DynoSys: A Dynamic Systems Framework for Multimodal Integration of Genetic, Environmental, and Neurobiological Signals

Abstract: Understanding the development of adolescent behavioral and mental health outcomes requires integrating genetic predisposition, environmental exposures, and neurobiological processes over time. Here, we present a unified quantitative framework that models the human body as a dynamic system, where genetic factors form the foundational state, environmental exposures act as time-varying inputs, the brain might serve as a mediation processor, and behavioral phenotypes emerge as system outputs. Using longitudinal data from the Adolescent Brain Cognitive Development (ABCD) Study, we construct harmonized multi-domain representations across six phenotypes: externalizing behavior, internalizing behavior, and four substance use initiation outcomes (alcohol, nicotine, cannabis, and any substance use). We integrate polygenic risk scores (PRS), multi-domain environmental features, and multimodal neuroimaging representations derived through stability selection and dimensionality reduction. Our framework supports both continuous longitudinal modeling and survival-based event modeling through a unified data structure. We further develop interpretable domain-level representations using principal components, weighted risk scores, and cluster-based summaries. These representations enable downstream modeling using survival analysis, state-space models, and machine learning approaches. This work establishes a scalable and interpretable framework for studying how genetic and environmental factors interact over time to shape behavioral outcomes, providing a foundation for identifying modifiable risk factors and informing early intervention strategies.

• The paper introduces a time-aware framework that integrates polygenic risk scores, environmental exposures, neuroimaging features, and longitudinal outcomes in one analytical pipeline.
• The framework predicts continuous behavior more effectively than substance-use initiation, achieving R²=0.298 for externalizing and R²=0.143 for internalizing symptoms, with environmental features providing the strongest stable signals.
• Substance-use models show good discrimination, including AUROC=0.823 for cannabis initiation, but sparse events and low precision limit stable feature discovery, underscoring the need for larger samples and advanced survival models.

DynoSys is a systems-level analytical framework that models adolescent behavioral development as a dynamic system, integrating polygenic risk scores (PRS), multi-domain environmental exposures, and structural neuroimaging features within a single time-indexed longitudinal data structure (2605.02952). Applied to the Adolescent Brain Cognitive Development (ABCD) Study, the framework jointly supports continuous trajectory modeling (externalizing and internalizing behavior) and discrete-time survival modeling of substance-use initiation (alcohol, nicotine, cannabis, any substance). The central empirical finding is an asymmetry between outcome classes: continuous behavioral phenotypes yield both moderate predictive performance and stable interpretable features dominated by environmental representations, whereas substance-use initiation outcomes achieve good discrimination but no stable feature-level predictors under conservative stability selection.

Conceptual framework

The framework formalizes behavior as the output of a dynamic system in which genetic variation constitutes the foundational state, environmental exposures act as time-varying inputs, the brain serves as a mediation processor, and behavioral phenotypes emerge as outputs. This conceptualization aligns with the bioecological model, developmental psychopathology, and imaging genetics traditions, but distinguishes itself by enforcing explicit temporal structure: lagged predictors at time $t-1$ predict outcomes at time $t$, change scores capture within-individual dynamics, and landmark-based designs anchor survival analyses at 0, 24, and 48 months.

Two outcome classes are handled within one pipeline. Continuous outcomes are CBCL-derived T-scores for externalizing and internalizing behavior across baseline and follow-up waves. Survival outcomes are substance-use initiation events defined on an interval basis with censoring for non-events, following the neuroanatomical variability study of Miller et al. (2605.02952). The unified panel structure permits identical subject-level cross-validation and covariate handling across both outcome types.

Data construction and feature engineering

The analytic sample comprises unrelated individuals of genetically defined European ancestry (projected onto the 1000 Genomes reference via PCA and k-nearest-neighbor classification), yielding 2,621 individuals and 11,108 longitudinal observations for continuous phenotypes, and 2,620 individuals with 7,860 interval-level observations per initiation phenotype. PRS were constructed with PRS-CS using LD from external reference panels for externalizing, internalizing (MDD, anxiety, neuroticism, youth internalizing symptoms, plus a composite), and substance use domains.

Environmental, family, physical health, technology use, and neurocognition variables passed through a three-stage pipeline: quality screening with LOCF imputation for static-like variables; individual-level cross-validated univariate selection requiring both selection frequency and sign consistency; and representation learning producing PCA components, effect-size-weighted scores, and risk/protective cluster summaries per domain. Structural MRI features (cortical area, sulcal depth, thickness, ICV-normalized volumes) from FreeSurfer processing at three waves underwent tiered quality control, residualization against demographic and genetic principal components, and stability-based selection via repeated subsampling Cox LASSO. An adaptive compression rule retains raw features when stable-feature counts are small and compact representations when counts are large.

Predictive performance

The strongest continuous result is externalizing behavior: LASSO on the full feature representation achieves mean $R^2 = 0.298$, RMSE = 8.14, and $r = 0.548$. Internalizing behavior is predicted more weakly ($R^2 = 0.143$, $r = 0.381$), which the authors attribute to greater subjectivity and reporting variability of internalizing symptoms. Across both outcomes, weighted representations retain most predictive signal while remaining interpretable, whereas PCA-only and cluster-only representations degrade performance—indicating that aggressive dimensionality reduction discards outcome-relevant information.

Outcome	Type	Best representation	Best model	Performance
Externalizing	Continuous	Full	LASSO	$R^2=0.298$; RMSE=8.14; $r=0.548$
Internalizing	Continuous	Full	LASSO/elastic net	$R^2=0.143$; RMSE=9.68; $r=0.381$
Alcohol initiation	Binary	PCA	Logistic elastic net	AUROC=0.782
Nicotine initiation	Binary	Full	Logistic elastic net	AUROC=0.793
Cannabis initiation	Binary	Weighted	Logistic elastic net	AUROC=0.823; AUPRC≈0.049
Any substance initiation	Binary	PCA	Logistic elastic net	AUROC=0.781; AUPRC=0.502

The cannabis initiation result warrants emphasis as internally tension-laden: the highest AUROC (0.823) coexists with an AUPRC near 0.05, reflecting very low event prevalence. High discrimination metrics therefore do not translate into stable feature discovery.

Domain contributions and temporal dynamics

Stable domain-level signal concentrates overwhelmingly in environmental features. For externalizing behavior, the cluster representation yields eight stable environmental features with total absolute coefficient weight of 71.78, versus PRS-foundation weights of only 0.68–0.83 across representations. For internalizing behavior, 19 stable environmental features appear in the full representation (total weight 16.09), with PRS contributions of 1.26–1.59. Lagged cross-domain analysis reinforces this ordering: the weighted environmental level score carries the largest coefficient for both phenotypes (3.878 for externalizing, 2.594 for internalizing), while externalizing PRS (0.716) and MDD/youth-internalizing/anxiety PRS (0.504/0.450/0.273) show phenotype-consistent positive associations. Brain features contribute smaller negative coefficients—for example, supramarginal and middle temporal structural measures for externalizing, and left banks of the superior temporal sulcus thickness and left frontal pole sulcal depth for internalizing—suggesting candidate but unproven neural substrates.

The authors are explicit that these are temporally ordered associations, not causal effects, and that brain features do not yet support a mechanistic claim of central processing despite the conceptual framing.

Multi-task learning

A shared-bottom multi-task network with two regression heads and four classification heads was trained to test whether shared structure exists across all six phenotypes. It did not outperform single-task benchmarks: best MTL $t$0 was $t$1 for externalizing and $t$2 for internalizing, versus 0.298 and 0.143 respectively for single-task penalized regression. Representation comparisons within MTL mirror the single-task pattern—full/all-feature and weighted representations perform well, PCA and cluster representations collapse toward zero variance explained for internalizing. Among binary outcomes under MTL, cannabis initiation shows the clearest signal (AUROC ≈ 0.73, AUPRC above its 0.022 prevalence baseline, though with large split-to-split variance), nicotine is modest (AUROC ≈ 0.60), and alcohol and any-substance initiation are near chance by AUROC. The authors interpret MTL not as a superior predictor but as evidence that some outcomes share detectable multi-domain structure while others require outcome-specific modeling.

Limitations

Several constraints bound the conclusions. PRS analyses are restricted to European ancestry, limiting generalizability. The design is observational, so no causal claims follow from the environmental dominance findings. Substance initiation events are sparse, and the absence of stable initiation features may reflect low event rates, short follow-up, distributed weak effects, or an overly conservative stability threshold rather than genuine absence of predictors—the authors correctly caution against interpreting this null as evidence of unimportance. Brain features contributed modestly, leaving the "brain as mediator" hypothesis empirically underdetermined. Whether competing-risk or multi-state survival models, domain-level grouped stability selection, or longer ABCD follow-up would recover stable initiation predictors remains an open question this paper does not resolve.

Conclusion

DynoSys demonstrates that genetic liability, environmental exposure, and neurobiological signals can be integrated within a single time-aware pipeline spanning continuous trajectories and survival-style events. Its empirical support is strongest for continuous behavioral outcomes, where environmental dynamics dominate prediction, PRS provide reproducible background susceptibility, and brain structure adds modest lagged information. The framework's principal unresolved challenge is translating good discrimination of sparse initiation events into stable, interpretable risk architecture—a problem the authors attribute to power rather than biology, and which larger samples and alternative survival formulations will need to address.