quality call in the whole set.  *PRIM2* is also the **nuclear** primase; mitochondrial primers are made by POLRMT, so even a genuine null allele there could not touch the mtDNA rate.

### 4. Cross-check in a second, independent call set

The eight damaging candidates in mtDNA-relevant pathways were re-genotyped in the MPI-EVA **combined exome VCF** — three Neanderthals, the Denisovan and nine present-day humans in one file, with a chimpanzee ancestral allele per site, explicitly filtered for low-quality deamination.

| candidate | ref>alt | chimp anc | Altai | Vindija | Sidrón | Denisovan | modern carriers |
| --- | --- | --- | --- | --- | --- | --- | --- |
| LIG1 R/C557 | G>**.** | G | 0/0 (55) | 0/0 (53) | 0/0 (11) | ./. (0) | none of 9 |
| LIG1 R/C821 | G>**.** | G | 0/0 (45) | 0/0 (18) | ./. (0) | ./. (0) | none of 9 |
| LIG1 T/M470 | G>**.** | G | 0/0 (38) | 0/0 (8) | ./. (0) | ./. (0) | none of 9 |
| LIG1 R/C672 | G>A | G | 0/0 (45) | 0/0 (20) | ./. (0) | **0/1 (13)** | none of 9 |
| CLPP L/M130 | C>**.** | C | 0/0 (38) | ./. (0) | ./. (0) | ./. (0) | none of 9 |
| CLPP Y/F138 | A>**.** | A | 0/0 (32) | ./. (0) | ./. (0) | ./. (0) | none of 9 |
| TOP1MT V/M360 | C>**.** | C | 0/0 (42) | 0/0 (72) | 0/0 (15) | ./. (0) | none of 9 |
| TOP1MT T/M554 | G>**.** | G | 0/0 (37) | 0/0 (34) | 0/0 (10) | ./. (0) | none of 9 |

**Seven of the eight sites are not polymorphic in that file at all** — the ALT column is `.`, meaning no alternate allele was called in any of the thirteen samples.  The one that is polymorphic, *LIG1* R/C672, is a **G→A heterozygote at DP 13** in the Denisovan while both covered Neanderthals are ancestral at DP 45 and DP 20.  That is the damage profile again.

State the limit honestly: `./.` in the Denisovan column is **absence of exome-capture coverage, not refutation**.  The positive evidence is the other direction — every *covered* archaic and modern sample at these sites carries the ancestral allele, and not one candidate is corroborated by a second sample.

### 5. The pathways are wrong anyway

Mitochondria run **base-excision repair** and their own **replication/maintenance** set.  They have **no nucleotide-excision repair and no mismatch repair**, and essentially no homologous recombination.  A damaged *ERCC6*, *MSH2*, *XRCC3*, *BRCA2*, *FANCA* or *SLX4* allele cannot change the mtDNA mutation rate whatever it does in the nucleus.  That removes most of the damaging list before any artifact argument is needed.

**Panel correction:** `genes.py` files *LIG1* under "base-excision repair (the mtDNA repair pathway)."  That is right for the nucleus and wrong for mitochondria — the mitochondrial ligase is **LIG3**.  The four *LIG1* hits should never have counted as mtDNA-relevant.

### 6. The core machinery is clean

Across the 21 genes that actually set mitochondrial replication fidelity or repair mtDNA lesions — *POLG, POLG2, TWNK, TFAM, SSBP1, MGME1, POLRMT, TEFM, DNA2, RNASEH1, MTERF1, PRIMPOL, LIG3, OGG1, UNG, NEIL1, MUTYH, APEX1, POLB* — the Denisovan carries **11 variants in 7 genes, and every one is PolyPhen-benign.**

**The highest PolyPhen score anywhere in the core set is 0.414.**  *POLG* itself carries a single benign P73A.  And the Denisovan does not carry *POLG* Leu580Pro, the one Neanderthal variant worth taking seriously — it is ancestral there at DP 21.

## Verdict

**There is no Denisovan DNA-repair defect.**  The lineage that needs the largest speed-up — 11 to 22 times — has the cleanest core machinery of the three archaic genomes examined.  Its apparently richer variant harvest is residual deamination plus mismapping in one duplicated gene, identified by three independent signatures and not corroborated by a second call set.

This closes the mechanism side of the hypermutation escape from both ends.  It matches the branch-length result exactly: the Denisovan mitochondrial branch measures **0.96** of the human branch, and there is nothing in its nuclear repair machinery that could have made it any faster.

Code: `code/repair-genes/denisovan_verdict.py`, `code/repair-genes/denisovan_crosscheck.py`.
Data: `data/repair-genes/denisovan-verdict.{txt,json}`, `data/repair-genes/denisovan-crosscheck.{txt,json}`.