Integrated Polygenic Scoring Sharpened Depression Prediction Without Losing Specificity
A new polygenic prediction strategy for depression moved beyond a single-disease score and into the broader internalizing spectrum. Investigators combined information from depression and genetically correlated psychiatric traits, using phenotype reconstruction, multi-trait genetic integration, and Bayesian effect estimation to build a unified score. In UK Biobank, that approach increased liability R2 for depression by 49.7% compared with the strongest publicly available depression-only score that excluded 23andMe data.
The gain was not confined to one dataset or one analytic frame. Within-family standardized associations with latent internalizing were up to 54% larger, a useful sign that the score retained signal even when comparisons were made among relatives rather than across unrelated individuals. In a field where population stratification and family structure can distort apparent performance, that pattern matters because it suggests the added predictive power was not merely an artifact of broader ancestry differences or environmental clustering.
Childhood and Adolescent Outcomes Tracked Across Mood, Sleep, and Social Function
The same integrated score was then tested in the ABCD study against 2,328 childhood and adolescent outcomes. Its association profile was almost identical to the depression-only score, with a correlation of r = 0.99, yet the new score was uniformly stronger. That combination is clinically important: the score expanded magnitude without changing the overall pattern of association, which argues for broader sensitivity rather than a shift toward a different phenotype.
A total of 1,494 outcomes linked to mood, behavior, social functioning, hobbies, and sleep showed association with the score, including 423 within-family associations. Those domains map onto the everyday prodrome of depressive illness and adjacent internalizing states, where sleep disruption, reduced engagement, and social withdrawal often emerge before full syndromic depression becomes obvious. By reflecting a wider internalizing liability, the score appears to capture early developmental expression of risk rather than a narrow endpoint alone.
One result stood out starkly. Adolescents in the highest score decile were 3.5 times as likely as those in the lowest decile to report suicidal thoughts by age 14, 23% versus 7%.
Risk Extended to Nicotine, Cannabis, and Adult Health Burden
The score also tracked later behavioral and somatic outcomes. By age 17, adolescents in the highest decile were 1.9 times as likely to report nicotine or cannabis use, 47% versus 25%. In adults, the score predicted worse general health, chronic pain, and life dissatisfaction. Those associations fit a broad liability model in which inherited vulnerability to mood disturbance overlaps with pain perception, substance use risk, and global self-rated health.
This is where biological plausibility becomes hard to ignore. Depression liability is not only about sadness; it often travels with altered reward processing, stress responsivity, sleep regulation, and pain sensitivity. Genetic signal that spans those domains may therefore be identifying a more fundamental vulnerability in how the brain and body process threat, reward, and recovery. The score’s breadth across development and adulthood suggests that the inherited architecture of depression is embedded early and persists across life stages in different forms.
Sex Differences Were Large, While Household Context Did Not Modify Effects
Associations with psychiatric symptoms were substantially stronger in girls than in boys, with boys’ coefficients at 0.68 times those of girls. The pattern was especially pronounced for girl-typical symptoms. That finding aligns with the well-established epidemiology of depression, which begins to diverge by sex during adolescence and often becomes more pronounced after puberty as hormonal, psychosocial, and developmental pressures change symptom expression.
The absence of moderation by household resources, hardship, or deprivation after correction for multiple testing is equally informative. Genetic liability remained visible across those social strata in this analysis, suggesting that the score’s predictive signal was not easily erased by measured family context. At the same time, social conditions still shape clinical expression, access to care, and the timing of symptom recognition, so lack of statistical moderation here should not be read as absence of environmental relevance.
Predictive Attenuation Followed Genetic Distance From Training Data
Prediction weakened in line with genetic distance from the training data for Hispanic and African American children. That pattern is not surprising in polygenic work, but it remains a central limitation for any model built largely from existing reference datasets. Scores trained in one ancestry background often transfer less efficiently to groups with different linkage disequilibrium structure, allele frequencies, and representation in discovery cohorts.
For clinical genomics, the practical implication is straightforward. A stronger score in one population does not automatically translate into equitable utility across all populations. Without broader and more representative discovery samples, performance gaps are likely to persist even when the underlying biology is shared. This analysis makes that problem visible rather than theoretical.
A Broader Internalizing Score May Be More Useful Than a Narrow Depression Model
The most consequential finding is that depression prediction improved substantially when the model incorporated the internalizing spectrum rather than depression alone, without detectable loss of specificity or within-family predictive ability. In other words, broadening the genetic input did not blur the phenotype; it sharpened it. The near-identical association profile between the integrated and depression-only scores suggests that much of the added power came from a more complete recovery of the same underlying liability structure.
For psychiatry, that raises a practical question about how depression risk should be modeled in large biobanks and eventually in clinical settings. A narrow case-control frame may miss the developmental, behavioral, and familial expression of inherited risk, while a broader internalizing model may better match the way depression unfolds in real life. The current findings remain a preprint and have not undergone peer review, so they should be interpreted as an early step rather than a ready-to-deploy clinical tool. Even so, the study leaves little doubt that depression liability can be predicted more strongly when the score is allowed to borrow signal from closely related psychiatric traits.