Transcriptomic signatures of early- versus late-diagnosed ADHD and implications for treatment heterogeneity

Abstract

Attention-deficit/hyperactivity disorder (ADHD) is heterogeneous in the age at which symptoms first lead to clinical diagnosis. Recent genomic analyses of a large Danish population-based cohort showed that individuals diagnosed in childhood differ from those diagnosed in adulthood in polygenic risk scores, rare-variant burden, and patterns of genetic correlation with other psychiatric traits [1]. Childhood cases exhibited stronger overlap with autism and higher rates of rare protein-truncating variants in constrained genes, whereas late-diagnosed cases showed greater genetic overlap with depression. The functional transcriptomic mechanisms underlying these differences have remained unclear.

We performed a transcriptome-wide association study (TWAS) using S-PrediXcan and GTEx v8 MASHR models from six adult brain tissues: amygdala, anterior cingulate cortex, caudate, frontal cortex, hippocampus, and nucleus accumbens. Analyses were based on GWAS summary statistics from the iPSYCH study for childhood ADHD (N = 14,878) and late-diagnosed ADHD (N = 6,961). Thirteen curated gene sets were examined, spanning synaptic plasticity, immune function, glial markers, and housekeeping controls.

The transcriptomic profiles of the two subgroups showed substantial overall similarity (Spearman ρ = 0.77 across 13,580 genes), with near-perfect directional concordance among nominally significant associations; because the two GWAS contrasted different case groups against the same 38,303 population controls, these similarity estimates should be interpreted as upper bounds rather than unbiased estimates of aetiological overlap. Despite this shared architecture, gene-set analyses suggested biologically interpretable divergences: late-diagnosed ADHD displayed stronger enrichment in long-term potentiation (enrichment ratio = 1.34, Mann–Whitney p = 1.42 × 10⁻⁸) and complement cascade pathways (1.22, p = 5.14 × 10⁻⁵), whereas childhood ADHD showed only nominally greater signals in microglia marker and MHC genes. Eighty-two genes differed in association strength between groups after FDR correction, although this delta-Z comparison should be read cautiously because the shared-control design violates the independence assumption of the test.

These findings provide a functional explanation for previously reported genetic and comorbidity differences by age at diagnosis and identify an 82-gene differential list for follow-up. The enrichment patterns generate cautious, hypothesis-level questions regarding potential variation in stimulant response and tolerability, but no direct treatment, adherence, or pharmacogenomic data were analysed. This represents the first TWAS to dissect ADHD subtypes by age at first diagnosis and offers a bridge between statistical genetic associations and biologically interpretable pathways relevant to clinical heterogeneity.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s44192-026-00540-2.

Introduction

ADHD across the lifespan

Attention-deficit/hyperactivity disorder (ADHD) is one of the more frequently encountered neurodevelopmental conditions, marked by ongoing difficulties with inattention, hyperactivity, and impulsivity that can disrupt daily functioning from childhood well into adulthood [2]. While diagnostic guidelines stress that symptoms should emerge early in life, people differ widely in when they actually receive a formal diagnosis. Some are picked up during the primary school years; others struggle quietly for much longer, only coming to clinical attention in adolescence or adulthood [3]. This spread in diagnostic timing has prompted a practical question: are childhood-diagnosed and late-diagnosed ADHD really the same condition, or do they differ in clinically and biologically meaningful ways?

Genetic evidence for diagnostic subgroups

Genetic research has started to shed light on this. Drawing on a large Danish population cohort from the iPSYCH study, Rajagopal et al. [1] divided ADHD cases by the age at which they were first diagnosed. Those identified in childhood (N = 14,878) shared more genetic overlap with autism spectrum disorder and carried a heavier load of rare protein-truncating variants in genes that are highly conserved across evolution. Individuals whose symptoms persisted from childhood into adulthood (N = 1,473) had the highest polygenic risk scores for ADHD overall, which fits with the notion that a stronger genetic push helps keep symptoms going over time. Late-diagnosed individuals (N = 6,961), by contrast, showed a noticeably lower genetic correlation with childhood ADHD (rg = 0.65) and stronger ties to polygenic scores for depression and alcohol use disorder. These genetic patterns lined up with what clinicians observe in practice: autism was far more common among childhood cases, whereas depression and substance-use problems were more frequent in the late-diagnosed group [1]. Taken together, the findings pointed toward genetic factors shaping not just who develops ADHD, but when symptoms are noticed and whether they stick around.

The mechanistic gap

Useful as these findings are, they leave a gap in our understanding of mechanism. Genome-wide association studies have turned up a number of risk loci for ADHD [4], and rare-variant work has flagged shared neurodevelopmental pathways with autism [5]. But statistical associations between genetic variants and a diagnosis do not, on their own, tell us how those variants get under the skin to affect brain function. Gene expression in the right tissues sits as a crucial middle layer between a person’s genotype and what we actually observe clinically. Transcriptome-wide association studies (TWAS) try to bridge this gap by combining GWAS summary statistics with data on how genetic variants influence gene expression in specific tissues, making it possible to ask whether the genetically predicted expression of particular genes tracks with disease risk [6]. When this approach is run across several brain regions and paired with carefully chosen gene sets, it can highlight coherent biological themes—synaptic plasticity, immune regulation, glial cell function, and so on—that might differ between diagnostic subgroups.

The present study

In this study, we applied a multi-gene-set TWAS approach to the summary statistics published by Rajagopal et al. [1], using GTEx v8 expression models drawn from six adult brain tissues. We looked at 13 gene sets chosen for their biological plausibility, covering housekeeping controls, monoaminergic systems, synaptic plasticity regulation, long-term potentiation, glutamatergic transmission, synapse pruning, the complement cascade, cellular senescence, telomere maintenance, and markers for adult astrocytes, oligodendrocytes, and microglia. We sourced the biological sets from established pathway databases [7–9] and from published single-cell transcriptomic atlases of the adult human brain [10, 11]. Our main expectation was that childhood ADHD would show relatively stronger signals in neurodevelopmental and microglial pathways, given its closer genetic relationship with autism and its higher burden of rare damaging variants. For late-diagnosed ADHD, we predicted that synaptic plasticity pathways—long-term potentiation in particular—and complement-related immune pathways would come through more strongly, in keeping with the greater load of internalising comorbidity in this group and the growing cognitive and emotional demands that come with navigating adult life.

Treatment relevance

Treatment heterogeneity remains clinically important because stimulant medications are widely used for ADHD and differ in tolerability and persistence across individuals [12–14]. The present study, however, did not analyse stimulant exposure, adherence, or pharmacogenomic outcomes. Any treatment-related interpretation is therefore deliberately framed as hypothesis generation rather than as evidence for subgroup-specific prescribing.

Aims

With this background in mind, the present investigation had two connected aims. First, we set out to characterise transcriptomic differences between childhood and late-diagnosed ADHD using a pathway-informed TWAS framework. Second, we considered how any differences that emerged might generate testable, explicitly speculative hypotheses about stimulant response or discontinuation for future registry-linked or pharmacogenomic studies. By bringing together statistical genetics and functional expression data within the same well-described cohort, we hoped to offer fresh insight into the biological diversity within ADHD and to provide a foundation for future work on personalised treatment strategies.

Methods

Participants and GWAS summary statistics

The analyses reported here relied on publicly available GWAS summary statistics from Rajagopal et al. [1]. The original sample was drawn from the iPSYCH2012 case-cohort, a population-based resource that includes all singletons born in Denmark between 1 May 1981 and 31 December 2008 who were still living in the country on their first birthday [15]. ADHD cases were identified through the Danish Psychiatric Central Research Register on the basis of an ICD-10 F90.0 code assigned by a psychiatrist. Rajagopal et al. [1] split these cases into three groups that did not overlap, defined by age at first diagnosis: childhood ADHD (N = 14,878; diagnosed before age 18 with no further adult diagnosis recorded by end of follow-up in 2016), persistent ADHD (N = 1,473; diagnosed both before and after age 18), and late-diagnosed ADHD (N = 6,961; first diagnosed at age 18 or later). Controls (N = 38,303) were selected at random from the same birth cohort and had no ADHD diagnosis on record. The childhood and late-diagnosed GWAS therefore used the same 38,303 controls, a design feature that improves comparability but creates statistical dependence between the two sets of summary statistics. We worked with the childhood and late-diagnosed summary statistics files, childhood.sumstats.gz and adulthood.sumstats.gz. Genotyping, quality control, imputation, and association testing procedures are described in full by Rajagopal et al. [1]. We focused on these two groups because they offered the largest and most straightforward comparison for examining how transcriptomic profiles might differ by age at first diagnosis.

Transcriptome-wide association analysis

Transcriptome-wide association studies were carried out with S-PrediXcan [16]. To estimate genetically predicted gene expression we used GTEx v8 MASHR models [17, 18] for six adult brain tissues that were available in the local model repository: the amygdala, anterior cingulate cortex (BA24), caudate basal ganglia, frontal cortex (BA9), hippocampus, and nucleus accumbens basal ganglia. We settled on these regions ahead of time because each has been linked to ADHD neurobiology and because they contained the largest numbers of well-predicted genes in GTEx v8. The six regions were prioritised for their relevance to frontostriatal, limbic, and hippocampal circuits implicated in ADHD and for adequate GTEx model availability. The cerebellum, although relatively well powered in GTEx and increasingly recognised in ADHD neurobiology, was not included in this initial analysis. Before running S-PrediXcan, we harmonised the GWAS summary statistics to the reference panel behind the prediction models: hg19 build, autosomal variants only. The analysis was run separately for each tissue and produced gene-level association statistics—Z-scores and p-values—reflecting how genetically predicted expression related to ADHD risk. Genes were kept for further analysis only if they had at least one SNP that appeared in both the GWAS and the prediction model.

Gene-set curation

We examined 13 gene sets in total; full gene lists and overlaps are provided in Supplementary Table S1. Two of these were intended as negative controls: a set of housekeeping genes (n = 182), compiled across ten categories that have been used before in psychiatric TWAS work, and a set of MHC/Histocompatibility Complex genes (n = 44) taken from HGNC nomenclature. Housekeeping genes and MHC/Histocompatibility Complex genes were included to benchmark against non-specific or technical signals rather than because they were expected to define ADHD subgroups. The other eleven sets captured biological processes we expected might differ between early- and late-diagnosed ADHD. These were: monoamines (n = 101), GOBP regulation of synaptic plasticity (n = 234), KEGG long-term potentiation (n = 70), GOBP synaptic transmission glutamatergic (n = 110), GOBP synapse pruning (n = 18), REACTOME complement cascade (n = 115), REACTOME cellular senescence (n = 197), REACTOME telomere maintenance (n = 112), and marker genes for adult astrocytes (n = 40), oligodendrocytes (n = 35), and microglia (n = 25). We sourced the biological sets from established pathway databases [7–9] and from published single-cell transcriptomic atlases of the adult human brain [10, 11].

Statistical analyses

Gene-level results from the six tissues were first brought together using Stouffer’s weighted Z-score method, yielding a single cross-tissue meta-Z and p-value for each gene [19]. For each of the 13 gene sets, we tested enrichment using a one-sided Mann–Whitney U test on absolute Stouffer meta-Z scores, comparing genes within the set against the background of genes with available TWAS models, approximately n = 13,580. This non-parametric approach evaluates whether set members show larger association magnitudes regardless of direction. We did not adjust for gene length, LD structure, or prediction-model accuracy, which is a limitation relative to competitive approaches such as MAGMA or stratified LDSC-style analyses. Small sets, including microglia markers (n = 25) and synapse pruning genes (n = 18), are particularly sensitive to individual outlying genes; results for these sets are therefore considered preliminary. Although the primary enrichment analysis focused on absolute Z-score magnitude, the sign of individual Z-scores may still offer mechanistic information about predicted up- or down-regulation and should be explored in future signed-enrichment analyses.

We also used two complementary tests: a Kolmogorov–Smirnov test on the full cumulative distribution of Z-scores, and a Fisher’s exact test comparing the proportion of nominally significant genes (p < .05) inside versus outside the set. Thirteen gene sets were tested per subgroup, giving 26 set-by-subgroup enrichment tests in total. No family-wise or FDR correction across gene sets was applied in the primary tables, so reported enrichment p-values should be treated as nominal unless otherwise stated. A simple Bonferroni threshold across these 26 tests is approximately 0.0019. The late-diagnosed long-term potentiation signal (p = 1.42 × 10⁻⁸) and complement cascade signal (p = 5.14 × 10⁻⁵) exceed this threshold, whereas the childhood microglia (p = .052) and MHC (p = .096) signals are suggestive only.

To compare childhood and late-diagnosed ADHD head to head, we took several steps. We computed Spearman and Pearson correlations of cross-tissue meta-Z scores across all genes that overlapped between the two sets of results. We also calculated a Z-difference score, delta-Z = Z_childhood minus Z_late, and tested it for significance both at a nominal level and after FDR correction. This delta-Z test was treated as an approximate comparison and assumes independence between the two Z-score vectors. Because the childhood and late-diagnosed GWAS shared the same controls, this assumption is violated; the resulting p-values may be miscalibrated, and the FDR-significant differential genes should be interpreted as a prioritised follow-up list rather than definitive subgroup-specific loci. Finally, we used rank-rank hypergeometric overlap (RRHO) to map out regions of concordant or discordant signal across the two diagnostic groups. Overlap of significant genes at nominal, FDR, and Bonferroni thresholds was evaluated with Fisher’s exact test.

Results

Overall similarity between childhood and late-diagnosed ADHD transcriptomic profiles

We began by looking at how closely the transcriptomic profiles of childhood and late-diagnosed ADHD resembled each other. Across the 13,580 genes for which usable prediction models existed in at least one of the six brain tissues, the cross-tissue Stouffer meta-Z scores correlated moderately to strongly between the two groups: Pearson r = .676, Spearman ρ = 0.773; both p < .001; Table 1. The direction of effect was nearly identical throughout: out of the 1,408 genes reaching nominal significance (p < .05) in both analyses, 1,406—or 99.9%—had associations pointing the same way. Rank-rank hypergeometric overlap analysis backed this up, with peak enrichment hitting 2.16-fold among the top 2,000 ranked genes. The degree of gene overlap also exceeded what chance alone would predict, and this held at increasingly strict thresholds. At nominal significance, 1,408 genes were shared: Fisher’s exact odds ratio = 2.78, p = 3.79 × 10⁻¹²⁴. At FDR < 0.05 the overlap was 590 genes: odds ratio = 3.55, p = 1.65 × 10⁻⁹⁷. After Bonferroni correction it was still 66 genes: odds ratio = 9.59, p = 2.16 × 10⁻³⁶. Even when we narrowed the focus to just the top 500 genes in each analysis, overlap ran five times higher than expected by chance: p = 1.59 × 10⁻³⁹.

Table 1.

Overall gene-level architecture (shared vs. differential)

Metric	Value	P-value
Pearson r (Stouffer-Z, all 13,580 genes)	+ 0.6764	0.00e + 00
Spearman ρ (Stouffer-Z)	+ 0.7725	0.00e + 00
Direction concordance (nominal sig genes)	99.9% (1,406/1,408)	0.00e + 00
RRHO peak enrichment	2.16× (top-2,000 genes)	—
Differential genes (FDR < 0.05)	82	—

Significant overlap (Fisher’s exact):

• Nominal P<0.05: 1,408 genes (OR=2.78, P=3.79e-124)

• FDR<0.05: 590 genes (OR=3.55, P=1.65e-97)

• Bonferroni: 66 genes (OR=9.59, P=2.16e-36)

• Top-500: 92 genes (5.00× enrichment, P=1.59e-39)

It is critical to note that both the childhood-diagnosed and late-diagnosed ADHD GWAS summary statistics were generated using the same 38,303 population controls from the iPSYCH cohort. This shared-control design induces positive covariance between the two TWAS Z-score vectors that is independent of shared biology. Consequently, the observed Spearman ρ = 0.77, directional concordance, and gene-overlap statistics in Table 1 should be interpreted as upper bounds on transcriptomic similarity attributable to aetiology. Under the null of no shared signal, the expected cross-trait covariance increases with control overlap, as described in analytical frameworks for bivariate GWAS [20]. The true aetiological transcriptomic correlation is therefore likely lower than reported here.

Pathway-level differences between diagnostic groups

That said, when we moved to the pathway level, a number of differences became apparent despite this overall resemblance (Table 2). Late-diagnosed ADHD stood out for noticeably stronger enrichment in the KEGG long-term potentiation gene set: enrichment ratio = 1.34×, Mann–Whitney p = 1.42 × 10⁻⁸, whereas childhood ADHD showed only modest enrichment in the same set: 1.10×, p = .035. A similar pattern appeared for the REACTOME complement cascade gene set, which was more clearly associated with late-diagnosed ADHD: 1.22×, p = 5.14 × 10⁻⁵, than with childhood ADHD: 1.09×, not significant. The picture flipped for a couple of other sets: childhood ADHD had nominally higher, but not multiple-testing-corrected, signals in adult microglia marker genes: 1.13×, p = .052, and in the HGNC histocompatibility complex set: 1.12×, p = .096, while these same sets showed no enrichment—and in some instances a slight numerical depletion—in the late-diagnosed group. Housekeeping genes, along with most of the remaining sets, including monoamines, regulation of synaptic plasticity, and glutamatergic transmission, had enrichment ratios hovering around 1.0 in both groups with no obvious difference between them. Full pathway results broken down by subgroup appear in Tables 3 and 4.

Table 2.

Pathway enrichment comparison (Early Dx / childhood vs. late Dx)

Gene Set	Early Dx Enrich. (MW-P)	Late Dx Enrich. (MW-P)	Δ (Late stronger)	Notes
KEGG_Long_Term_Potentiation	1.10 × (0.035)	1.34 × (1.4e-8)	+ 0.24×	Synaptic plasticity
REACTOME_Complement_Cascade	1.09× (NS)	1.22 × (5.1e-5)	+ 0.13×	Immune/inflammation
GOBP_Synaptic_Transmission_Glutamatergic	0.80× (NS)	1.01× (NS)	+ 0.21×	Glutamate signaling
Adult_Microglia_Markers	1.13 × (0.052)	0.73× (NS)	–0.40×	Neurodev./autism-like
HGNC_Histocompatibility_Complex	1.12 × (0.096)	0.65× (NS)	–0.47×	MHC/immune
Housekeeping / Monoamines / others	~ 1.0× (NS)	~ 1.0× (NS)	—	No difference

Bold values indicate statistical significance or near-significance driving the interpretation. NS = Not significant. P-values in Tables 2, 3 and 4are nominal unless otherwise stated. Across 26 gene-set tests, a simple Bonferroni threshold is approximately 0.0019; the late-diagnosed long-term potentiation and complement cascade signals exceed this threshold, whereas the childhood microglia and MHC signals are suggestive only

Table 3.

Early Dx (childhood) — multi-gene-set TWAS summary

Gene Set	Tested	Nominal sig	FDR sig	Enrich Ratio	MW P
Adult_Microglia_Markers	19	5	0	1.13×	5.19e-02
HGNC_Histocompatibility_Complex	15	2	0	1.12×	9.61e-02
GOBP_Synapse_Pruning	14	1	0	1.11×	6.36e-02
KEGG_Long_Term_Potentiation	47	7	0	1.10×	3.47e-02
REACTOME_Complement_Cascade	29	4	0	1.09×	3.22e-01
Housekeeping	123	24	0	1.00×	8.63e-01
Monoamines	60	8	0	0.99×	5.96e-01
GOBP_Regulation_of_Synaptic_Plasticity	137	18	0	1.01×	2.42e-01
GOBP_Synaptic_Transmission_Glutamatergic	66	7	0	0.80×	1.00e + 00

Table 4.

Late Dx — multi-gene-set TWAS summary

Gene Set	Tested	Nominal sig	FDR sig	Enrich Ratio	MW P
KEGG_Long_Term_Potentiation	47	7	0	1.34×	1.42e-8
REACTOME_Complement_Cascade	29	4	0	1.22×	5.14e-5
REACTOME_Telomere_Maintenance	46	7	0	1.06×	8.54e-02
GOBP_Synapse_Pruning	14	2	0	1.05×	3.45e-01
Adult_Oligodendrocyte_Markers	21	2	0	1.04×	1.05e-01
Adult_Astrocyte_Markers	27	4	0	1.03×	1.04e-01
Monoamines	60	11	0	1.02×	2.25e-01
GOBP_Regulation_of_Synaptic_Plasticity	137	21	0	1.01×	1.95e-01

Gene-level differential associations

At the individual gene level, 82 genes showed a statistically significant difference in association strength between the two diagnostic groups after FDR correction on the delta-Z test (Table 5). The single largest difference belonged to LRRCC1: delta-Z = + 10.81, FDR p = 2.95 × 10⁻¹⁰, which was more strongly tied to late-diagnosed ADHD by absolute association magnitude. Other genes with stronger late-diagnosed signals included MRPL3, RSPH4A, and CA13. On the other side, several genes were more robustly associated with childhood ADHD—most notably CTNNB1: delta-Z = + 8.31, along with FEZ2, STIP1, and TRPV2. The functions of these differentially associated genes spanned a range of biological processes, among them cytoskeletal organisation, mitochondrial function, and Wnt signalling. Because the delta-Z test did not model covariance induced by shared controls, these 82 genes should be treated as a prioritised follow-up list rather than definitive evidence of subgroup-specific causal genes.

Table 5.

Top 20 differential genes (early vs. late, FDR-ranked)

Gene	ΔZ	Diff FDR	Early Z	Late Z	Stronger in
LRRCC1	+ 10.805	2.95e-10	-0.686	-11.491	Late
MRPL3	-9.891	1.81e-08	+ 0.827	+ 10.719	Late
CTNNB1	+ 8.306	1.94e-05	+ 8.622	+ 0.316	Early
RSPH4A	-8.019	4.87e-05	+ 0.306	-7.713	Late
COQ9	+ 7.936	5.46e-05	+ 6.177	-1.758	Early
CA13	-7.912	5.01e-05	+ 0.354	+ 8.267	Late
FEZ2	-7.305	4.67e-04	-8.282	-0.977	Early
STIP1	+ 7.183	6.46e-04	+ 7.120	-0.062	Early
TRPV2	+ 7.063	8.02e-04	-7.134	-0.070	Early
TNR	-6.677	0.0020	-0.433	-7.110	Late
LATS1	+ 6.716	0.0020	+ 7.099	+ 0.383	Early
LIN7B	+ 6.658	0.0020	+ 7.068	+ 0.410	Early
RANBP6	-6.626	0.0035	-0.474	-7.385	Late
LRRK2	-3.229	0.046	+ 2.516	+ 3.524	Late
ARPC5L	-3.220	0.046	+ 0.410	+ 6.570	Late

ΔZ = Early Z − Late Z

Consistency across brain regions

The overall pattern of similarity between childhood and late-diagnosed ADHD held steady across all six brain regions we examined. Spearman correlations of the cross-tissue meta-Z scores ranged from 0.756 in the hippocampus to 0.765 in the amygdala; all p < .001; Table 6. No single tissue appeared to be driving the pathway-level differences on its own.

Table 6.

Tissue-level spearman correlations (early vs. late stouffer-Z)

Tissue	Spearman ρ	P-value
Brain_Amygdala	+ 0.7650	0.00e + 00
Brain_Anterior_cingulate_cortex_BA24	+ 0.7601	0.00e + 00
Brain_Caudate_basal_ganglia	+ 0.7584	0.00e + 00
Brain_Frontal_Cortex_BA9	+ 0.7623	0.00e + 00
Brain_Hippocampus	+ 0.7562	0.00e + 00
Brain_Nucleus_accumbens_basal_ganglia	+ 0.7623	0.00e + 00

A shared transcriptomic core

Finally, despite the differences between subgroups, a sizeable core of genes ranked highly in both analyses. One hundred and fifty-six genes fell within the top 5% of associations in both the childhood and late-diagnosed TWAS. Among the more prominent were CDC20 and PPP1CB, both of which are connected to synaptic plasticity and long-term potentiation pathways, alongside several ribosomal and mitochondrial genes (Table 7). These shared signals suggest that a set of transcriptomic changes is common to ADHD risk regardless of when a person first receives their diagnosis.

Table 7.

Top shared genes (Top 5% in BOTH analyses)

Gene	Early Z	Late Z	Pathways (if any)
CDC20	+ 7.818	+ 4.358	GOBP_Regulation_of_Synaptic_Plasticity
PPP1CB	-5.720	-4.558	KEGG_Long_Term_Potentiation
PROS1	-6.946	—	REACTOME_Complement_Cascade
TNR	-0.433	-7.110	GOBP_Regulation_of_Synaptic_Plasticity
RPL5	—	-7.793	Housekeeping
LRRK2	+ 2.516	+ 3.524	GOBP_Synaptic_Transmission_Glutamatergic
ST3GAL3	+ 9.857	+ 5.110	—
PPP1CC	—	+ 6.570	KEGG_Long_Term_Potentiation
RAB8A	-5.313	—	GOBP_Regulation_of_Synaptic_Plasticity

Partial list (first 20 of 156 genes)

Figure 1 shows the three-state alluvial diagram (Up | Null | Down) for n = 50 prioritised genes drawn from the top shared candidates and the top differentially associated genes. Full pathway enrichment statistics, tissue-level |Z| matrices, top genes, and cross-cohort comparisons are provided in Appendix A.

Fig. 1.

Three-state alluvial diagram (Up | Null | Down) for n = 50 prioritised genes drawn from the top shared candidates (n = 30; |Z| ≥ 6.8 in both cohorts; concordant ribbons) and the top differentially associated genes (n = 20; FDR-significant Z-difference). Ribbon thickness is proportional to gene count. Up denotes Stouffer Z ≥ + 2 (positive direction), Down denotes Z ≤ − 2, and Null denotes |Z| < 2 in that cohort. The two largest ribbons (Up → Up = 16; Down → Down = 14) recapitulate the high cross-cohort concordance observed across the full dataset (99.9% directional agreement among genes nominally significant in both runs; Spearman ρ = +0.773, p < .001). Smaller cohort-specific ribbons (Up → Null = 9; Down → Null = 2; Null → Up = 3; Null → Down = 5) summarise genes whose effect collapses to the null in the other cohort, while the single Down → Up reversal (RP5-1025A1.3) illustrates the rare class of sign-discordant genes (total FDR-differential = 82). Counts: Early Up = 25, Null = 8, Down = 17; Late Up = 20, Null = 11, Down = 19

Discussion

Transcriptomic differences in context

The TWAS analysis reported here puts a functional face on the genetic differences between childhood and late-diagnosed ADHD that Rajagopal et al. [1] first described. At the broad level, the transcriptomic profiles of the two groups correlated moderately to strongly with each other, which is not surprising given that they share a clinical diagnosis, although the magnitude of this correlation is likely inflated by the shared-control design and should not be equated with an unbiased aetiological transcriptomic correlation. Beneath that overall resemblance, pathway-level differences were still apparent. Late-diagnosed ADHD was marked by stronger enrichment in long-term potentiation and complement cascade pathways. Childhood ADHD showed nominally stronger signals in adult microglia marker genes and MHC genes. Although the microglia and MHC results do not survive multiple-testing correction, and MHC was initially included as a negative control to assess non-specific immune or MHC-related artefacts, the pattern is consistent with the higher autism overlap and rare-variant burden reported by Rajagopal et al. [1]. We therefore interpret these childhood-enriched signals as suggestive rather than confirmatory (Fig. 2).

Fig. 2.

Two-pillar mechanistic schematic distilling the discussion. From a common GWAS source, transcriptome-wide association testing reveals two biologically interpretable axes. The childhood-diagnosed branch (blue) is characterised by nominal microglial- and MHC-related enrichment, consistent with developmental synaptic-pruning processes, the elevated rare-variant burden in constrained genes, and high genetic overlap with autism. The late-diagnosed branch (red) is defined by Bonferroni-significant enrichment in long-term potentiation and complement cascade pathways, indicating an adult plasticity- and neuroimmune-axis signature, paired with elevated depression polygenic loading and a reduced hyperactivity genetic correlation. A shared core of 156 genes (centre) recapitulates the moderate-to-strong overall transcriptomic correlation (Spearman ρ = 0.77), interpreted as an upper bound owing to the shared-control design.. Any connection between these transcriptomic signatures and stimulant response, tolerability, or discontinuation remains hypothetical because no direct treatment data were analysed in the present study

Long-term potentiation, complement, and subgroup biology

The stronger long-term potentiation signal in late-diagnosed ADHD points toward synaptic plasticity as one possible biological distinction between diagnostic subgroups. Long-term potentiation in prefrontal, striatal, and hippocampal circuits is closely tied to learning and cognitive control, both of which are central to ADHD across the lifespan. The present finding does not show that late-diagnosed ADHD is more treatment-responsive, but it does suggest that plasticity-related regulatory architecture may be more prominent in this group than in childhood-diagnosed cases.

The complement cascade finding in late-diagnosed ADHD raises a different possibility. Complement-related genes are involved in immune signalling, synaptic refinement, and neuroinflammatory processes. In the context of the original iPSYCH findings, where late-diagnosed ADHD showed stronger genetic overlap with depression and alcohol use disorder [1], this enrichment may reflect a transcriptomic profile partly shared with internalising or stress-related liability. This interpretation remains cautious, as TWAS association does not establish causal direction and complement enrichment was not linked here to any direct biomarker or clinical treatment outcome.

Childhood ADHD and neurodevelopmental pathways

For childhood ADHD, the relatively stronger microglial and MHC signals fit within the broader neurodevelopmental picture. These pathways have come up repeatedly in autism spectrum disorder research [5, 21], and the much higher rate of real-world autism comorbidity in early-diagnosed cases—23% compared to 6% in those diagnosed late—suggests that inflammatory or pruning-related processes may play a bigger part when symptoms show up early in life [1]. Because the childhood microglia and MHC signals were only nominal, they should not be read as evidence of a confirmed immune subtype. Rather, they provide a biologically plausible lead that matches the autism and rare-variant patterns already reported in the same cohort.

Gene-level and pathway-level findings

The 82 differential genes span diverse biological processes. LRRCC1 and related cytoskeletal genes may reflect differences in cellular organisation, CTNNB1 points toward Wnt signalling, and MRPL3 suggests a possible mitochondrial component. Some of these genes may act as specific effectors within the enriched long-term potentiation or complement pathways, while others may represent independent signals. A dedicated pathway analysis of the 82-gene differential list would help reconcile the single-gene and gene-set levels of evidence.

Strengths and limitations

The study has a number of things going for it. It was carried out within the same well-described Danish population-based cohort that produced the original genetic findings, which means the childhood and late-diagnosed groups are directly comparable. The multi-gene-set approach allowed us to interpret results in terms of coherent biological themes rather than isolated single-gene hits, and the cross-TWAS comparison framework—combining correlation, delta-Z testing, and rank-rank overlap—gave us a quantitative way to pin down both what the two groups share and where they diverge. All analyses used standard, publicly available tools, which should make replication straightforward.

A fundamental limitation is the exclusive use of adult post-mortem GTEx v8 expression models for both subgroups, including the childhood-diagnosed arm. Given the neurodevelopmental nature of early-onset ADHD, adult regulatory architecture may not fully capture the gene-expression patterns relevant to childhood disease processes. This limitation could contribute to the observed transcriptomic similarity between groups and tempers interpretation of all age-at-diagnosis differences. We therefore view the current results as a necessary first step and strongly advocate replication using developmental resources such as BrainSpan, PsychENCODE developmental datasets, fetal single-cell atlases, or induced-pluripotent-stem-cell models timed to early neurodevelopment.

Two related statistical limitations arise from the shared-control design. First, the primary correlation and overlap metrics are inflated by covariance that is not necessarily biological. Second, the delta-Z test assumes independence between the childhood and late-diagnosed Z-scores. Because the two GWAS share the same controls, the variance of delta-Z was not directly estimated and the test may be miscalibrated. The 82 FDR-significant genes in Table 5 should therefore be interpreted with appropriate caution. Future analyses could model the full bivariate covariance structure or apply permutation-based approaches that preserve the shared-control design.

The gene-set analyses also require conservative reading. Enrichment p-values were reported as nominal values, and no family-wise or FDR correction was applied across the 26 set-by-subgroup tests in the primary tables. The late-diagnosed long-term potentiation and complement signals survive a simple Bonferroni threshold, but the childhood microglia and MHC findings do not. Small gene sets, especially microglia markers and synapse pruning genes, may be driven by a small number of influential genes. In addition, the enrichment approach did not adjust for gene length, LD structure, or prediction accuracy. These results should therefore be treated as pathway-level prioritisation rather than definitive competitive pathway testing.

S-PrediXcan identifies genes whose genetically predicted expression is associated with ADHD risk, but it cannot distinguish causal eQTL effects from LD-driven co-association. The highlighted differential genes, including LRRCC1 and CTNNB1, should therefore be regarded as provisional pending colocalisation analyses using approaches such as COLOC, ENLOC, or FOCUS.

Analyses were not stratified by sex. Childhood ADHD is diagnosed more frequently in males, whereas late-diagnosed cases may show different sex distributions or symptom profiles. Sex differences could therefore confound some between-group comparisons and should be addressed in future sex-stratified or sex-interaction TWAS analyses.

We also had no direct access to data on stimulant treatment or treatment response within these subgroups, so everything we say about medication implications is indirect and grounded in the existing literature rather than in the data at hand. The sample is of European ancestry, and we cannot say how well these findings would hold in other populations. The persistent ADHD group was also too small for a separate TWAS analysis here; bringing it into future work would allow a fuller picture of the developmental trajectory from childhood through to adult persistence.

Hypothesis-generating implications for stimulant treatment heterogeneity

Although no direct treatment, adherence, or pharmacogenomic data were analysed in this study, the divergent transcriptomic profiles raise testable hypotheses about treatment heterogeneity. Stronger long-term potentiation enrichment in late-diagnosed ADHD may relate to preserved plasticity and potentially more robust acute stimulant response, while elevated complement-cascade signal, together with higher depression PGS reported in the original iPSYCH analysis, could contribute to greater mood instability or inflammation-related tolerability issues [1, 13, 14, 22, 23]. However, it is currently unknown whether discontinuation rates or stimulant efficacy differ systematically between childhood- versus late-diagnosed ADHD subgroups in the epidemiological literature. The 39% one-year discontinuation rate cited from register studies is a population average and may not apply differentially to these subgroups [13]. These ideas remain speculative and are presented solely to motivate future studies that integrate TWAS signatures with longitudinal treatment records, for example from iPSYCH or national registries.

Conclusion and future directions

To sum up, this TWAS analysis shows that childhood and late-diagnosed ADHD share a large core transcriptomic signature while parting ways in biologically meaningful respects. Late-diagnosed ADHD is characterised by stronger enrichment in long-term potentiation and complement-related pathways; childhood ADHD, by relatively greater but nominal microglial and MHC signals. These differences offer a candidate functional account of the genetic and comorbidity patterns that Rajagopal et al. [1] reported and produce hypothesis-level questions about whether the two groups might follow different treatment trajectories with stimulants.

Replication in independent cohorts will matter—ideally with larger samples of persistent and late-diagnosed cases than were available here. Colocalisation analyses tying back to the original GWAS loci, followed by functional validation in iPSC-derived neurons or animal models of long-term potentiation and complement activation, would go a long way toward strengthening causal claims. Looking further ahead, combining TWAS-derived expression scores with immune biomarkers and clinical predictors could help move the field closer to genuinely personalised approaches—ones that reduce discontinuation rates and improve long-term outcomes for the growing number of adults who receive an ADHD diagnosis later in life. Such clinical applications, however, require direct testing and should not be inferred from the present transcriptomic results alone. There is clearly much still to be done, but the findings presented here illustrate how transcriptomic data can serve as a useful bridge between statistical genetics and the clinical heterogeneity that characterises this complex, lifespan disorder.

Authors’ contributions

N.C. is the sole author of this manuscript. N.C. conceptualized the study, designed the analytical framework, acquired and harmonized the GWAS summary statistics, performed all transcriptome-wide association analyses (S-PrediXcan, cross-tissue meta-analysis), conducted the gene-set enrichment tests, delta-Z comparisons, correlation analyses and rank-rank hypergeometric overlap, interpreted the results in their biological and clinical context, drafted the original manuscript, prepared all tables and the conceptual figure, and critically reviewed and approved the final version for submission.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Data availability

The datasets analysed during the current study are available from the iPSYCH Downloads repository. Summary statistics for the childhood ADHD and late-diagnosed ADHD genome-wide association studies were obtained from Rajagopal et al. Nature Genetics 2022. The iPSYCH Downloads repository lists the ADHD subtype summary statistics, readme files, and checksums for the childhood, persistent, and late-diagnosed ADHD analyses. Analysis code, including the S-PrediXcan pipeline, gene-set definitions, Stouffer meta-Z calculation, Mann–Whitney enrichment scripts, and delta-Z testing scripts, will be deposited in a public GitHub repository or Zenodo upon acceptance.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.