Fixing biological data without erasing it?

In single-cell genomics, data collected in different labs often carry technical issues that confuse results. Traditional methods correct these errors but sometimes erase true biological information. A new method, ConDo, fixes technical data without touching biological info. It works like a smart adapter: instead of aligning all data together, it adjusts each pair of batches based on cell type. This avoids losing important details. To handle many batches at once, researchers use a compatibility graph: batches sharing cell types are merged step by step. The result? Corrected data ready for advanced analysis, preserving original biological richness. Tested on multiple datasets, ConDo outperforms other gene-preserving methods and is competitive with more complex ones. A step forward for studying diseases and development with greater precision.

The main challenge in single-cell batch integration is the confounding between technical effects and biological signals. When cell types vary across batches, methods aligning marginal distributions risk overwriting genuine biological variation. ConDo addresses this by applying Conditional Distribution matching, a confounder-aware domain adaptation method that corrects conditional distributions given cell-type annotations. This preserves cell-type-specific biological information, avoiding distortions from heterogeneous cellular compositions. To extend ConDo to multi-batch scenarios, a compatibility-graph integrator was developed: batches are graph nodes, connected when sharing a cell type. The algorithm performs agglomerative merges, selecting the best-scoring batch and adapting it to the current reference via a ConDo adapter. This iterative process ensures consistent and scalable correction. Evaluated on the Open Problems Batch Integration Benchmark, ConDo excels as the top feature-space method, ranking first on five of six datasets. It also competes with deep embedding methods, leading overall on four of six datasets. Its ability to return corrected gene expressions rather than opaque embeddings makes it especially valuable for downstream analyses like biomarker discovery and differential expression studies.