#20 · ML / AI methods

Null-model specification governs false discovery and bias in omics inference

The dominant thread is that statistical conclusions in omics depend on the correctness of the underlying data model. Two published studies address this directly for count data: a proportionality analysis quantifies bias in empirical metrics on count compositional data and proposes logit-normal multinomial model estimates, while a simulation study shows that scale misspecification inflates false discovery rates in high-throughput sequencing differential expression tools, with ALDEx2 and ALDEx3 scale models reducing this effect. The remaining preprints extend the theme more loosely: TCRdenoise applies an unsupervised similarity-based denoising step with a modified silhouette score to TCR-pMHC specificity data, AltraFlowSOM embeds partial expert labels into a SOM framework for imaging mass cytometry, and a benchmark of 16 dimensionality reduction methods uses 12 path-preservation metrics across point-density thresholds. What is new is the explicit treatment of model specification as the locus of error, rather than any single algorithm.