rcompat-locfit

rcompat-locfit is a Rust crate for selected R locfit-compatible one-dimensional local regression behavior. The first target is the local dispersion trend path used by DESeq2-style workflows.

This is a clean-room Rust implementation. It is not a binding to R locfit, does not call R at runtime, and does not contain code copied or translated from R locfit, DESeq2, or other copyleft projects. R locfit is used only as a black-box oracle for generating numeric test fixtures.

Current status: validated for targeted DESeq2 local-dispersion compatibility work. The DESeq2 wrapper matches black-box R locfit very closely on the real DESeq2-derived fixtures used during development. Broader exact R locfit parity is still work in progress, especially around adaptive evaluation grids, ties, and singular local designs.

Scope

• Pure Rust local polynomial regression in one predictor.
• Tricube kernel, adaptive nearest-neighbor bandwidth, and prior weights.
• Local fitted values and local slopes, so future R-style Hermite interpolation can be implemented on top of the direct fits.
• DESeq2-oriented wrapper:
  • normal-scale means and dispersions as input;
  • log(mean) and log(dispersion) fit internally;
  • original means as prior weights;
  • an approximate R locfit-style Hermite evaluation grid for larger fits;
  • floor-based nearest-neighbor bandwidth selection for R-style prediction;
  • R-style handling for tiny rank-deficient quadratic cells;
  • R-style weighted-quadratic extrapolation outside the evaluation range;
  • normal-scale dispersion predictions as output;
  • optional log-column constructor and log-dispersion predictions for callers that already have R-computed log(mean) / log(dispersion) values.

This crate is not a full port of R locfit and does not attempt to clone its public API.

Parity

The current implementation is validated against black-box R locfit fixtures. Recent local measurements were run with R 4.6.0 and locfit 1.5-9.12:

The synthetic matrix intentionally includes edge cases that are outside the main DESeq2-sized path. Public-wrapper rows report normal-scale relative error. The 2026 log-API rows use stored log columns to avoid normal-scale TSV rounding; their reported errors are absolute log-dispersion differences.

Usage

use rcompat_locfit::{
    fit_deseq2_local_dispersion_trend, fit_deseq2_local_dispersion_trend_from_logs,
};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let means = vec![1.0, 2.0, 5.0, 10.0, 100.0, 1000.0];
    let disps = vec![0.5, 0.3, 0.2, 0.12, 0.06, 0.03];

    let trend = fit_deseq2_local_dispersion_trend(&means, &disps, 1e-8)?;
    let predicted = trend.predict(&[3.0, 30.0, 300.0])?;

    let log_means: Vec<_> = means.iter().map(|mean| f64::ln(*mean)).collect();
    let log_disps: Vec<_> = disps.iter().map(|disp| f64::ln(*disp)).collect();
    let log_trend =
        fit_deseq2_local_dispersion_trend_from_logs(&log_means, &log_disps, &means, 1e-8)?;
    let predicted_log_disp = log_trend.predict_log_dispersion(&[30.0_f64.ln()])?;

    println!("{predicted:?}");
    println!("{predicted_log_disp:?}");
    Ok(())
}

Fixture Generation

Fixtures are generated with R locfit as a black-box oracle. The script does not require DESeq2.

Rscript fixtures/r/generate_locfit_fixtures.R
cargo test --test fixtures -- --ignored

The committed real DESeq2-derived subset can be regenerated from the ignored local /data/ debug tables with:

Rscript fixtures/r/generate_real_deseq_subset_fixture.R

The committed 2026 hard-real subset can be regenerated from the ignored 2026 hard-real bundle with:

Rscript fixtures/r/generate_hard_real_2026_subset_fixture.R

The default cargo test run includes the committed synthetic fixture matrix and the committed real-derived subset parity checks. It does not require R.

License

MIT. See LICENSE.