Benchmarking single cell transcriptome matching methods for incremental growth of reference atlases DOI Creative Commons
Joyce Hu, Beverly Peng, Ajith V. Pankajam

et al.

Published: April 16, 2025

Abstract Background The advancement of single cell technologies has driven significant progress in constructing a multiscale, pan-organ Human Reference Atlas (HRA) for healthy human cells, though challenges remain harmonizing types and unifying nomenclature. Multiple machine learning artificial intelligence methods, including pre-trained fine-tuned models on large-scale atlas data, are publicly available the community users to computationally annotate match their clusters reference atlas. Results This study benchmarks four computational tools type annotation matching – Azimuth, CellTypist, scArches, FR-Match using two lung datasets, Lung Cell (HLCA) LungMAP single-cell (CellRef). Despite achieving high overall performance while comparing algorithmic annotations expert annotated variations accuracy were observed, especially annotating rare types, underlining need improved consistency across prediction methods. benchmarked methods used cross-compare incrementally integrate 61 from HLCA 48 CellRef, resulting meta-atlas 41 matched 20 HLCA-specific 7 CellRef-specific types. Conclusion reveals complementing strengths presents framework incremental growth inventory atlases, leading 68 unique CellRef HLCA. benchmarking analysis contributes improving coverage quality HRA construction by assessing reliability approaches transcriptomics datasets.

Language: Английский

Benchmarking single cell transcriptome matching methods for incremental growth of reference atlases DOI Creative Commons
Joyce Hu, Beverly Peng, Ajith V. Pankajam

et al.

Published: April 16, 2025

Abstract Background The advancement of single cell technologies has driven significant progress in constructing a multiscale, pan-organ Human Reference Atlas (HRA) for healthy human cells, though challenges remain harmonizing types and unifying nomenclature. Multiple machine learning artificial intelligence methods, including pre-trained fine-tuned models on large-scale atlas data, are publicly available the community users to computationally annotate match their clusters reference atlas. Results This study benchmarks four computational tools type annotation matching – Azimuth, CellTypist, scArches, FR-Match using two lung datasets, Lung Cell (HLCA) LungMAP single-cell (CellRef). Despite achieving high overall performance while comparing algorithmic annotations expert annotated variations accuracy were observed, especially annotating rare types, underlining need improved consistency across prediction methods. benchmarked methods used cross-compare incrementally integrate 61 from HLCA 48 CellRef, resulting meta-atlas 41 matched 20 HLCA-specific 7 CellRef-specific types. Conclusion reveals complementing strengths presents framework incremental growth inventory atlases, leading 68 unique CellRef HLCA. benchmarking analysis contributes improving coverage quality HRA construction by assessing reliability approaches transcriptomics datasets.

Language: Английский

Citations

0