# JAK-centric explainable few-shot gene-expression diagnosis framework for alopecia via MultiPLIER priors and relation-style set-to-set comparison

## TL;DR

The paper introduces a new AI framework for diagnosing alopecia using gene expression data, identifying a four-gene loop linked to JAK-STAT signaling in alopecia areata.

## Contribution

A novel explainable AI framework using MultiPLIER priors and a relation-style comparator for small-cohort genomic diagnosis is proposed.

## Key findings

- A four-gene axis (IL2RB, IL2RG, EOMES, GZMA) is selectively activated in alopecia areata at both mRNA and protein levels.
- The proposed AI framework enables robust diagnosis with limited samples and retains mechanistic interpretability.
- The JAK-STAT hyperactivation in alopecia areata is linked to cytotoxic T-cell activity and potential drug targets like JAK inhibitors.

## Abstract

Alopecia areata (AA) and androgenetic alopecia (AGA) both present with hair loss but require different therapies, and reliable biomarkers to guide treatment remain lacking. We integrated bulk and single-cell RNA-seq to compare JAK–STAT signaling in AA versus AGA. In AA, 257 immune-enriched differentially expressed genes (DEGs) were identified; WGCNA and consensus machine learning (LASSO, SVM-RFE, random forest) yielded six candidate hub genes, and external validation narrowed these to four key genes–granzyme A (GZMA), interleukin-2 receptor
β
(IL2RB) and
γ
(IL2RG) chains, and eomesodermin (EOMES). Building on these biology-anchored features, we introduced an interpretable few-shot deep learning classifier as an explainable AI alternative to a nomogram: bulk expression profiles are projected onto pathway/cell-type–aligned MultiPLIER latent variables (a frozen prior), the latent channels are re-weighted via element-wise multiplication with the expression levels of the key hub genes, and a Relation-style set-to-set comparator then aggregates support–query similarities (Hadamard mapping + permutation-invariant aggregation) before a shallow head predicts AA versus control. This prior-informed approach enables robust discrimination under limited sample conditions while retaining mechanistic interpretability, thereby exemplifying a next-generation XAI solution for small-cohort genomic diagnosis. Cross-database functional annotation and wet-lab validation (RT-qPCR and Western blot) in independent AA/AGA/healthy scalp samples confirmed that the IL2RB/IL2RG–EOMES–GZMA axis is selectively activated at both mRNA and protein levels in AA. Single-cell analysis localized GZMA to cytotoxic T cells and IL2RG to proliferating lymphocytes, outlining an
EOMES+

CD8+
T-cell GZMA–IL2RB/IL2RG cytotoxic loop driving JAK–STAT hyperactivation in AA. Drug–gene network analysis linked these targets to JAK inhibitors and cyclosporine. AGA showed no comparable JAK–STAT perturbation, consistent with its androgen-centric biology. In summary, this four-gene loop provides a non-invasive AA biomarker and a tractable target for precision JAK blockade, while the proposed few-shot framework offers a general, prior-driven alternative to nomograms for transcriptomic diagnosis in small cohorts, illustrating an XAI-driven diagnostic approach for precision medicine.