Phenotype-defined gene modules become biologically interpretable only after broad clinical terms, gene-set composition, developmental timing, and replicate structure are controlled. Stable convergence is concentrated in developmental programs, not in one universal “neurodevelopmental” cell type.
Executive interpretation
Many rare neurodevelopmental disorders share terms such as global developmental delay, intellectual disability, hypotonia, seizures, or abnormal behavior. Those terms are clinically important, but they are also common across genetically distinct conditions. If genes are grouped by broad vocabulary alone, almost any neuronal atlas can produce an apparently plausible enrichment.
We find that convergence is most credible when it appears at the level of a developmental program and survives both phenotype and gene-set perturbation. The analysis does not support a single cellular mechanism for all neurodevelopmental disease genes. It supports narrower modules that align with specific transitions such as progenitor proliferation, neuronal specification, migration, synaptogenesis, or excitatory-neuron maturation.
Phenotype module construction
Phenotype sets are represented as graphs, not bags of labels. Exact Human Phenotype Ontology terms are retained with onset, frequency, negation, and uncertainty when available. Ancestors are propagated for recall, while information-content weighting prevents very broad ancestors from dominating similarity. We compare Resnik-style most-informative-common-ancestor similarity with normalized alternatives and report where rankings depend on the metric.
Gene modules are created from phenotype-to-disease and disease-to-gene relationships under explicit evidence filters. A gene can participate in more than one module, which is biologically realistic but statistically consequential. Module overlap is therefore quantified, and enrichment is compared with degree-matched and expression-matched null sets so heavily annotated genes do not win merely because the knowledge graph contains more edges for them.
Developmental atlas analysis
The cortical references are organized along two simultaneous axes: cell lineage and developmental progression. Broad labels such as radial glia, intermediate progenitor, excitatory neuron, inhibitory neuron, glia, and vascular cell support cross-atlas transfer. Finer labels capture regional identity, maturation stage, and transient states. We retain both scales because broad agreement with fine-state disagreement is itself informative.
Module activity is evaluated with rank-based gene-set scoring and donor-aware pseudobulk summaries. Genes with extremely high abundance or broad housekeeping roles are capped or tested separately. We also calculate leading-edge membership: the subset of genes repeatedly responsible for a module score. A module that transfers only because one ubiquitous gene dominates is not treated as convergent biology.
Robustness and null architecture
Four perturbations are central to the design. We vary ontology propagation depth, remove one phenotype term at a time, resample the gene set while matching expression and annotation degree, and leave out each atlas or donor group in turn. Stability is evaluated by direction, cell-program neighborhood, leading-edge overlap, and rank concordance rather than by one significance threshold.
We also test generic neuronal, ribosomal, cell-cycle, stress, and synaptic gene programs as competing explanations. A disease module must show information beyond these controls. This is particularly important for synaptic genes, which are numerous, highly studied, and broadly detected across maturing neurons.
Technical findings
Broad developmental-delay modules show high recall but low localization: they are distributed across multiple neuronal lineages and are sensitive to highly annotated genes. When the phenotype representation is refined with seizure type, onset, morphologic features, movement abnormalities, or specific cognitive and behavioral features, the modules become smaller and more developmentally resolved.
The most reproducible convergence is observed as program-level agreement across stages of neuronal differentiation and maturation. Some modules localize to proliferative or early specification programs, while others become coherent after neuronal identity is established and synaptic machinery is assembled. The exact atlas label is less stable than the developmental ordering, which argues for interpreting convergence along trajectories rather than treating clusters as fixed biological entities.
What would falsify the interpretation
The shared-program hypothesis would be weakened if the module score vanished after removal of one broad phenotype term, if matched random sets produced equivalent localization, if the leading edge were dominated by one or two ubiquitous genes, or if the developmental ordering reversed across independent references. It would also be weakened if perturbation of representative genes produced unrelated cellular phenotypes under a matched experimental context.
These failure conditions are part of the output. They distinguish a reusable mechanistic module from a retrospective narrative assembled around a visually compelling embedding.
What the analysis establishes
Broad terms recover candidates but blur mechanism
High-level phenotype ancestors are useful for recall but are too nonspecific to localize a developmental program on their own.
Trajectory position transfers better than cluster name
Independent atlases disagree in taxonomy more often than they disagree in the ordering from progenitor to differentiated neuronal programs.
Leading-edge stability is essential
A module is credible when a coherent subset of genes repeatedly carries the signal, not when a different dominant gene appears in every reference.
Convergence is modular, not universal
Distinct phenotype-defined gene groups align with different developmental windows; they should not be collapsed into one pan-neurodevelopmental mechanism.
Conclusion and experimental handoff
We conclude that phenotype similarity can identify shared neurodevelopmental biology, but only after ontology depth, gene-set composition, and developmental time are made explicit. The useful output is a set of bounded developmental modules with named leading-edge genes and failure conditions—not a claim that all genes producing similar clinical language share one pathway.
The experimental handoff is a matched perturbation panel performed at the predicted developmental stage. Readouts should include cell-state progression, morphology, electrophysiology or network activity when appropriate, and module-specific molecular markers. Testing the same perturbation too early, too late, or in an unrelated lineage would not adequately evaluate the hypothesis.
Limitations
- Developmental atlases incompletely sample spatial, temporal, and ancestry diversity.
- Disease-gene annotations are uneven and favor well-studied disorders and genes.
- Transcriptomic convergence does not establish identical molecular direction across genes.
- In-vitro developmental systems may not reproduce human cortical timing or cell-cell interactions.
Glossary
- Information content
- A measure of term specificity based on how rarely an ontology term or its descendants occur in annotations.
- Leading edge
- The subset of genes that repeatedly contributes most strongly to a gene-set enrichment or activity score.
- Ontology propagation
- Expansion of an observed term to include defined ancestors in the ontology graph.
- Trajectory
- An inferred ordering of cellular states that approximates a biological progression rather than direct elapsed time.
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.