Reading 23andMe Raw Data for CYP2D6 Star Alleles in Python — Why DTC Often Misses *5 Deletion
Step-by-step Python tutorial, run on real public 23andMe files: find the CYP2D6 markers on GRCh37, read them on the forward strand, turn them into a CPIC activity score, and see what a SNP chip cannot tell you (phasing, the *5 deletion, gene duplications).
글 SYSOFT 연구팀 · 공식 지침과 논문을 출처로 정리한 정보 글이며, 의사의 감수를 거치지 않았습니다.
Before you read. This post summarises published research and guidelines for general information. It is not medical advice; talk to your doctor before making decisions about tests, treatment or supplements.
Corrected in October 2026: coordinates changed to GRCh37 and alleles to the forward strand that 23andMe uses (the old code could not find or call any variant); phenotypes now follow the CPIC/DPWG activity score (Caudle 2020, Clin Transl Sci 13:116–124), and rs1065852 alone is no longer read as *10. The code was re-run on public CC0 Personal Genome Project files and the output shown is real. The *5 frequencies were corrected, and tools and prices we could not verify were removed.
TL;DR (Quick Answer)
Can you use 23andMe data for CYP2D6 pharmacogenomics? Only partly, and it depends on your chip. On the older-chip file we tested, the *4 and *10 markers were there and the code could make a cautious call. On both newer v5 files we tested, the *4 (rs3892097) and *10 (rs1065852) markers were not in the file at all. No SNP chip can see the *5 whole-gene deletion or gene duplications.
- Why CYP2D6 matters — it metabolizes many commonly prescribed drugs, including codeine and tramadol (prodrug activation), many antidepressants, and tamoxifen.
- What a chip can give you — genotypes at a handful of defining positions. From those you can get a chip-consistent diplotype and a CPIC activity score, if the key markers are on your chip.
- What a chip cannot give you — phasing (which variants sit on the same chromosome), the *5 deletion, duplications (xN) and CYP2D6/CYP2D7 hybrids.
- Clinical bottom line — treat a DTC CYP2D6 result as a screen, not a diagnosis; confirm any actionable finding with a clinical PGx test.
What CYP2D6 Star Alleles Mean
CYP2D6 is a highly polymorphic liver cytochrome-P450 enzyme. Each person inherits two alleles, named with "star" nomenclature (*1, *2, *4, *5 …). Each allele has an activity value, the two values add up to an activity score, and the score maps to a metabolizer phenotype. Star-allele definitions are maintained by PharmVar (CYP2D6), drug–gene evidence by PharmGKB, prescribing guidance by the CPIC guidelines, and label biomarkers by the FDA Table of Pharmacogenomic Biomarkers in Drug Labeling.
The CPIC/DPWG consensus (Caudle et al. 2020) translates the activity score like this:
| Activity score | Phenotype | Example diplotypes (from Caudle 2020) |
|---|---|---|
| 0 | Poor metabolizer | *4/*4, *5/*5, *3/*4 |
| 0 < score < 1.25 | Intermediate metabolizer | *4/*10 (0.25), *10/*10 (0.5), *1/*5 (1.0) |
| 1.25 – 2.25 | Normal metabolizer | *1/*10 (1.25), *1/*1, *1/*2 (2.0) |
| > 2.25 | Ultrarapid metabolizer | *1/*1xN |
Activity values used below come from CPIC's current allele table: *1 and *2 = 1; *3, *4, *5, *6 = 0; *17 = 0.5; *10 and *41 = 0.25. (Caudle 2020 introduced the 0.25 value for *10; CPIC's table now lists *41 at 0.25 as well.)
"Can I Use My 23andMe Data for Pharmacogenomics?"
Short answer: for some CYP2D6 alleles, sometimes; for the *5 whole-gene deletion and duplications, never. Understanding which is which prevents the most common DTC-PGx interpretation mistake.
This tutorial walks through:
- How to load a 23andMe raw data file in Python and find the CYP2D6 markers by GRCh37 position
- How to read those markers on the forward strand and turn them into a cautious star-allele call and activity score
- What the chip cannot tell you: phasing, *5 and duplications
- Where automated tools fit, and their limits
It's a worked example of the broader principle covered in 약물유전체학이란? 실제 사례로 보는 유전자와 약 반응: SNP-based DTC tests catch some common variants and miss structural variants.
The Files Used Here
Everything below was run on real files, not made-up examples. We used public 23andMe raw data from the Harvard Personal Genome Project. PGP participants share their genomic data publicly under a CC0 waiver or equivalent public-domain licence (PGP Global Network guidelines). The PGP data portal was not loading in October 2026, so we took the files from the Internet Archive's copies of the PGP download links.
- PGP participant hu5BB600: a 23andMe download generated in 2017, 960,614 rows (an older chip).
- PGP participant hu1C1368: a 23andMe v5 download generated in 2019, 638,570 rows.
A second 2019 v5 download (PGP participant huE50E31, 638,531 rows) was used only to confirm the v5 finding.
The code was run with Python 3.12.3 and pandas 3.0.6. Save the code blocks below, in order, as one file (say cyp2d6.py) and run python cyp2d6.py your_23andme_file.txt.
What a 23andMe Raw Data File Looks Like
After downloading your raw data from 23andMe you get a tab-separated text file (sometimes zipped). The header of the hu5BB600 file includes these lines, which matter for everything below:
# individually validated for accuracy. As such, this data is suitable only for research,
# educational, and informational use and not for medical or other use.
...
# (an rsid or an internal id), its location on the reference human genome, and the
# genotype call oriented with respect to the plus strand on the human reference sequence.
# We are using reference human assembly build 37 (also known as Annotation Release 104).
...
# rsid chromosome position genotype
So: build 37 (GRCh37/hg19) coordinates, forward (plus) strand genotypes. Some rows use 23andMe's own internal IDs (i4001453 and so on) instead of an rsid.
Loading the File in Python
import sys
import pandas as pd
RAW_FILE = sys.argv[1] # your 23andMe download (.txt, or the .zip as downloaded)
snps = pd.read_csv(
RAW_FILE,
sep='\t',
comment='#',
names=['rsid', 'chromosome', 'position', 'genotype'],
dtype={'rsid': str, 'chromosome': str, 'position': int, 'genotype': str},
)
print(f"Total SNPs: {len(snps):,}")
print(snps.head(3).to_string())
Output for hu5BB600:
Total SNPs: 960,614
rsid chromosome position genotype
0 rs4477212 1 82154 AA
1 rs3094315 1 752566 AA
2 rs3131972 1 752721 GG
CYP2D6 — Where It Lives in the Genome
CYP2D6 is on chromosome 22q13.2. On GRCh37, the build 23andMe files use, the gene spans roughly chr22:42,522,500–42,526,900 (Ensembl GRCh37 places rs3892097 at 42,524,947 and rs1065852 at 42,526,694). The range 42,126,000–42,131,000 you may see elsewhere is a GRCh38 location (Ensembl GRCh38 puts the gene at 42,120,962–42,131,727) and matches nothing in a 23andMe file.
# GRCh37 window around CYP2D6 (23andMe files use build 37)
cyp2d6_region = snps[
(snps['chromosome'] == '22') &
(snps['position'].between(42_522_000, 42_527_500))
]
print(f"Rows in the CYP2D6 window: {len(cyp2d6_region)}")
print(f" of which 23andMe internal IDs (i...): {cyp2d6_region['rsid'].str.startswith('i').sum()}")
Output for hu5BB600:
Rows in the CYP2D6 window: 105
of which 23andMe internal IDs (i...): 50
For the v5 file hu1C1368 the same window has 50 rows, 12 of them internal IDs. Most of these rows are not the markers that define the common star alleles, and half of them on the older chip have no rsid at all. That is why the next step matches on position.
Star Allele Lookup Table
CYP2D6 star alleles are named haplotypes (combinations of variants). The markers below are the ones this tutorial looks for. Bases are on the forward strand, as 23andMe reports them. CYP2D6 is on the minus strand, so gene-level names are complemented: *4's 1846G>A shows up as C>T, *10's 100C>T as G>A. Positions and forward-strand alleles were checked against Ensembl GRCh37 and CPIC's allele definitions.
| Star allele | Function (CPIC) | Marker: normal → variant (forward strand) | Allele frequency (CPIC table) |
|---|---|---|---|
| *2 | Normal | rs16947 G → A (also carried by *17 and *41, so not decisive alone) | — |
| *3 | No function | rs35742686 T → deleted | European 1.6% |
| *4 | No function | rs3892097 C → T; *4 often also carries rs1065852 A (CPIC's *4 definition allows either base) | European 18.5% |
| *5 | No function (whole-gene deletion) | No marker: not detectable by SNP array | East Asian 4.8%, European 3.0%, Sub-Saharan African 6.2% |
| *6 | No function | rs5030655 A → deleted | European 1.1% |
| *10 | Decreased (activity 0.25) | rs1065852 G → A | East Asian 42.8% |
| *17 | Decreased (activity 0.5) | rs28371706 G → A | Sub-Saharan African 19.4% |
| *41 | Decreased (activity 0.25) | rs28371725 C → T | European 9.2% |
Frequencies are CPIC's weighted average allele frequencies (api.cpicpgx.org, checked October 2026).
Which markers are on your chip? It varies, and it matters. In the older-chip file (hu5BB600) all seven markers were present, two of them only under internal IDs (i4001453 for *4, i4000452 for *3). In both v5 files we checked, rs3892097 (*4) and rs1065852 (*10) were missing, as were rs16947 and rs35742686.
Python — Read the Markers on the Forward Strand
# Key CYP2D6 markers. Positions are GRCh37. Bases are on the forward (+) strand,
# which is how 23andMe reports genotypes. CYP2D6 sits on the minus strand, so the
# gene-level names are complemented here (e.g. *4's 1846G>A appears as C>T).
# Indels: 23andMe writes I/D; for these 1-bp deletions, D is the variant.
markers = pd.DataFrame([
# rsid GRCh37 pos normal variant what the variant base tags
('rs3892097', 42_524_947, 'C', 'T', '*4'),
('rs1065852', 42_526_694, 'G', 'A', '*10, often on *4'),
('rs16947', 42_523_943, 'G', 'A', '*2, *17, *41 (not decisive alone)'),
('rs28371706', 42_525_772, 'G', 'A', '*17'),
('rs28371725', 42_523_805, 'C', 'T', '*41'),
('rs35742686', 42_524_244, 'I', 'D', '*3'),
('rs5030655', 42_525_086, 'I', 'D', '*6'),
], columns=['rsid', 'position', 'normal', 'variant', 'tags'])
# Match on position, not rsid: 23andMe stores some of these under internal "i" IDs.
chr22 = snps[snps['chromosome'] == '22'].drop_duplicates('position')
calls = markers.merge(chr22[['position', 'rsid', 'genotype']],
on='position', how='left', suffixes=('', '_in_file'))
def count_variant(gt, variant):
if pd.isna(gt) or gt == '--':
return pd.NA # not on this chip, or a no-call
return gt.count(variant)
calls['copies'] = pd.array(
[count_variant(g, v) for g, v in zip(calls['genotype'], calls['variant'])], dtype='Int64')
print(calls[['rsid', 'rsid_in_file', 'genotype', 'variant', 'copies', 'tags']]
.astype(object).fillna('-').to_string(index=False))
Output for hu5BB600:
rsid rsid_in_file genotype variant copies tags
rs3892097 i4001453 CT T 1 *4
rs1065852 rs1065852 AG A 1 *10, often on *4
rs16947 rs16947 GG A 0 *2, *17, *41 (not decisive alone)
rs28371706 rs28371706 GG A 0 *17
rs28371725 rs28371725 CC T 0 *41
rs35742686 i4000452 II D 0 *3
rs5030655 rs5030655 II D 0 *6
Two things to notice. The *4 marker is in the file only as i4001453, so a lookup by rsid would have missed it. And the genotypes are forward-strand letters (CT, AG), so a lookup that expects the gene-strand bases (A for *4, T for *10) would have reported "no variant" for everyone.
Python — From Markers to an Activity Score (Carefully)
The tempting shortcut is "rs1065852 A = *10". It is wrong: the same base is often on *4 alleles too, and the chip does not tell you which variants share a chromosome (phase). The code below only counts a *10 for rs1065852 copies that *4 does not already explain, refuses to call when the key markers are missing, and prints its assumptions.
# CPIC activity values (cpicpgx.org allele functionality table)
ACTIVITY = {'*1': 1.0, '*3': 0.0, '*4': 0.0, '*5': 0.0, '*6': 0.0,
'*10': 0.25, '*17': 0.5, '*41': 0.25}
def phenotype(score):
# Caudle et al. 2020, CPIC/DPWG consensus thresholds
if score == 0:
return 'poor metabolizer'
if score < 1.25:
return 'intermediate metabolizer'
if score <= 2.25:
return 'normal metabolizer'
return 'ultrarapid metabolizer'
def chip_diplotype(calls):
n = dict(zip(calls['rsid'], calls['copies']))
missing = [r for r in ('rs3892097', 'rs1065852') if pd.isna(n[r])]
if missing:
return None, f"Cannot call *4 or *10: {', '.join(missing)} not on this chip."
alleles = []
for rsid, star in [('rs35742686', '*3'), ('rs3892097', '*4'), ('rs5030655', '*6'),
('rs28371706', '*17'), ('rs28371725', '*41')]:
if not pd.isna(n[rsid]):
alleles += [star] * int(n[rsid])
# rs1065852-A sits on *4 as well as *10, so only copies *4 does not explain count as *10
alleles += ['*10'] * max(0, int(n['rs1065852']) - int(n['rs3892097']))
if len(alleles) > 2:
return None, f"Variants {alleles} do not fit on two chromosomes without phasing."
alleles += ['*1'] * (2 - len(alleles)) # '*1' = none of the tested variants (*1, *2, or untested)
return sorted(alleles), None
alleles, problem = chip_diplotype(calls)
if problem:
print(problem)
else:
score = sum(ACTIVITY[a] for a in alleles)
print(f"Chip-consistent diplotype: {'/'.join(alleles)} "
f"(activity score {score} -> {phenotype(score)})")
print("Assumes: standard haplotypes, variants phased as above, exactly two gene copies.")
untested = calls.loc[calls['copies'].isna(), 'rsid'].tolist()
print("Not tested on this chip:", ', '.join(untested) if untested else 'none')
het = calls.loc[calls['copies'] == 1, 'rsid'].tolist()
if het:
print(f"Heterozygous at {', '.join(het)}: two different copies were read, "
"so a *5 deletion is unlikely here.")
elif alleles:
a = alleles[0]
s = ACTIVITY[a]
print(f"All calls homozygous: a hidden *5 would make this {a}/*5 "
f"(activity score {s} -> {phenotype(s)}).")
print("Not visible on any SNP chip: gene duplications (xN) and CYP2D6/CYP2D7 hybrids.")
Output for hu5BB600 (older chip):
Chip-consistent diplotype: *1/*4 (activity score 1.0 -> intermediate metabolizer)
Assumes: standard haplotypes, variants phased as above, exactly two gene copies.
Not tested on this chip: none
Heterozygous at rs3892097, rs1065852: two different copies were read, so a *5 deletion is unlikely here.
Not visible on any SNP chip: gene duplications (xN) and CYP2D6/CYP2D7 hybrids.
Output for hu1C1368 (v5 chip), same code:
Cannot call *4 or *10: rs3892097, rs1065852 not on this chip.
Not tested on this chip: rs3892097, rs1065852, rs16947, rs35742686
Not visible on any SNP chip: gene duplications (xN) and CYP2D6/CYP2D7 hybrids.
How to read the hu5BB600 result. One *4 marker (T) and one rs1065852 A. The simplest reading is that both sit on the same chromosome, a normal *4, with a *1 or *2 on the other: *1/*4, activity score 1.0, which CPIC/DPWG now classify as an intermediate metabolizer. The chip cannot rule out the other phase, a *4 without the rs1065852 change plus a *10 on the other chromosome: *4/*10, score 0.25. That is also intermediate, but with a quarter of the score. Only phased sequencing or a clinical genotyping assay settles it.
Because both markers are heterozygous, two different copies of the gene were read at those positions, which argues against a *5 deletion (array artefacts aside). What the chip cannot exclude is a duplication: if the *1 were duplicated (*1x2/*4), the score would be 2.0, a normal metabolizer.
How to read the hu1C1368 result. Nothing to call. On this v5 file the code can only say that *6 (rs5030655), *17 (rs28371706) and *41 (rs28371725) were not seen. A *4 or *10 carrier would look exactly like this. We saw the same gaps in a second 2019 v5 download (PGP participant huE50E31), which also lacked rs28371706.
Why *5 Is Invisible to 23andMe (And Every Other SNP Array)
CYP2D6 *5 is a whole-gene deletion: the entire CYP2D6 gene is missing on that chromosome. No SNP exists to detect "absence." A SNP array probes specific positions; where one copy is deleted, the probe reads only the remaining copy, which looks like an ordinary homozygous call.
How common is it? CPIC's table gives *5 allele frequencies of about 4.8% in East Asian, 3.0% in European and 6.2% in Sub-Saharan African populations. The share of people who carry at least one *5 is roughly double the allele frequency: with an allele frequency of 4.8%, about 1 − 0.952² ≈ 9% of people carry one.
The DTC interpretation trap. Take someone who is really *1/*5. The chip reads only the *1 copy, so every marker comes back homozygous normal, and they look like *1/*1 (activity score 2.0, normal metabolizer). Their true score is 1.0, an intermediate metabolizer. A *10/*5 person looks like *10/*10 (score 0.5) but is 0.25. Both are intermediate, but the chip overstates the activity. This is why the code above warns when all calls are homozygous.
What about CYP2D6 duplications (ultra-rapid metabolizers)?
Same problem in the opposite direction. CYP2D6 can be present in extra copies on one chromosome (gene duplication). More functional copies mean a higher activity score; above 2.25 is an ultrarapid metabolizer, and for codeine that means faster conversion to morphine and a higher risk of toxicity.
SNP arrays detect alleles, not copy number. 23andMe raw data cannot show duplications. Someone who is *1x2/*1 (score 3, ultrarapid) looks like *1/*1 on the chip.
How Clinical PGx Labs Detect *5 and Duplications
Clinical tests check whether the CYP2D6 gene is present and how many copies exist:
- Copy-number assays (e.g. TaqMan copy-number assays)
- Long-range PCR with primers spanning the deletion or duplication
- Whole-genome or long-read sequencing (PacBio HiFi or Oxford Nanopore), which can also resolve hybrids and phase
These are not what DTC services offer. For clinical decisions involving CYP2D6, a proper PGx test through a clinical lab is needed.
Automated Tools
You don't have to write Python. Tools such as Promethease (which cross-references your raw file against SNPedia; its site lists most reports at $25 as of October 2026) will read a 23andMe file. Whatever the tool, it can only report on markers that are in your file. If your chip lacks rs3892097 and rs1065852, no tool can call *4 or *10 from it, and no tool can see *5 or duplications from chip data. If you're using DTC data for actual prescription decisions, work with a doctor who orders a clinical PGx test.
Beyond CYP2D6 — Other Important PGx Genes on 23andMe
Same approach applies to other genes, with the same caveat: check that the markers are actually in your file before reading anything into their absence.
| Gene | Star alleles on 23andMe? | Detect deletion? |
|---|---|---|
| CYP2C19 | *2, *3, *17 — common variants detected | No CNV detection |
| CYP2C9 | *2, *3 detected; ethnic-specific harder | No CNV detection |
| CYP3A5 | *3 detected | No CNV detection |
| DPYD | Common variants (*2A, *13) sometimes | No CNV detection |
| TPMT | Common variants detected | No CNV detection |
| VKORC1 | rs9923231 detected (relevant for warfarin) | N/A |
| HLA-B*57:01 (abacavir) | Partial via tagging SNPs | Imputation only |
| HLA-B*15:02 (carbamazepine, Asian risk) | NOT reliably detectable by DTC | Needs clinical typing |
HLA typing in particular is poorly served by DTC arrays. For HLA-B*15:02 (severe SJS/TEN risk with carbamazepine) or HLA-B*57:01 (abacavir hypersensitivity), clinical typing is required.
Practical Workflow — From 23andMe to Useful PGx Info
- Download your 23andMe raw data
- Run the Python code above and look first at which markers are missing
- Identify which medications in your future or current regimen involve CYP2D6, CYP2C19 or CYP2C9 metabolism (your doctor can confirm)
- For those medications, check your detected variants, but treat them as a starting hypothesis, not a clinical conclusion
- If a critical medication is involved (especially codeine/tramadol, clopidogrel, warfarin, abacavir, carbamazepine, certain SSRIs/antipsychotics), request a clinical PGx test from your physician
- Never stop or change a prescription based on DTC results alone
FAQ
Q: Will the Python code above work on AncestryDNA raw data too? The file layout differs (AncestryDNA splits the two alleles into separate columns), so the loading step needs changing. The marker logic is the same. Check which of the seven positions your file actually contains.
Q: My file uses hg19 vs hg38 coordinates — does that matter?
Yes, for anything done by position. 23andMe files are GRCh37 (build 37), and the code above uses GRCh37 positions. We match by position rather than rsid because 23andMe stores some CYP2D6 markers under internal IDs (in hu5BB600, the *4 marker is i4001453). If you use a GRCh38 file, change the positions.
Q: Why do the alleles in the code look "wrong" compared with papers? Papers and PharmVar name CYP2D6 variants on the gene's own strand, which is the minus strand of chromosome 22. 23andMe reports the forward strand. So *4's 1846G>A is C>T in your file, and *10's 100C>T is G>A.
Q: How accurate is 23andMe's own pharmacogenetic report? It covers a limited set of variants. Read the report's own description of what it does and does not test; anything outside that list, including *5 and duplications, needs clinical testing.
Q: Can I use 23andMe data to choose between antidepressants? Hypothesis-generating only. Many SSRIs/SNRIs are CYP2D6 substrates (paroxetine, fluoxetine, venlafaxine). If you carry *4 or are *10/*10, discuss it with your psychiatrist; they may order a clinical PGx test before prescribing.
Q: Why does my DTC report disagree with another DTC tool on the same raw data? Different tools use different variant-to-star-allele rules, handle strand and internal IDs differently, and make different phasing assumptions. PharmVar definitions and the CPIC activity score are the standard.
Q: Are there tools that try to detect *5 by imputation? Some research-grade methods use surrounding SNP patterns to probabilistically infer CNVs. They're not validated for clinical use. For clinical decisions, get a real PGx test.
Q: Can I run this in R instead of Python?
Yes — readr::read_tsv() and dplyr::filter() do the same loading and filtering. The marker table and activity-score logic carry over directly.
Q: Where do I find the latest star allele definitions? PharmVar (https://www.pharmvar.org) for definitions, and CPIC (https://cpicpgx.org) for activity values and phenotype rules.
Closing — Key Takeaways
- Check what is on your chip first. In our test, the older-chip file had all seven markers; both v5 files lacked the *4 and *10 markers entirely.
- Read the file the way 23andMe wrote it: GRCh37 positions, forward-strand bases, and some markers under internal IDs.
- Don't equate one SNP with one star allele. rs1065852 A is on *4 and *10; rs16947 A is on *2, *17 and *41. Without phasing, a chip gives you a chip-consistent guess, not a diplotype.
- The chip cannot see *5 or duplications, which can move someone between normal, intermediate and ultrarapid.
- For prescription decisions, especially codeine/tramadol, clopidogrel, warfarin, abacavir and carbamazepine, get a clinical-grade PGx test.
Related posts:
- 약물유전체학이란? 실제 사례로 보는 유전자와 약 반응
- 유전자 검사 결과지, 어떻게 읽나 — 위험도 숫자의 진짜 의미
- BRCA1/2와 유전성 암 — 유전자 검사·예방 전략 완전 가이드 2026
References:
- PharmVar — star allele definitions: https://www.pharmvar.org
- PharmGKB — clinical pharmacogenomics knowledge base: https://www.pharmgkb.org
- CPIC — CYP2D6 allele functionality and frequency tables: https://cpicpgx.org
- Gaedigk, A. et al. (2008). The CYP2D6 activity score: translating genotype information into a qualitative measure of phenotype. Clinical Pharmacology & Therapeutics, 83(2), 234–242. doi:10.1038/sj.clpt.6100406
- Caudle, K. E. et al. (2020). Standardizing CYP2D6 genotype to phenotype translation: consensus recommendations from the Clinical Pharmacogenetics Implementation Consortium and Dutch Pharmacogenetics Working Group. Clinical and Translational Science, 13(1), 116–124. doi:10.1111/cts.12692
- Harvard Personal Genome Project public data (CC0), participants hu5BB600, hu1C1368 and huE50E31: https://pgp.med.harvard.edu/
- Ensembl GRCh37 (grch37.ensembl.org) — variant positions and forward-strand alleles.
검사 결과 해석 가이드 · 약물유전체(PGx)
다른 검사 해석 글 전체 보기