1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
//! Negative binomial differential expression for bulk and single-cell RNA-seq.
//!
//! A Rust port of the edgeR numerical stack (normalisation, Cox-Reid dispersion
//! estimation, the Levenberg-damped negative binomial GLM, quasi-likelihood
//! weights, the exact test), the limma linear model stack (`squeezeVar`, the
//! F-distribution fits, voom, `lmFit`, `contrasts.fit`, `eBayes`, `topTable`),
//! and NEBULA, a negative binomial gamma mixed model for single cell.
//!
//! The bulk stack follows edgePython, itself a port of edgeR and limma, with
//! the R winning wherever the two disagree. NEBULA is ported from the `nebula`
//! package's own C++ instead.
//!
//! ### Numeric policy
//!
//! Public containers and fits are generic over [`prelude::EdgeFloat`], so
//! single-cell counts can be held as `f32` and halve the memory. Likelihood
//! evaluation, Cox-Reid log-determinants, the optimisers and every p-value run
//! in `f64` regardless of `T`. A negative binomial deviance is a difference of
//! large logs, and accumulating one in `f32` loses parity with edgeR well
//! before it saves anything worth having.
//!
//! ### Parallelism
//!
//! Genes are the parallel axis almost everywhere. Counts are stored gene-major,
//! so one gene is a contiguous slice, and the fan-out is a rayon iterator over
//! genes with a per-thread scratch buffer.
//!
//! Normalisation is the exception, and deliberately so: TMM has no gene axis to
//! fan out over, since the unit of work is one sample compared against a
//! reference column. It parallelises over samples and transposes the counts
//! once on entry so each sample is contiguous. Any module that departs from the
//! gene-major rule says so in its own header.