Title: scBatchProx: Leveraging Federated-Inspired Optimization to Ensure Stable Cell-Type Discriminability Amidst Heterogeneous Batch Compositions

Abstract: Standard single-cell integration pipelines typically generate low-dimensional cell embeddings, which are subsequently refined using post-hoc techniques to mitigate batch effects. However, this refinement stage often proves unstable when cell-type distributions differ significantly across batches, particularly when certain populations are sparse or entirely missing in specific datasets. This challenge is exacerbated in dynamic single-cell data ecosystems, where incoming batches may introduce shifts in both technical parameters and cellular composition. Such instability can compromise downstream cell-type classification accuracy and diminish robustness against imbalance perturbations. To address this, we present scBatchProx, a lightweight post-hoc refinement strategy designed to stabilize single-cell latent embeddings within heterogeneous and evolving environments. Operating on precomputed embeddings, scBatchProx adopts a federated-inspired optimization framework, treating each batch or study as an individual client. The method employs a batch-conditioned FiLM adapter to learn local latent updates, while proximal and identity-preserving regularization constraints ensure these modifications remain conservative. Evaluations across multi-batch and cross-study single-cell datasets demonstrate that scBatchProx enhances downstream cell-type classification performance regardless of the upstream embedding method. Furthermore, under controlled imbalance perturbations, the method preserves more stable F1 scores for affected cell types even when specific populations are downsampled or removed from a batch. In scenarios involving cumulative retraining and continual integration, scBatchProx proves effective as new datasets are incorporated over time. Collectively, these findings indicate that conservative, federated-inspired refinement can sustain stable single-cell embeddings despite changing batch compositions across datasets and temporal shifts.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC