Genetic associations with learning over 100 days of practice

The study investigates how static genetic differences, summarized as polygenic scores (PGS) for educational attainment (EA) and cognitive performance (CP), relate to the dynamic process of learning over time. While intelligence, EA, and academic performance are heritable and sensitive to learning, it is unclear whether genetic influences on performance are magnified, reduced, or remain constant during sustained practice. Theoretical frameworks (e.g., reaction norms) suggest genetic effects can shape phenotypic responses to environmental input, but this has been rarely tested for human cognition. The authors test multiple potential patterns: magnification (rich-get-richer), compensation (catch-up), and constant effects, and whether patterns are universal or task-dependent. They examine these dynamics across nine cognitive tasks over 100 days of practice in younger and older adults using EA/CP PGS derived from large GWAS.

Prior work shows substantial heritability of intelligence, EA, and academic performance. Ackerman’s work indicates that practice effects and inter-individual variability can differ by task demands, with some motor-speed tasks showing convergence and more complex tasks showing divergence depending on performance bounds (closed vs open tasks). Genetic studies of learning are rare: Fox et al. reported increasing heritability across motor skill learning; Hambrick & Tucker-Drob and Mosing et al. found that music practice is heritable and genetically related to expert performance, suggesting genetic effects on rate of skill acquisition. Genetic PGS from large GWAS predict portions of variance in cognitive outcomes (EA PGS ~9%, CP PGS up to ~4%) but prediction varies by sample and outcome. The literature motivates testing whether and how genetic associations with performance change during learning and whether these patterns vary by cognitive domain or task characteristics.

Design and sample: 131 adults from the COGITO study (51 younger: ages 20–31; 80 older: ages 65–80) completed ~100 days of practice on nine cognitive tasks spanning episodic memory (Word List, Number-Noun, Object Position), working memory (Alpha Span, Memory Updating, N-Back), and perceptual speed (Numerical, Verbal, Figural/Spatial comparisons). Extensive pre- and post-tests bookended the practice phase. In total, 155,002 observations were collected. Tasks (except perceptual speed) had individualized difficulty during practice based on pre-test performance (presentation times, masking, interstimulus intervals), to deconfound individual differences in distance to performance ceilings and to maintain challenge across ability levels. Measures: For episodic and working memory, pre/post scores were averages of accuracy across four presentation times. For speed tasks, pre/post scores were the reciprocal of mean correct response times (higher is better). Genetic measures: EA and CP PGS were constructed from large-scale GWAS (EA N=1,131,881; CP N=257,841), excluding BASE-II and 23andMe where required. Clumping window 500 kb, pruning R²=0.25, p threshold=1.0. PGS standardized (M=0, SD=1) and adjusted for top 10 ancestry PCs; no ancestry outliers detected. Analytic approach: - Pre/post regressions: Linear regressions of PGS with pre-test, post-test, and pre–post difference for each task; standardized scores relative to pre-test; one-tailed tests; Bonferroni considered. - Latent difference score modeling (Mplus v8): Latent factors for episodic memory, working memory, and perceptual speed with measurement invariance; paths from PGS and age group to baseline latent factors and latent change; included PGS×age interactions; used full sample via ML for measurement stability. - Trajectory analyses: Correlated pre-test with performance at each practice day; assessed shapes of correlations (upward, downward, constant). - Genetic dynamics over sessions: Correlated PGS with performance at each day (residualized for age and ancestry PCs) and in 10-day blocks; also within groups sharing identical presentation times to rule out artifacts; bootstrapped SEs (1,000 resamples) for day 1 and day 100 Spearman correlations. - Growth curve modeling: Linear, logarithmic, and exponential models of individual performance trajectories over sessions using SAS NLMIXED; PGS as predictor of random-effect parameters; compared AIC across models. - Extreme-group comparisons: Compared observed trajectories between lowest 15 vs highest 15 EAPGS participants; outcomes scaled relative to day 1 means/SDs; examined parallelism or divergence; considered exponential/logarithmic fits when converged. Data preprocessing: Residualized PGS and cognitive scores (pre, post, and all waves) for continuous age and ancestry PCs to reduce confounding; results robust to age as categorical. Power considerations recognized potential low power given underperformance of PGS in-sample.

• Learning occurred with variability: All nine tasks showed positive mean pre-to-post improvements; SDs of change ~0.75 to >1 indicated substantial individual differences; 95% ranges largely positive but included some declines. - Pre/post regressions (EAPGS): Small, mostly nonsignificant effects after multiple-testing correction. Notable estimates: positive associations with N-Back Spatial pre-test (Estimate=0.159, SE=0.092) and Number-Noun pre- and post-test (Pre: 0.189, SE=0.091; Post: 0.134, SE=0.098). Negative association with Figural/Spatial Comparison pre-test response time (Estimate=-0.199, SE=0.086), suggesting slower RT (tradeoff with accuracy). EAPGS negatively related to N-Back pre–post difference (Estimate=-0.245, SE=0.112), consistent with catch-up (lower PGS improved more). - Latent difference score models: Paths from PGS to baseline latent factors or latent change were small and nonsignificant; PGS×age interactions nonsignificant. Age had consistent significant effects on baseline and latent change. - Baseline–session correlations: Three qualitative patterns in correlations between pre-test and session performance across days—upward (e.g., Alpha Span), downward (Object Position, Memory Updating, Numerical Comparison), and constant (others). - PGS–performance dynamics over 100 days (EAPGS): Three qualitative trajectories—upward, downward, constant—across tasks. Alpha Span showed upward trend from ~-0.25 to ~0.05 (initial lower EAPGS advantage that equalized). Memory Updating and N-Back trended downward toward r≈0, indicating decreasing PGS-associated differences with practice. Other tasks were roughly constant over time. - Day 1 vs Day 100 correlations: EAPGS correlations at day 1 were small (up to ~0.155). By day 100, more than half the tasks (Word List, Alpha Span, Memory Updating, Verbal Comparison, Figural/Spatial Comparison) showed increased EAPGS correlations relative to day 1; similar pattern for CPPGS. - Growth models: Substantial variability in shapes and rates of individual learning trajectories; no statistically significant associations between EAPGS and trajectory parameters. - Extreme PGS groups (top 15 vs bottom 15): Most tasks showed parallel trajectories (genetic effects roughly constant). Exceptions: Word List—high PGS group exhibited steeper gains over time; N-Back—low PGS group learned more and ultimately outperformed high PGS group. - Overall: Genetic associations with learning were modest, imprecise, and task-dependent; patterns did not consistently mirror baseline performance–learning relations.

The study directly probed how polygenic predispositions for education and cognitive performance relate to learning trajectories during 100 days of intensive cognitive practice. Findings indicate that genetic associations with performance can change during learning and differ by task, but effects are generally small and imprecise in this sample. Contrary to a universal magnification expectation, patterns included catch-up (declining PGS–performance correlations), constant effects, and occasional increases, depending on the task. Importantly, these dynamics did not simply recapitulate baseline performance associations, implying that mechanisms linking genetics to learning can differ from those linking genetics to initial ability. The robust influence of age on baseline and change underscores known age-related differences in cognitive plasticity. The results suggest that pre–post summaries can obscure when learning occurs and how genetic influences may vary across the learning process. These findings highlight the need for intensive longitudinal designs and larger samples to precisely quantify genetic contributions to learning dynamics and to test whether task characteristics (e.g., bounded vs unbounded performance) systematically moderate genetic effects.

This exploratory study provides one of the first intensive longitudinal examinations of how EA/CP polygenic scores relate to learning over 100 days across nine cognitive tasks. While participants improved on average, genetic associations with learning were small, inconsistent across tasks, and often nonsignificant, with qualitative patterns including catch-up, constant effects, and occasional increases. The work emphasizes that genetic influences on learning are dynamic and task-specific, and that pre–post differences may mask important within-course changes. Future research should deploy larger, more diverse samples; leverage stronger and more portable PGS; examine task design features as moderators; and continue using dense within-person measurement to map when and how genetic associations with performance evolve during learning.

• Limited statistical power due to small sample (N=131) for genetic association analyses; estimates imprecise with wide confidence intervals. - PGS underperformance in predicting educational attainment within this sample (R² ≈ 1.97% [0.01%, 9.75%]) reduced expected power to detect associations with learning. - Sample restricted to individuals of European ancestry in Germany, limiting generalizability. - Exploratory nature and absence of pre-registration. - Ambiguity regarding the extent to which practice gains reflect task-specific improvements versus general cognitive ability gains; transfer to other tasks or real-world functions remains uncertain. - Although individualized task difficulty was designed to deconfound ceiling proximity, interpretation of between-person differences in daily data can still be complex; however, analyses within identical presentation-time groups suggested patterns were not artifacts of individualization.