VoxelCoder: Classification of human cellular phenotypes via autoencoder batch alignment and hyperdimensional representation of cytometry data.

Technical variations in sample processing and instrument calibration, known as batch effects, can obscure true biological signals in cytometry data, impeding the integration of large-scale datasets. We present an autoencoder neural network approach that achieves batch correction comparable to that in existing methods while better preserving biological variation. Following alignment, cellular datasets are projected into a hyperdimensional voxel space that maintains interpretable marker-based features without requiring abstract latent dimensions, ensuring that identified cell populations remain fully interpretable. We benchmark this approach using a purpose-generated mouse splenocyte dataset with synthetic batch effects, demonstrating superior biological signal preservation compared to existing tools. Applied to clinical datasets, our method enables identification of cellular phenotypes associated with cytomegalovirus serostatus and COVID-19/sepsis discrimination, outperforming alternative approaches in downstream classification tasks. This framework addresses key technical limitations in integrating multi-batch cytometry datasets and provides a foundation for machine-learning applications in cytometry-based diagnostics.