AltraFlowSOM: A Semi-Supervised Framework for Imaging Mass Cytometry Phenotyping

Imaging Mass Cytometry (IMC) allows for the simultaneous quantification of multiple protein markers within individual cells in tissue samples. However, accurately phenotyping cells at scale remains a significant challenge. Traditional unsupervised clustering methods often fail to maintain coherent populations, while supervised classifiers impose limitations on the vocabulary of cell types.

Rare cell subtypes, which may encode clinically relevant information, can also negatively impact detection performance when combined with abundant cell types.

The research team introduces AltraFlowSOM, a semi-supervised framework built upon the FlowSOM algorithm. AltraFlowSOM incorporates expert annotations into the self-organizing map training process using a two-layer SuperSOM architecture. This approach balances the guidance provided by labeled reference points with the unsupervised discovery of cell populations.

By anchoring the map to biologically labeled points, AltraFlowSOM eliminates the need for batch correction before clustering. In a study comparing AltraFlowSOM to unsupervised and supervised baselines on two independent IMC cohorts - Lupus Nephritis (n=22 ROIs) and Sjogren syndrome (n=10 ROIs) - the semi-supervised method outperformed all other approaches.

Metrics such as Adjusted Rand Index, F1 scores (both macro and weighted), weighted purity, and the identification of rare cell populations were significantly improved with AltraFlowSOM.

One of the most notable findings was the median Treg cell recovery, which surpassed that of all other methods. The study demonstrates that AltraFlowSOM effectively resolves the scalability-alignment-discovery trilemma in high-dimensional IMC phenotyping by offering a generalizable semi-supervised SOM framework.