EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Arithmetic power geometry adds mathematically explicit structural information to classical and principal-component representations of single-cell PBMC profiles

Mohammad Amir Khusru Akhtar

Frontiers in Bioinformatics · 18.11.2026 · 6

DOI: 10.3389/fbinf.2026.1961687 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Frontiers in Bioinformatics, online erschienen am 18.11.2026. Verfasst von Mohammad Amir Khusru Akhtar. Die Arbeit nennt 27 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Single-cell analyses require representations that retain biological structure while remaining mathematically transparent. Arithmetic power geometry (APG) describes a nonnegative molecular profile through normalized squared weights and an exponent-indexed deformation curve, but its value depends on showing complementarity rather than claiming to replace latent embeddings. We therefore rebuilt the evaluation around three tests: direct ablation against classical summaries, explicit comparison with principal-component representations, and validation using labels defined independently of APG. In the 10x Genomics Fresh 68k PBMC transcriptomic benchmark (68,579 cells), five-fold classification showed that program-level classical summaries achieved macro-F1 0.2838 ± 0.0078 , program APG achieved 0.3806 ± 0.0135 , and their combination reached 0.4127 ± 0.0108 . Across gene- and program-level summaries, adding APG increased the macro-F1 from 0.4258 ± 0.0100 to 0.5333 ± 0.0035 . A feature-number-matched nonlinear control reached only 0.4445 ± 0.0056 at the same 38-feature dimensionality; APG exceeded this control in all five folds by a mean macro-F1 difference of 0.0887 (95% fold-level t interval 0.0858–0.0917). In a stronger nested-CV sensitivity analysis, the equal-dimensional polynomial/log and random Fourier controls reached a macro-F1 of 0.4594 ± 0.0079 and 0.4544 ± 0.0118 , respectively, versus 0.5404 ± 0.0031 for APG; learning curve and per-cell-type analyses likewise preserved the APG advantage. High-exponent gene-level defects saturated, identifying 2 < p ≲ 3 as the useful transition regime. An external analysis used a compact Azimuth/Hao PBMC reference containing 36,433 cells, frozen eight-class annotations, 228 normalized antibody-derived features, and a 50-dimensional RNA reference principal component analysis (PCA). With fold-specific, 30-component, antibody-derived tag (ADT) PCA, adding APG increased the macro-F1 score from 0.9382 ± 0.0017 to 0.9590 ± 0.0033 ; after classical ADT summaries were included, APG still increased the macro-F1 scores from 0.9468 ± 0.0026 to 0.9594 ± 0.0030 . These results support APG as a compact, mathematically explicit structural layer that contributes measurable information beyond matched classical summaries and augments conventional principal-component representations in the tested settings. Fold-level comparisons are treated as exploratory rather than as independent biological replication, and the external same-modality claim is limited to ADT because the compact Azimuth object lacks an RNA expression matrix.

Zitieren

APA

Akhtar, M. A. K. (2026). Arithmetic power geometry adds mathematically explicit structural information to classical and principal-component representations of single-cell PBMC profiles. Frontiers in Bioinformatics, 6. https://doi.org/10.3389/fbinf.2026.1961687

Vancouver

Akhtar MAK. Arithmetic power geometry adds mathematically explicit structural information to classical and principal-component representations of single-cell PBMC profiles. Frontiers in Bioinformatics. 2026;6. doi:10.3389/fbinf.2026.1961687

BibTeX
@article{Akhtar2026,
  author = {Akhtar, Mohammad Amir Khusru},
  title = {Arithmetic power geometry adds mathematically explicit structural information to classical and principal-component representations of single-cell PBMC profiles},
  journal = {Frontiers in Bioinformatics},
  year = {2026},
  volume = {6},
  doi = {10.3389/fbinf.2026.1961687}
}
Kostenlos starten →