Comparative evaluation of reference-free transcriptomic deconvolution highlights the importance of biological...

Astrocytes are central regulators of neuronal energy metabolism and redox homeostasis-processes that become progressively disrupted across the Alzheimer's disease (AD) continuum. However, bulk transcriptomic data obscure cell-type-specific signals, and existing reference-free deconvolution methods often prioritize either statistical robustness or quantitative accuracy without fully integrating both dimensions.

Here, we present a comparative framework evaluating two complementary unsupervised approaches, CDSeq and DECODER, to reconstruct astrocyte-associated transcriptomic profiles from human hippocampal samples spanning control, mild cognitive impairment (incipient and moderate), and AD (severe) stages (GSE28146; n = 30). The inferred profiles were functionally contextualized through integration into a genome-scale metabolic model of human astrocytes, enabling the assessment of system-level metabolic alterations associated with disease progression.

Our results reveal consistent dysregulation of key astrocytic pathways, including impairment of the astrocyte-neuron lactate shuttle, disruption of glutamine metabolism, and reduced glutathione-mediated an oxidant capacity. Methodological benchmarking showed distinct yet complementary performance profiles: DECODER achieved higher accuracy in reconstructing global expression magnitudes, whereas CDSeq exhibited greater stability and preservation of gene-gene relationships. Crucially, external validation using independent single-nucleus RNA-seq astrocyte data demonstrated that CDSeq-derived profiles achieve moderate but robust concordance with reference signatures (r ≈ 0.43-0.44), substantially exceeding DECODER-derived concordance (r ≈ 0.19-0.23), with higher concordance with astrocyte-associated signatures, alongside preservation of canonical astrocyte markers and enrichment of astrocyte-specific pathways, indicating superior biological coherence.

Together, these findings demonstrate that technical accuracy does not necessarily translate into biological validity and highlight CDSeq as the method that more reliably captures astrocyte-specific transcriptional programs in this context. While DECODER remains valuable for detecting absolute expression changes, CDSeq provides a more consistent recovery of astrocyte-associated transcriptional patterns. More broadly, our results support the incorporation of biological validation alongside statistical benchmarking when selecting deconvolution methods for downstream systems biology and metabolic modeling applications. This framework establishes a reproducible strategy for evaluating deconvolution methods and their functional consequences, advancing the interpretation of bulk transcriptomic data in neurodegenerative disease.