Nutrigenomics — Personalised Nutrition from Genetic Data

Skill ID: nutrigenomics Version: 0.3.12 Status: Beta Author: David de Lorenzo Requires: Python 3.11+ (standard library only for the analysis; pandas, numpy, matplotlib and seaborn are needed only for figures)

---

What This Skill Does

The Nutrigenomics generates a personalised nutrition report from consumer genetic data (23andMe, AncestryDNA raw files or VCF). It interrogates a curated set of nutritionally-relevant SNPs drawn from GWAS Catalog, ClinVar, and peer-reviewed nutrigenomics literature, then translates genotype calls into actionable dietary and supplementation guidance — all computed locally.

Key outputs

• Markdown nutrition report with risk scores and per-SNP genotype calls
• Radar chart of nutrient risk profile
• Gene × nutrient heatmap
• Reproducibility bundle (README_reproducibility.txt, environment.yml, checksums.txt, provenance.json)

---

Trigger Phrases

The Bio Orchestrator should route to this skill when the user says anything like:

• "personalised nutrition", "nutrigenomics", "diet genetics"
• "what should I eat based on my DNA"
• "nutrient metabolism", "vitamin absorption genetics"
• "MTHFR", "APOE", "FTO", "BCMO1", "VDR", "FADS1/2"
• "folate", "omega-3", "vitamin D", "caffeine metabolism", "lactose", "gluten"
• Input files: .txt or .csv (23andMe), .csv (AncestryDNA), .vcf

---

Curated SNP Panel

Implemented panel: 28 SNPs across 24 genes, 12 nutrient domains (see data/snp_panel.json). The tables below also list 3 documented candidate variants not yet in the scoring panel — ADRB2 rs1042713, HLA-DQ2 (proxy SNPs), and GSTT1 (deletion) — which require genotyping/scoring approaches not yet implemented. They are shown for scientific context; the scorer evaluates only the 28 SNPs present in the JSON.

Macronutrient Metabolism

Gene	SNP	Nutrient Impact	Evidence
FTO	rs9939609	Energy balance, fat mass, carb sensitivity	Strong (GWAS)
PPARG	rs1801282	Fat metabolism, insulin sensitivity	Moderate
APOA5	rs662799	Triglyceride response to dietary fat	Strong
TCF7L2	rs7903146	Carbohydrate metabolism, T2D risk	Strong
ADRB2	rs1042713	Fat oxidation, exercise × diet interaction	Moderate

Micronutrient Metabolism

Gene	SNP	Nutrient	Effect of risk allele
MTHFR	rs1801133	Folate / B12	↓ 5-MTHF conversion (~70%)
MTHFR	rs1801131	Folate / B12	↓ enzyme activity (~30%)
MTR	rs1805087	B12 / homocysteine	↑ homocysteine risk
BCMO1	rs7501331	Beta-carotene → Vitamin A	↓ conversion (~50%)
BCMO1	rs12934922	Beta-carotene → Vitamin A	↓ conversion (compound het)
VDR	rs2228570	Vitamin D absorption	↓ VDR function
VDR	rs731236	Vitamin D	↓ bone mineral density response
GC	rs4588	Vitamin D binding	↑ deficiency risk
SLC23A1	rs33972313	Vitamin C transport	↓ renal reabsorption
ALPL	rs1256335	Vitamin B6	↓ alkaline phosphatase activity

Omega-3 / Fatty Acid Metabolism

Gene	SNP	Nutrient	Effect
FADS1	rs174546	LC-PUFA synthesis	↑/↓ EPA/DHA from ALA
FADS2	rs1535	LC-PUFA synthesis	Modulates omega-6:omega-3 ratio
ELOVL2	rs953413	DHA synthesis	A allele: lower EPA→DHA conversion
APOE	rs429358	Saturated fat response	ε4 → ↑ LDL-C on high SFA diet
APOE	rs7412	Saturated fat response	Combined with rs429358 for ε typing

Caffeine & Alcohol

Gene	SNP	Compound	Effect
CYP1A2	rs762551	Caffeine	Slow/Fast metaboliser
AHR	rs4410790	Caffeine	Modulates CYP1A2 induction
ADH1B	rs1229984	Alcohol	Acetaldehyde accumulation risk
ALDH2	rs671	Alcohol	Asian flush / toxicity risk

Food Sensitivities

Gene	SNP	Sensitivity	Effect
MCM6	rs4988235	Lactose intolerance	Risk allele G (-13910C) is non-persistence; the persistence allele A (-13910T) is dominant
HLA-DQ2	Proxy SNPs	Coeliac / gluten	HLA-DQA1/DQB1 risk haplotypes

Antioxidant & Detoxification

Gene	SNP	Pathway	Effect
SOD2	rs4880	Manganese SOD	↓ mitochondrial antioxidant
GPX1	rs1050450	Selenium / GSH-Px	↓ glutathione peroxidase
GSTT1	Deletion	Glutathione-S-trans	Null genotype → ↑ oxidative risk
NQO1	rs1800566	Coenzyme Q10	↓ CoQ10 regeneration
COMT	rs4680	Catechol / B vitamins	Met/Val → methylation load

---

Algorithm

1. Input Parsing (parse_input.py)

Accepts:

• 23andMe .txt or .csv (tab-separated: rsid, chromosome, position, genotype)
• AncestryDNA .csv
• Standard VCF (extracts GT field)

Auto-detects format from header lines. Normalises alleles to forward strand using a hard-coded reference table (avoids requiring external databases).

2. Genotype Extraction (extract_genotypes.py)

For each SNP in the panel:

1. Look up rsid in parsed data
2. Return genotype string (e.g. "AT", "TT", "AA")
3. Flag as "NOT_TESTED" if absent (common for chip-to-chip variation)

3. Risk Scoring (score_variants.py)

Each SNP is scored on a 0 / 0.5 / 1.0 scale by default:

• 0.0 — homozygous reference (lowest risk)
• 0.5 — heterozygous
• 1.0 — homozygous risk allele

Lactose (rs4988235) instead declares inheritance: dominant_protective in the panel. Lactase non-persistence is autosomal recessive, so one copy of the persistence allele is sufficient: AA and AG both score 0.0, and only GG scores 1.0. Any panel entry without an inheritance field is scored additively as above.

Composite Nutrient Risk Scores (0–10) are computed per nutrient domain by summing weighted SNP scores. Weights are derived from reported effect sizes (beta coefficients or OR) in the primary literature.

Risk categories:

• 0–3: Low risk — standard dietary advice applies
• 3–6: Moderate risk — dietary optimisation recommended
• 6–10: Elevated risk — consider testing and targeted supplementation

Important caveat: These are polygenic risk indicators based on common variants. They are not diagnostic. Rare pathogenic variants (e.g. MTHFR compound heterozygosity with high homocysteine) require clinical confirmation.

4. Report Generation (generate_report.py)

Outputs a structured Markdown report with:

• Executive summary (top 3 personalised findings)
• Per-nutrient sections: genotype table → interpretation → recommendation
• Radar chart (matplotlib) of nutrient risk scores
• Gene × nutrient heatmap (seaborn)
• Supplement interactions table
• Disclaimer section
• Reproducibility block

5. Reproducibility Bundle (repro_bundle.py)

Exports to the output directory (not committed to the repo):

• README_reproducibility.txt — step-by-step instructions to reproduce the analysis manually
• environment.yml — pinned conda environment
• checksums.txt — SHA-256 checksums of the SNP panel and output report (input file intentionally excluded to avoid persisting a fingerprint of genetic data)
• provenance.json — timestamp, version, and format arguments (input filename intentionally omitted)

Note: No executable scripts are generated. The reproducibility bundle contains only text files for documentation and integrity verification.

---

Execution

To run the analysis on a user-provided genetic file:

python <skill-dir>/nutrigenomics.py --input <path_to_genetic_file> --format auto

To run a demo without real genetic data (synthetic patient file included with the skill):

python <skill-dir>/nutrigenomics.py --input <skill-dir>/tests/synthetic_patient.csv --format 23andme

Two rules make these commands work, and both matter.

1. Replace <skill-dir> with the absolute path to this skill's folder. Each platform supplies it:

• Hermes states it as [Skill directory: ...] when it loads this skill, and substitutes ${HERMES_SKILL_DIR} anywhere that token appears.
• OpenClaw substitutes {baseDir} at runtime — e.g. python {baseDir}/nutrigenomics.py --input .... Do not substitute it manually.
• Otherwise, use the literal absolute path of the skill folder.

2. Run from the user's working directory — never cd into the skill folder. The script confines output to the current working directory (path_safety.validate_output_dir), so the working directory decides where the report lands. Running from the user's directory puts it there, which is correct. Running from inside the skill folder makes the skill folder the only permitted destination, writing a report full of per-SNP genotype calls into the skill itself — where it is liable to be swept into a publish, a commit, or a shared bundle.

Output is written to the directory named by --output (default: nutrigenomics_results/), created under the current working directory, and persists until manually deleted. OpenClaw's structured entry point (openclaw_adapter.py, declared in openclaw.json) instead defaults to a timestamped nutrigenomics_output_YYYYMMDD_HHMMSS/ directory.

Supported --format values: auto (default), 23andme, ancestry, vcf.

Usage

Ask the agent in plain language — it routes to this skill and runs the command above:

• "Generate my personalised nutrition report from genome.csv"
• "Run a nutrigenomics analysis on variants.vcf and flag any folate pathway risks"
• "What does my APOE status mean for my saturated fat intake?"
• "Run a demo nutrigenomics report using the synthetic patient file"

---

File Structure

skills/nutrigenomics/
├── SKILL.md                      ← this file (agent instructions)
├── nutrigenomics.py            ← main entry point
├── parse_input.py                ← multi-format parser
├── extract_genotypes.py          ← SNP lookup engine
├── score_variants.py             ← risk scoring algorithm
├── generate_report.py            ← Markdown + figures
├── repro_bundle.py               ← reproducibility export
├── data/
│   └── snp_panel.json            ← curated SNP definitions
├── tests/
│   ├── synthetic_patient.csv     ← fixed 23andMe-format test data (for pytest)
│   └── test_nutrigenomics.py           ← pytest suite
└── examples/
    ├── generate_patient.py       ← random patient generator (demo use)
    ├── data/                     ← generated patient files land here (gitignored)
    └── output/
        ├── nutrigenomics_report.md     ← pre-rendered demo report
        ├── nutrigenomics_radar.png     ← demo radar chart (nutrient risk profile)
        └── nutrigenomics_heatmap.png   ← demo gene × nutrient heatmap

Note: Runtime output directories and randomly generated patient files are excluded from version control. Only the pre-rendered demo report in examples/output/ is committed.

---

Privacy

All computation runs locally — no genetic data is ever transmitted to external servers or third-party services.

What the report contains: The Markdown report includes per-SNP genotype calls (e.g. AT, TT) for each of the 28 panel SNPs analysed. This is intentional: knowing your specific genotype at each nutrition-related locus is what makes the report actionable. Full raw genome data from the input file is not reproduced in the report; only the 28 panel SNPs are included.

The input filename appears in the report header. It is shown so you can tell which file a report came from, but a name like jane_smith_genome.csv identifies a person, and the report is the artefact most likely to be shared with a clinician. Rename the input file first if that matters to you. The filename is sanitised before it is written, so a name containing Markdown or HTML cannot inject content into the report; the full path is never recorded anywhere.

File persistence: Output files (report, figures, reproducibility bundle) are written to a timestamped nutrigenomics_output_YYYYMMDD_HHMMSS/ directory under the working directory and persist on disk until manually deleted. The input file is read-only and is never copied into the output directory.

If you are running this skill on behalf of others or in a shared environment, delete the output directory once the user has downloaded their results.

---

Evidence Quality

Not every SNP in this panel rests on equally strong evidence, and the report does not weight them by evidence quality — only by effect size. Every citation in data/snp_panel.json has been checked against PubMed, and the panel splits into two tiers.

Tier 1 — GWAS-backed. A catalogued genome-wide association exists for this exact variant and a trait matching its nutrient domain: MTHFR rs1801133 and rs1801131, GC rs4588, FADS1 rs174546, FADS2 rs1535, APOE rs429358 and rs7412, SLC23A1 rs33972313, ALPL rs1256335, FTO rs9939609, TCF7L2 rs7903146, PPARG rs1801282, APOA5 rs662799, AHR rs4410790, ADH1B rs1229984, ALDH2 rs671, MCM6 rs4988235.

Tier 2 — candidate-gene evidence only. No catalogued genome-wide association for the assigned nutrient domain; the entry rests on functional or candidate-gene studies: BCMO1 rs7501331 and rs12934922, VDR rs2228570 and rs731236, MTR rs1805087, CYP1A2 rs762551, SOD2 rs4880, GPX1 rs1050450, NQO1 rs1800566, COMT rs4680.

Four Tier 2 entries carry caveats that a reader should know about:

• rs762551 (CYP1A2, caffeine) — the widely used "fast vs slow caffeine metaboliser" interpretation is weaker than usually presented. In the source study (Sachse et al. 1999), caffeine metabolite ratios did not differ significantly between genotypes in 185 non-smokers; the effect appeared only in 51 smokers, and is best understood as an inducibility effect. A contemporaneous study (Welfare et al. 1999) concluded that CYP1A2 shows no functionally significant polymorphism at all.
• rs731236 (VDR, TaqI) — the standard review of VDR polymorphisms describes the BsmI/ApaI/TaqI cluster as probably non-functional, with observed associations attributed to linkage disequilibrium with functional variants elsewhere in the gene. It is scored here as a haplotype tag, not as a causal variant.
• rs1050450 (GPX1, antioxidant) — the phenotype study for this variant found no significant difference in erythrocyte glutathione peroxidase activity between genotypes and no association with stroke. It is retained for completeness; the evidence for a functional effect on antioxidant defence is weak.
• rs4680 (COMT, antioxidant) — the three-to-four-fold effect on COMT enzyme activity is well established, but COMT inactivates catecholamines and catechol drugs. Its placement in the antioxidant domain reflects catechol handling rather than direct antioxidant defence.

rs1805087 (MTR) is a genuine but modest determinant of homocysteine; its source study reports an effect independent of folate and B12 status, and larger in people with low vitamin B6.

rs953413 (ELOVL2, omega-3) sits between the tiers and is described on its own. The ELOVL2 locus is genome-wide significant for n-3 fatty acids (Lemaitre et al. 2011), but through other, correlated SNPs; the GWAS Catalog holds no omega-3 association for rs953413 itself. The direct evidence is:

• Association. In Tanaka et al. 2009 (PMID 19148276) DHA fell from GG through AG to AA in two cohorts, and the minor A allele went with higher EPA and DPA and lower DHA. The discovery signal did not reach genome-wide significance; the DHA association replicated.
• Mechanism. In liver cell lines the G allele gives higher ELOVL2 enhancer activity than A (iScience 2020, PMID 31928966), consistent with A carriers converting less EPA to DHA.
• Supplementation. Mixed. A carriers showed larger rises in blood EPA and DHA in the MARINA trial (PMID 24292947) and a small 2025 study (PMID 40114193), but not in the seAFOod trial (PMID 42135634).

It is scored as an association with lower EPA-to-DHA conversion, not as a requirement: none of these studies shows that any genotype is deficient or needs a particular daily intake. The effect is modest, about 0.2 percentage points of DHA between GG and AA. Other ELOVL2 variants such as rs3734398, rs2236212 and rs3798713 are strongly correlated with it in European, East and South Asian ancestry, so they are not added as independent risks; doing so would count one signal several times.

Treat Tier 2 scores as weaker signals than Tier 1 scores when interpreting a report.

---

Threat Model and Limits

The input is a genetic data file. It is treated as untrusted: it may not have come from the person reading the report, and everything derived from it reaches a Markdown document that is likely to be shared onward. Four classes of risk are handled explicitly, so that new instances are covered by an existing rule rather than fixed one at a time.

1. Untrusted values reaching the report. Every value written into the report is filtered before rendering, not just the ones known to be attacker-controlled today: genotype calls, the input filename, and the gene, rsID and effect text taken from the SNP panel. Genotypes are additionally validated at parse time and non-nucleotide calls are discarded rather than scored.

2. Resource exhaustion. Parsing streams, but a genotype table still grows with the number of variants, so streaming alone bounds nothing. Explicit caps in path_safety.py:

Limit	Value	Why
MAX_INPUT_BYTES	512 MB	a 23andMe export is ~25 MB
MAX_VARIANTS	10,000,000	a consumer export holds ~600,000
MAX_LINE_BYTES	64 KB	a real variant line is under 100 bytes
MAX_HEADER_LINES	1,000	only the header can identify a format

Exceeding the file limit fails with a clear error; the others discard the excess and carry on.

3. Filesystem safety. Output is confined to the working directory, writes refuse to follow a symbolic link at either the file or its parent directory, and output directories are created exclusively with entropy added on collision, so concurrent runs cannot overwrite one another's report.

4. Provenance leakage. No absolute path is written to any artefact. The input filename appears in the report header only, and is sanitised.

Out of scope. Dependencies are declared with version bounds but not cryptographic hashes; the skill installs nothing itself, and the four packages involved are needed only for figures. The report necessarily contains per-SNP genotype calls — that is the point of it — and output files persist until deleted.

---

Limitations & Disclaimer

1. Not a medical device. This skill provides educational, research-oriented nutrigenomics analysis. It does not constitute medical advice.
2. Common variants only. The panel covers SNPs with MAF > 1% in at least one major population. Rare pathogenic variants are out of scope.
3. Population context. Effect sizes are predominantly derived from European GWAS cohorts. Risk estimates may not generalise equally across all ancestries.
4. Gene–environment interaction. Genetic risk scores interact with baseline diet, lifestyle, microbiome, and epigenetic state. A "high risk" score does not mean a nutrient deficiency is present — it means the individual may benefit from monitoring.
5. Simpson's Paradox note. Population-level associations used to derive weights may not reflect individual trajectories (see Corpas 2025, Nutrigenomics and the Ecological Fallacy).

---

Roadmap

• v0.2: Microbiome × genotype interaction module (16S rRNA input)
• v0.3: Longitudinal tracking — compare reports across time
• v0.4: HLA typing for immune-mediated food reactions (coeliac, gluten sensitivity)
• v1.0: Multi-omics integration (metabolomics + genomics + dietary recall)

---

References

This skill's SNP panel and methodology are informed by peer-reviewed nutrigenomics research. For verification and additional details, consult:

• PubMed MEDLINE: https://pubmed.ncbi.nlm.nih.gov/
• GWAS Catalog: https://www.ebi.ac.uk/gwas/ (published genome-wide association studies)
• ClinVar: https://www.ncbi.nlm.nih.gov/clinvar/ (variant interpretations)

Users are encouraged to verify specific claims through these authoritative sources and with qualified healthcare providers.

---

Contributing

The SNP panel (data/snp_panel.json) is maintained by the skill author. To suggest additions or corrections, contact David de Lorenzo directly via GitHub (@drdaviddelorenzo) or open an issue on GitHub.