Converts biomarker expression profiles into pairwise log-ratio images and classifies them using a convolutional neural network. The encoding is additive-shift-invariant in log-space: any per-sample additive constant applied uniformly across all features cancels exactly in the pairwise difference pixel(i,j) = log2(x_i) - log2(x_j). This yields a reference-cohort-free representation that preserves within-sample relationships while removing uniform sample-level offsets.
Details
For a panel of \(p\) biomarkers, each sample becomes a \(p \times p\) image where pixel(i,j) = log2(biomarker_i / biomarker_j). On log-scale data this is simply the difference \(x_i - x_j\). The default CNN uses two 3x3 convolutions, adaptive average pooling to 3x3, a 64-unit dense layer with dropout 0.3, and a sigmoid output head.
The torch backend is used when available. The learner wrapper stores the fitted torch module and applies the same training-fold normalization to new samples during prediction.
References
Aitchison J. (1986). The Statistical Analysis of Compositional Data. Chapman and Hall.
Egozcue JJ, Pawlowsky-Glahn V, Mateu-Figueras G, Barcelo-Vidal C. (2003). Isometric logratio transformations for compositional data analysis. Mathematical Geology, 35(3), 279-300.
Gloor GB, Macklaim JM, Pawlowsky-Glahn V, Egozcue JJ. (2017). Microbiome datasets are compositional: and this is not optional. Frontiers in Microbiology, 8, 2224.
Quinn TP, Erb I, Richardson MF, Crowley TM. (2020). Understanding sequencing data as compositions: an outlook and review. Bioinformatics, 36(16), 4424-4432.