mintyr
mintyr turns "many groups × many variables" data into analysis-ready pieces — and back into files — with data.table speed and no surprises.
A lot of real-world analysis follows the same loop: take a wide table, split it by trait and group, fit something on every piece (often with cross-validation), then write each piece or result to its own file or sheet. mintyr packages that loop into a small set of consistent functions:
files ──► import ──► reshape & nest ──► cross-validate / summarise ──► export ──► files
(xlsx/csv) (by trait × group) (per piece) (txt/csv/xlsx)
It was born in animal breeding, where this loop is everywhere (multi-trait, multi-breed, multi-farm data feeding ASReml-R, HIBLUP or DMU), but nothing in it is specific to genetics.
Who is it for?
| Field | Typical use |
|---|
| Animal & plant breeding | Per-trait / per-breed models, genetic correlations across traits or environments, input files for HIBLUP or DMU |
| Multi-site / multi-batch experiments | The same analysis for every site × variable, with results written back per site |
| Grouped machine learning | Reproducible (stratified) k-fold CV run separately within every subgroup |
| Reporting | Read dozens of Excel workbooks, clean them in one table, write them back to the original file/sheet layout |
| Any grouped analysis | Pairwise correlations between many variables, top/bottom-N% summaries per group |
Installation
# From CRAN
install.packages("mintyr")
# Development version from GitHub
pak::pak("tony2015116/mintyr")
Quick start
Fit a model for every trait within every group, using 4-fold cross-validation, and export each subset to its own folder:
library(mintyr)
library(data.table)
# 1. Wide -> long, nested by trait (`name`) and group (`am`)
nested <- w2l_nest(mtcars, cols = c("mpg", "qsec"), by = "am")
nested
#> name am data
#> 1: mpg 1 <data.table[13x9]>
#> 2: mpg 0 <data.table[19x9]>
#> 3: qsec 1 <data.table[13x9]>
#> 4: qsec 0 <data.table[19x9]>
# 2. Reproducible 4-fold CV inside every trait x group
cv <- nest_cv(nested, v = 4, seed = 2026)
# 3. Predictive ability per fold, then averaged
cv[, r := mapply(function(tr, va) {
fit <- lm(value ~ wt + hp, data = tr)
cor(predict(fit, va), va$value)
}, train, validate)]
cv[, .(mean_r = round(mean(r), 3)), by = .(name, am)]
#> name am mean_r
#> 1: mpg 1 0.816
#> 2: mpg 0 0.881
#> 3: qsec 1 0.884
#> 4: qsec 0 0.858
# 4. One file per trait x group: <path>/<name>/<am>/data.txt
export_nest(nested, path = file.path(tempdir(), "by_trait"))
What's inside
Import and export — a lossless round trip
| Function | What it does |
|---|
import_xlsx() | Read many workbooks and sheets into one table, with excel_name / sheet_name source columns; optional parallel reading (workers) |
import_csv() | Read many CSV/TXT files with fread(), with a source-file column |
export_xlsx() | Write a table back to the original file/sheet layout, or a list of tables into one workbook or one file each |
export_nest() | Write every nested table to <path>/<group1>/<group2>/<column>.txt |
export_list() | Write every element of a named list to its own file (names may contain / for sub-folders) |
Reshape and nest
| Function | What it does |
|---|
w2l_nest() / w2l_split() | Wide → long, then nest (list-column) or split (named list) by variable and group |
c2p_nest() | All pairs (or triples, …) of variables, renamed to value1, value2, … for uniform pairwise analysis |
r2p_nest() | Same variable across levels of another column (e.g. one column per environment), aligned by an ID |
Cross-validation and summaries
| Function | What it does |
|---|
split_cv() / nest_cv() | Repeated, optionally stratified k-fold CV for a list of tables or a nested table; returns fold indices and (optionally) the subsets |
top_perc() | Statistics of the top / bottom X% per variable and group |
format_digits() | Format numeric columns for reports (decimals, percentages) |
get_path_info() | Extract file names or path segments, cross-platform |
Examples by use case
Pairwise correlations between many variables, per group
pairs <- c2p_nest(mtcars, cols = c("mpg", "hp", "wt"), by = "am")
pairs[, .(r = sapply(data, function(d) cor(d$value1, d$value2))), by = .(pairs, am)]
The same trait in different environments (e.g. genotype × environment)
# growth: one row per animal x farm
r2p_nest(growth, names_from = "farm", cols = c("adg", "backfat"), id = "animal")
# -> per trait: animal | farm1 | farm2, ready for a cross-environment correlation
Clean many Excel files and put them back where they came from
dt <- import_xlsx(list.files("raw", pattern = "\\.xlsx$", full.names = TRUE))
dt <- dt[!is.na(weight)] # any data.table processing
export_xlsx(dt, path = "cleaned") # same file names, same sheet names
Input files for command-line breeding software
export_nest(nested, path = "hiblup_runs", na = "NA") # HIBLUP
export_nest(nested, path = "dmu_runs", na = "-9999", # DMU: numeric missing code,
col.names = FALSE) # no header
Design principles
- No side effects. Your input objects are never modified by reference.
- No silent data loss. File names are sanitised; if two outputs would land on the same path, mintyr stops instead of overwriting.
r2p_nest() refuses ambiguous IDs instead of returning counts.
- Consistent arguments.
data is the input, cols are the columns to work on, by are grouping columns, out_type chooses data.table ("dt") or data.frame ("df"), path is where files go.
- Light dependencies.
data.table, readxl, writexl and base R; cross-validation is implemented in the package itself.
- Reproducible.
seed makes CV folds reproducible without disturbing your global random-number stream.
Upgrading from 0.1.x
Version 0.2.0 unifies argument names (cols2l, cols2bind, trait, nest_dt, group_cols, export_path, rbind, … → cols, data, by, path, combine, …); the old names are no longer accepted. Cross-validation no longer depends on rsample: the splits column is replaced by train_idx / validate_idx. See NEWS for the complete old → new table.
Cheat sheet
A one-page quick reference for every function (printable as a landscape PDF):
English · 中文
Use with AI agents
mintyr ships an Agent Skill, mintyr-pipelines, that lets Claude and other agents run mintyr workflows from a plain-language request ("combine these Excel files and give me mean ± SD by breed"). Upload mintyr-pipelines.zip (Chinese: mintyr-pipelines-zh.zip) in Claude under Settings › Capabilities › Skills, or copy SKILL.md to ~/.claude/skills/mintyr-pipelines/ for Claude Code. Full instructions: English · 中文.
Acknowledgments
AI assistants helped turn the initial ideas behind mintyr into code, refine its structure and write its documentation.
