Merging nucleus datasets by correlation-based cross-training