Top2All: Benchmarking sparse-to-transcriptome reconstruction from high-expression profiles across sex- and sex-steroid-defined conditions

Abstract

Transcriptome-wide profiling at single-cell resolution is limited by sparse detection of gene expression, raising the question of how much transcriptomic information can be recovered from a small set of highly expressed genes. We developed Top2All, a Mouse-Geneformer-based framework for reconstructing transcriptome-wide expression from sample-specific sparse inputs. Using 24 cell-number-normalized RNA-seq profiles of mouse adrenal zona fasciculata cells across eight experimental conditions defined by sex and sex-steroid milieu, we evaluated condition-aware and no-condition models under replicate-wise out-of-fold and leave-one-condition-out validation. Top2All achieved near-unity sample-wise correlation with measured expression, but similarly high correlations were obtained using simple reference profiles, whereas gene-wise evaluation revealed substantially greater differences among methods. Explicit condition information did not consistently improve reconstruction, indicating that sex- and sex-steroid-associated condition information was already strongly encoded in the sparse expression profiles. Nearest-neighbor retrieval provided a stronger baseline than learned reconstruction, and input-depth analysis showed that increasing the number of input genes did not monotonically improve performance. These findings show that a relatively small set of highly expressed genes contains substantial information about transcriptomic state, while accurate reconstruction of gene-specific biological variation remains challenging. Sparse-to-transcriptome models should therefore be evaluated using gene-wise metrics, strong reference baselines, and tests of generalization beyond conditions represented during training.