Skip to contents

Build A Continuous Autofluorescence Basis From Raw Events

Source: R/get_af_basis_empirical.R
get.af.basis.empirical.Rd

As get.af.basis(), but derives the basis directly from the panel-oblique residual of real unstained single-cell events rather than from a SOM-derived spectral library. A SOM library is itself a quantisation of the unstained sample's cell-to-cell diversity into a fixed set of node shapes, so building the basis one step further upstream, from the events the SOM was trained on, removes that quantisation before it happens rather than fitting around it.

Structurally identical to get.af.basis(): both take the panel-oblique residual of a set of reference spectra (there, SOM nodes; here, raw events), take its SVD, and keep the leading out-of-span directions. Ranking by out-of-span variance rather than total variance matters here more than in the library case, because raw single-cell data also carries shot noise and any residual spillover contamination, both of which add variance that has nothing to do with autofluorescence shape.

Usage

get.af.basis.empirical(
  unstained.exprs,
  spectra,
  n.components = NULL,
  var.explained = 0.99,
  trim.quantile = 0.99,
  max.events = 50000,
  seed = 1
)

Arguments

unstained.exprs

Numeric matrix of raw unstained expression data, cells in rows and detectors in columns. Columns must match spectra.

spectra

Fluorophore spectral signatures, fluorophores in rows and detectors in columns.

n.components

Integer, number of basis directions to retain. Default NULL, in which case the count is chosen from var.explained.

var.explained

Numeric in (0, 1], default 0.99. Fraction of the out-of-span variance of the trimmed, subsampled event set the retained basis must capture. Ignored when n.components is supplied.

trim.quantile

Numeric in (0, 1], default 0.99. Events whose out-of-span residual norm exceeds this quantile are excluded before the basis is built. Real unstained samples can carry debris, doublets, or residual spillover-contaminated events; the SVD has no built-in robustness to them, so they are trimmed explicitly rather than left to dominate the leading components. Set to 1 to disable.

max.events

Integer, default 5e4. If more events remain after trimming, a random subsample of this size is used to build the basis. The basis is then applied to every event via unmix.af.basis(), so this only bounds the cost of basis construction, not of applying it.

seed

Integer, default 1. Seed for the subsampling step.

Value

A list with basis (detectors x components, real combinations of the sampled events), directions (detectors x components, the orthonormal out-of-span directions), sigma (singular values), var.explained (cumulative fraction captured), n.components, and n.trimmed (events excluded by trim.quantile). Compatible with unmix.af.basis().

See also

get.af.basis, unmix.af.basis