Mashford, B. | Search bioRxiv

bioRxiv2025

Benchmark of Wide Range of Pairwise Distance Metrics for Automated Classification of Mouse Mutant Phenotypes from Flow Cytometry Data

Precision medicine requires a comprehensive mapping of genotype to phenotype to provide patients with individually tailored treatment. However, when using flow cytometry to identify phenotypes, such as the quantity of various immune cell populations in tissue and blood used to identify autoimmune disorders, it is often unclear which cellular phenotypes are from healthy and disease individuals, especially when including the effects of population diversity, due to the high-dimensional nature of the data. To identify and segregate healthy phenotype from various disease phenotypes, we use pairwise distance metrics between each samples cell populations. By comparing distance metrics between C57BL/6 clone mice with mutations of known phenotype, we find that cosine similarity is best suited for segregating wildtype from mutant samples while respecting minute differences in already small cell populations, and that standardised Euclidean distance is best suited for machine-learning input due to its sensitivity. Both metrics outperform other tested metrics (including Aitchison, Euclidean, Manhattan, Earth-Movers Distance, and squared Euclidean). We demonstrate the utility of these different pairwise metrics through their application to a classification task of known mutant phenotypes: using an existing FACS phenotype dataset derived from X000 inbred C57BL/6 mice that harbour potentially phenotypic genetic variation introduced through ENU mutagenesis of individual pedigree-founding G0 male mice.

bioinformatics↗

bioRxiv2021

Linked machine learning classifiers improve species classification of fungi when using error-prone long-reads on extendedmetabarcodes

BackgroundThe increased usage of error-prone long-read sequencing for metabarcoding of fungi has not been matched with adequate public databases and concomitant analysis approaches. We address this gap and present a proof-of-concept study for classifying fungal taxa using linked machine learning classifiers. We demonstrate the capability of linked machine learning classifiers to accurately classify species and strains using real-world and simulated fungal ribosomal DNA datasets, including plant and human pathogens. We benchmark our new approach in comparison to current alignment and k-mer based methods based on synthetic mock communities. We also assess real world applications of species identification in complex unlabelled datasets. ResultsOur machine learning approach assigned individual nanopore long-read amplicon sequences to fungal species with high recall rates and low false positive rates. Importantly, our approach successfully distinguished between closely-related species and strains when individual read errors were higher than the genetic distance between individual taxa, which the alignment and k-mer methods could not do. The machine learning approach showed an ability to identify key species with high recall rates, even in complex samples of unknown species composition. ConclusionsA proof of concept machine learning approach using a tree-descent approach on a decision tree of classifiers can identify known taxa with high accuracy, and precisely detect known target species from complex samples with high recall rates. We propose this approach is suitable for detecting the known knowns of pathogens or invasive species in any environment of mostly unknown composition, including agriculture and wild ecosystems.

bioinformatics↗