Replicable dosage-sensitive programs are defined by conserved biological neighborhoods and coherent cell contexts, not by a single marker, one atlas, or one label ontology. Gene-level constraint alone does not establish a shared cellular program.
Executive interpretation
Dosage-sensitive genes are often discussed as a single class, but they participate in very different biological systems. A constrained transcription factor, a ribosomal protein, a chromatin regulator, and a structural protein may all be depleted for predicted loss-of-function variation while acting in different cell types and developmental windows.
This study asks which expression programs travel across independently produced atlases. We find that replication is strongest when defined at the level of a biological neighborhood—coherent sets of genes active in related cell contexts—rather than a universal dosage-sensitivity signature.
Program definition
Seed genes are drawn from curated dosage-sensitivity and population-constraint resources under explicit inclusion rules. Programs are then expanded using within-reference co-expression neighborhoods, but seed identity and neighbor identity remain distinct. This prevents a large inferred network from being mislabeled as direct dosage evidence.
Neighborhoods are estimated within donor-aware cell strata and summarized across references. Highly ubiquitous housekeeping modules, mitochondrial read fractions, ribosomal abundance, and cell-cycle programs are modeled as competing explanations. Program size is capped and resampled so large modules do not obtain an automatic enrichment advantage.
Cross-atlas harmonization
Cell labels are represented hierarchically. Broad classes establish whether a program transfers across studies; fine states determine where the mechanism may be specific. Original labels, mapped labels, and mapping confidence are retained together. We do not force a one-to-one correspondence when atlases describe different biological resolution.
Expression values are normalized within source, and replication is assessed through ranks, direction, neighborhood overlap, and cell-context concordance. Raw effect sizes are not directly pooled across incompatible platforms. A random-effects summary is used only when the underlying contrast and biological unit are sufficiently aligned.
Replication criteria
A program is considered transferable when its leading biological neighborhood remains enriched in the same broad lineage and compatible fine states across independent references, with no single atlas carrying the conclusion. Leave-one-atlas-out analysis must preserve the direction and the central functional interpretation.
A marker that is reproducible without its neighbors is classified as a replicated gene signal, not a replicated program. Conversely, modest gene-level overlap can still support replication when orthologous or functionally equivalent neighborhoods preserve the same process. This distinction is essential in atlases with different feature detection and annotation depth.
Technical findings
Broad cellular processes associated with transcriptional regulation, proteostasis, translation, and developmental control frequently appear among dosage-sensitive genes, but their cell-context specificity varies. Some neighborhoods are widespread and likely reflect general cellular vulnerability; others localize to developmental or lineage-specific transitions and are more useful for disease mechanism design.
The apparent strength of a dosage program often decreases after matching for baseline expression and annotation degree. The programs that remain coherent are carried by multiple genes and preserve a functional neighborhood across references. Atlas-specific programs are retained as hypotheses with narrow scope rather than discarded or generalized.
Interpretation and use
A replicated program can identify the cell system in which dosage perturbation is most likely to produce a measurable state change. It cannot by itself establish that reduced dosage is the disease mechanism for every seed gene. Variant consequence, inheritance, allelic series, and gene-specific functional evidence remain necessary.
The practical output is a versioned program catalog with source provenance, neighborhood membership, transfer status, null-model performance, and explicit non-transfer conditions. That catalog is designed to support experimental selection and external reanalysis.
What the analysis establishes
Neighborhoods replicate better than isolated markers
Conserved functional neighborhoods are more portable across atlases than one highly expressed gene or one source label.
Constraint does not imply one program
Dosage-sensitive genes partition into distinct biological and cellular contexts.
Label resolution changes specificity, not necessarily direction
Broad lineage agreement can coexist with disagreement over fine-state nomenclature.
Non-transfer is informative
A program limited to one tissue, developmental period, or protocol defines a bounded hypothesis rather than a failed study.
Conclusion and experimental handoff
We conclude that dosage sensitivity becomes mechanistically useful when it is attached to a replicated cell-contextualized program. The most defensible programs are multi-gene, preserve their biological neighborhood, and survive removal of any one atlas.
Experimental follow-up should perturb dosage in the cell context predicted by the transferable program and measure both the program-level response and gene-specific phenotypes. Rescue by restoring dosage provides substantially stronger evidence than reproduction of a generic stress response.
Limitations
- Atlas coverage is uneven across tissues, developmental periods, and disease states.
- Co-expression is compatible with shared regulation but does not prove direct molecular interaction.
- Harmonized labels can obscure source-specific biology if mapping confidence is ignored.
- Dosage sensitivity inferred from population depletion is not equivalent to proven haploinsufficiency for a specific disorder.
Glossary
- Dosage sensitivity
- A relationship in which altered gene copy number or expression level affects biological function or phenotype.
- Biological neighborhood
- A group of genes that repeatedly cohere by expression, pathway, or cell context.
- Random-effects summary
- A meta-analytic model that allows true effects to vary across studies.
- Transfer status
- A record of where a signal reproduces, partially reproduces, or fails to reproduce.
Reproducibility and evidentiary scope
This study is a REELD public-data analysis and methods interpretation. It does not report a newly recruited clinical cohort, classify an individual variant, or replace clinical review. A release-ready execution of the workflow includes accession-level provenance, source and ontology versions, code state, environment locks, predefined sensitivity analyses, and output checksums.