#9 · ML / AI methods

Tailored representations for dimensionality reduction and sequence encoding in biology

The dominant thread is replacing generic high-dimensional representation steps with alternatives matched to the structure of biological data. Dimensional Cophenetic Integrity provides a metric for judging whether dimensionality reduction preserves structure in mass spectrometry imaging data. The Komlós-Hadamard transform encodes DNA and RNA sequences as an alternative to One-Hot encoding, reporting lower compute cost. The Rayleigh quotient preprint extends contrastive PCA with a kernel-weighted spatial variant (k-rhoPCA) and a basis-coefficient functional variant (f-rhoPCA). What is new across these items is that the representation objective is conditioned on modality, whether spatial and functional structure, cophenetic structure in imaging, or sequence-level interpretability. These items differ in data type and method class, so the thread is not uniform across all of them. Two items are published and one is a preprint, so the results remain preliminary.