Interferon-stimulated genes were being scored as batch effects: separating condition from batch before ranking genes

Gene-level batch-effect assessment can now subtract condition effects before deciding whether a gene is truly batch-sensitive, so expression differences caused by disease stage, drug treatment or stimulation are not booked as batch variation. The earlier GTE scored each gene from the difference between within-group variation and across-batch variation inside the same cell type, mixing condition-driven differences with batch fluctuation; its authors had already noted that genes flagged as highly batch-sensitive in interferon-beta data were enriched in interferon activation pathways. The evaluation used the paired PBMC dataset of Kang et al. (2018), with samples from eight systemic lupus erythematosus donors each split into untreated control and 6-hour interferon-beta stimulation, totaling 29,065 cells; 3,000 genes were randomly sampled from the transcriptome, and repeated sampling gave qualitatively consistent score distributions and correction results.

The correction first gives each gene a condition-driven score, then uses that score to shrink its original batch sensitivity, suppressing strongly condition-driven genes more. The condition score comes from scDisInFact, a disentanglement model that splits expression into shared biological factors, condition-specific biological factors and batch factors in three latent spaces, and uses a group lasso to make the first-layer weights of the non-shared encoder sparse at the gene-group level. Each gene's score is the l2 norm of the corresponding column vector in that weight matrix; the score distribution is strongly right-skewed, with most genes near zero and LY6E and IRF7 at 1.38 and 0.84 in the right tail, and LY6E consistently high in stimulated cells across all eight donors.

Penalty strength is set by an adaptive threshold rule that simultaneously limits two things: how far known condition-responsive genes are pushed down and how much technical control genes are disturbed. A larger penalty exponent k suppresses more aggressively but risks penalizing genes driven by technical noise, while a smaller k under-corrects, so a data-driven trade-off is needed. With 18 canonical type-I interferon-stimulated genes as the positive set and 15 RPS/RPL ribosomal protein genes as the negative set, the 99th percentile of negative-set rank shifts at k=1 was 12.9; mean negative-set shifts stayed within that threshold for k=1, 2 and 3 but exceeded it from k=4, and among feasible values k=3 maximized the positive-set mean shift at 870.2, versus 534.2 and 692.0 for k=1 and k=2.

The demonstrated scope is one paired interferon-beta-stimulated human PBMC dataset with eight donors and 29,065 cells, analyzed on 3,000 randomly sampled genes. The method is positioned as a diagnostic gene-screening tool before downstream dimensionality reduction or visualization, not a replacement for data integration or batch correction; tasks that must preserve original expression levels should still use raw data. The source proposes extending the framework to multi-condition datasets with more complex designs and evaluating it at whole-genome scale.

The adaptive threshold prioritizes controlling false positives in the negative set, which may make the correction conservative and under-correct condition-driven genes in some datasets. The negative set assumes ribosomal genes' batch sensitivity is mainly technical; if ribosomal expression is itself biologically perturbed in a given context, the threshold would be underestimated and k selection affected. Validation covers one interferon-beta-stimulated dataset and a random gene subset, so generality across datasets, tissues and experimental designs remains to be validated; when the condition score exceeds 1 the penalty sets the adjusted score to zero, an operational rule of the correction procedure rather than evidence that all of that gene's variation is biological.