Identification and external validation of a problem cannabis risk network

Abstract

Background:

Cannabis use is common, particularly during emerging adulthood when brain development is ongoing, and its use is associated with harmful outcomes for a subset of people. An improved understanding of the neural mechanisms underlying risk for problem-level use is critical to facilitate the development of more effective prevention and treatment approaches.

Methods:

The current study applied a whole-brain, data-driven, machine-learning approach to identify neural features predictive of problem-level cannabis use in a non-clinical sample of college students (n=191, 58% female) based on reward task functional connectivity data. We further examined whether the network identified would generalize to predict cannabis use in an independent sample of European adolescents/emerging adults (n=1320, 53% female), whether it would predict clinical characteristics among adults seeking treatment for cannabis use disorder (n=33, 9% female), and whether it was specific for predicting cannabis versus alcohol use outcomes across datasets.

Results:

Results demonstrated (i) identification of a problem cannabis risk network, which (ii) generalized to predict cannabis use in an independent sample of adolescents, and (iii) linked to increased addiction severity and poorer treatment outcome in a third sample of treatment-seeking adults; further, (iv) the identified network was specific for predicting cannabis versus alcohol use outcomes across all three datasets.

Conclusions:

Findings provide insight into neural mechanisms of risk for problem-level cannabis use among adolescents/emerging adults. Future work is needed to assess whether targeting this network can improve prevention and treatment outcomes.

Clinical Trials Registration:

“Computer Based Training in Cognitive Behavioral Therapy Web-based (Man VS Machine)”, https://www.clinicaltrials.gov/ct2/show/NCT01442597, registration number: NCT01442597. “Maximizing the Efficacy of Cognitive Behavior Therapy and Contingency Management”, https://www.clinicaltrials.gov/ct2/show/NCT00350649, registration number: NCT00350649. “Computer-Based Training in Cognitive Behavior Therapy (CBT4CBT)”, https://www.clinicaltrials.gov/ct2/show/NCT01406899, registration number: NCT01406899.

Introduction

Cannabis use is extremely common (1–3). In 2022, 22% of the US population reported using cannabis in the past year (4). Use is associated with harmful outcomes for some (5, 6), including lower educational attainment (6, 7), elevated risk for psychiatric disorders (6, 8, 9), poor clinical course of co-occurring psychopathology (10), and elevated risk for cannabis (CUD; 11, 12) and other substance use disorders (12, 13). Nonetheless, many use cannabis without experiencing negative consequences, and cannabinoids have been proposed to have therapeutic potential for certain conditions (14, 15). Critically, the neural mechanisms underlying risk for the development of problem-level cannabis use (i.e., patterns of use associated with negative consequences) remain poorly understood, impeding the design of targeted prevention/intervention approaches to mitigate cannabis-related problems for persons at risk. In light of substantial increases in cannabis potency (16–18), and ongoing changes in cannabis policy (19), it is urgent to uncover mechanisms underlying cannabis-related negative outcomes and identify markers of risk.

Cannabis use peaks in late adolescence/emerging adulthood, i.e., ages 18–25 (4), and evidence suggests that youth are more vulnerable to neural effects of cannabis and cannabis-related negative outcomes (5), in part due to ongoing brain development (5, 20). Nonetheless, just as heightened neuroplasticity gives rise to increased risk for harm, it also creates a window of opportunity when individuals can maximally benefit from interventions (21). Therefore, identifying neuromarkers of risk for problem cannabis use (defined here based on frequency of use associated with meeting criteria for CUD; 22) at this point in development is critical to facilitate the development of targeted prevention and intervention approaches to foster positive developmental trajectories for individuals at risk.

Altered structure and function of brain regions implicated in executive functioning, emotions, reward processing, and memory have been identified among individuals with versus without cannabis use (23), which have been hypothesized as potential mechanisms of cannabis-related negative outcomes. However, most studies do not differentiate between recreational and problem-level use and rely on cross-sectional designs that cannot distinguish neural risk factors for use from the neurobiological sequelae of exposure (24). A small number of studies have identified prospective neural predictors of cannabis use, including volume (25, 26) and cortical thickness (25) of the orbitofrontal cortex (OFC), as well as reduced cingulate-OFC connectivity (27). Most recently, a machine-learning approach identified patterns of structure and activation of widely-distributed brain regions that predicted cannabis use initiation and demonstrated that brain-based variables improved the accuracy of predictive models relative to models including genetic and psychosocial factors alone (28), illustrating the relevance of investigating brain-based mechanisms of adolescent cannabis use.

Adolescent neurodevelopment is characterized by cascading changes in neural network connectivity (29, 30) as whole-brain connectivity patterns become increasingly stable and unique (31, 32). Therefore, whole-brain circuit-based approaches are critical to adequately capture neurodevelopmental patterns relevant to problem cannabis use (21). However, existing developmental studies of cannabis use have largely focused on individual regions and brain structure/activation, and none have taken an entirely data-driven approach, nor examined whole-brain functional connectivity (FC). Furthermore, prior studies have focused on neural correlates of cannabis use broadly (24). While informative, prediction of problem-level cannabis use is far more relevant within the context of understanding vulnerabilities for cannabis-related negative outcomes. Accordingly, recent studies have applied data-driven, whole-brain, machine-learning approaches to identify neural features associated with chronic cannabis use (33) and CUD (34) in adult samples. The current study applies one such approach, connectome-based predictive modelling (CPM), to identify neural features predictive of problem-level cannabis use in a non-clinical sample of late adolescents/emerging adults.

Prior work has demonstrated that patterns of FC are not static but change dynamically in the context of different cognitive states and/or task demands (35–39). Furthermore, FC data collected during task performance has been shown to outperform data acquired during rest for predicting behavioral outcomes (40, 41). Accordingly, different brain states may be optimal for elucidating different brain-behavior relationships (42, 43). Our prior work found that CPM analyses using data collected during reward or cognitive control tasks were differentially successful for identifying neural substrates of cocaine versus opioid abstinence (43, 44), providing evidence that different brain states are indeed better-suited for predicting different substance use behaviors. Alterations in neural reward processing have been implicated in the pathophysiology of addiction broadly (45, 46), and problem-level cannabis use specifically (47, 48). Therefore, the current study used FC data acquired during performance of a common reward processing fMRI paradigm, the monetary incentive delay (MID) task, to identify neural substrates of problem-level cannabis use.

Finally, to assess the generalizability, specificity, and clinical relevance of our results, we examined: (i) whether the neural features identified are specifically related to cannabis versus alcohol use; (ii) if they would generalize across brain states (i.e. examining whether the same features relate to problem cannabis use using data collected during a distinct cognitive task); (iii) whether the identified neural features would extend to predict cannabis use status in an entirely independent, geographically distinct sample of late adolescents/emerging adults; and (iv) if the same neural features would predict clinical characteristics in an additional independent sample of adults entering treatment for CUD.

Methods and Materials

Participants

Problem cannabis risk network identification

To identify FC markers of problem-level cannabis use, we analyzed data from the Brain and Alcohol Research in College Students (BARCS) study (49–55, 56; see Supplement for additional study information). Briefly, the BARCS study enrolled ~2,000 first-year college students (age 18–20) who completed comprehensive baseline assessments, and n~400 also participated in neuroimaging. Participants also completed monthly surveys assessing alcohol and substance use for two years. Participants with cannabis use and fMRI task data meeting quality control standards (details below) at baseline (n=191) were included in the present analyses. See Table 1 for demographics.

Table. 1.

Participant Characteristics.

	BARCS (n=191)		IMAGEN (n=1320)		Clinical (n=33)
	N	%	N	%	N	%
Sex
Female	110	57.6	695	52.7	3	9.1
Male	81	42.4	625	47.3	30	90.9
Ethnicity
Asian	9	4.7	0	0	0	0
Black	23	12	0	0	17	51.5
Hispanic	21	11	0	0	2	6.1
Other	0	0	0	0	5	15.1
White	138	72.3	1320	100	9	27.3
Substance Use
Cannabisa	53	27.7	276	20.9	32	97
Nicotineb	15	7.9	297	22.5	20	60.6
Alcoholc	118	61.8	1158	87.7	24	72.7
	M	SD	M	SD	M	SD
Age	18.37	0.583	19.05 C7	0.744	27.6	8.4

Demographic and substance use information for participants in the BARCS (n=191), IMAGEN (N=1320), and clinical replication (N=33) samples.

^aCannabis: any past-month cannabis use at baseline, based on monthly surveys of alcohol and drug use (BARCS), the ESPAD (IMAGEN), or the Addiction Severity Index (Clinical Replication sample).

^bNicotine: daily nicotine use, based on the Fagerstrom Test for Nicotine Dependence (BARCS), the ESPAD (IMAGEN), or baseline assessment (Clinical Replication sample).

^cAlcohol: any past-month alcohol use at baseline, based on monthly surveys of alcohol and drug use (BARCS), the ESPAD (IMAGEN), or the Addiction Severity Index (Clinical Replication sample).

External Replication Sample

To assess the generalizability of our findings, we also examined data from the IMAGEN study, a large-scale study of adolescent neurodevelopment including ~2,000 youth from four European countries (57; see Supplement for additional study information). Participants were enrolled at age 14 and followed into emerging adulthood. The current analyses included participants with fMRI task data that passed quality control and cannabis use data at age 19 (n=1,320; additional details below; see Table 1 for demographics).

Clinical Replication Sample

To investigate whether our findings would generalize among individuals with CUD, we used data from an independent sample of adults entering treatment for CUD (n=33; see Supplement and 58 for additional information). This sample was drawn from clinical trials of cognitive-behavioral therapy for individuals with substance use disorders. Participants were included if they had fMRI task data that met quality control standards and cannabis was their primary drug at treatment entry. See Table 1 for demographics.

All study procedures were approved by relevant ethics committees (BARCS: Central Connecticut State University, University of Connecticut, Trinity College, Hartford Hospital, and Yale University IRBs; IMAGEN: Ethics committees at each of 8 sites; Clinical sample: Yale School of Medicine IRB), and all experiments were performed in accordance with relevant guidelines and regulations.

Cannabis Use

Participants reported their monthly frequency of cannabis use over 2 years on a 6-point scale: 1=never, 2=1–2 times, 3=3–5 times, 4=6–9 times, 5=10–19 times, 6=20 or more times. To identify individuals with problem-level cannabis use, we used a frequency cutoff of 12 or more days/month, based on prior findings that this frequency of use is a strong predictor of meeting criteria for CUD in a nationally representative sample of older adolescents (22). Accordingly, participants who rated their past-month cannabis use as a 5 or 6 at any time point were classified as exhibiting problem-level use. Based on this criterion, 44 participants met criteria for problem-level use and 147 did not (the latter subgroup includes both individuals with non-use (n=96) and those who reported recreational cannabis use but no problem-level use (n=51)).

Neuroimaging Data

Monetary Incentive Delay (MID) Task

Participants in all 3 datasets completed the MID task, a fMRI paradigm used to examine neural correlates of reward processing that has frequently been administered to study neural mechanisms underlying addictive behaviors (59–64). On each trial, participants were presented with a cue indicating the potential outcome of that trial, e.g., winning money, no money at stake. They then responded to a target with a button press within a certain time window to achieve the outcome. At the end of each trial, participants received feedback indicating whether they were successful (see Supplement for details on MID task design across datasets).

Neuroimaging acquisition and preprocessing

Neuroimaging data for participants in all three datasets were collected on a 3T MRI scanner (see Supplemental Materials for acquisition parameters). Reward task data from all datasets were preprocessed with SPM8 and Bioimage Suite (43, 44, 58, 65), excluding scans with mean frame-to-frame motion greater than 0.3mm, as in prior work (43, 44, 58). The Shen atlas (66) was used to parcellate the data into 268 nodes, and pairwise correlations between the mean time course across the entire MID task for each node were conducted. Consistent with prior CPM work, the task design was not regressed out to capture patterns of FC during the brain state induced by the MID task (40). R values were then normalized with the Fisher’s r-to-z transformation and used to create 268×268 connectivity matrices summarizing connectivity strength between each node pair (see Supplement for additional preprocessing details).

Problem Cannabis Risk Network Identification

Connectivity matrices derived from BARCS data were used as the input for the CPM (67, 68) predicting problem-level cannabis use. CPM was conducted with 5-fold cross-validation using custom Python scripts (https://github.com/fye92/abcd_fy/tree/main/CPM_code), including the following steps: (1) edge selection, in which logistic regression was used to relate each edge in the matrix to problem-level cannabis use in the training dataset; (2) edge summarization, in which weights of edges identified in Step 1 are summed to create a summary value for each subject; (3) model building, in which summary scores are linearly related to the behavioral variable (i.e., y = mx+b); (4) model application, in which resultant coefficients (from Step 3) are applied to the testing dataset to generate behavioral predictions; and (5) model evaluation, in which the predictive ability of the CPM is evaluated using the correlation between predicted and observed behavioral values. To avoid overfitting to a random split of the data, the CPM was repeated 100 times. Model performance was assessed based on mean Spearman correlations between model-predicted and actual values across the 100 repetitions, and permutation testing with 1000 iterations was used to evaluate the significance of observed Spearman rho values. Finally, features present in ≥75% of CPM repetitions were included in the final problem cannabis risk network, and network anatomy was characterized based on overlap with macroscale brain regions and canonical neural networks (43, 44, 58, 65; see Supplement for features present in 100% of CPM repititions).

Specificity and Generalizability

Specificity to Cannabis versus Alcohol

Most cannabis-using BARCS participants also consumed alcohol (50). Thus, we performed sensitivity analyses to determine whether the identified problem cannabis risk network was specific to problem cannabis use or more broadly related to problem alcohol and/or substance use. Spearman correlation was used to assess the association between problem cannabis risk network strength and past 30-day alcohol use frequency. Furthermore, problem cannabis risk network strength and past 30-day alcohol use frequency were entered into a binary logistic regression analysis predicting problem cannabis use to assess whether controlling for alcohol use would diminish the relationship between problem cannabis risk network strength and problem cannabis use. Finally, we assessed whether problem cannabis risk network strength would relate to past 30-day alcohol use in a subset of participants with no recent cannabis use (n=138).

Generalizability across samples

Strength of the problem cannabis risk network (identified in BARCS data) was extracted from connectivity matrices derived from IMAGEN reward task data. Among IMAGEN participants, cannabis use was quantified using the European School Survey Project on Alcohol and Drugs (ESPAD) (69), as in prior work (e.g., 70, 71, 72). The ESPAD collects self-reported use of cannabis, alcohol, and other substances during the lifetime, past year, past month, and past week on a 7-point scale: 0: never, 1: 1–2 times, 2: 3–5 times, 3: 6–9 times, 4: 10–19 times, 5: 20–39 times, and 6: 40 or more times. Following preprocessing and quality control of the connectivity matrices, 1380 IMAGEN participants had usable age 19 reward task data. Ten participants were excluded as outliers for problem cannabis risk network strength (+/− 3 SD); of the remaining participants, 1320 had data on lifetime cannabis use collected at age 19.

Finally, strength of the problem cannabis risk network (identified in BARCS data) was extracted from connectivity matrices derived from baseline reward task data from our clinical replication sample (n=33; no outliers detected). CUD severity was assessed using the Addiction Severity Index (73) at baseline and treatment outcome was assessed based on the percentage of cannabis-negative urines provided during treatment (58).

Results

Problem Cannabis Risk Network Identification

CPM successfully identified a neuromarker of problem cannabis use among emerging adults in the BARCS study (rho=.21, p=.009; see Figure 1a; see Supplement for CPM model coefficients).

Figure 1. Problem Cannabis Risk Network.

Panel A displays model performance for positive, negative, and combined problem cannabis risk network CPM models. Panel B illustrates network anatomy based on overlap with macroscale brain regions. Edges of the positive network are depicted with red lines and edges of the negative network are depicted with blue lines; plots are thresholded to include only nodes with a minimum of 10 significant edges. Panel C illustrates network anatomy based on overlap with canonical neural networks. The positive network is shown in red, and the negative network is shown in blue; darker shading indicates that network connections account for a greater percentage of the total network.

Network Anatomy

Edges that were present in at least 75% of CPM repetitions were retained in our final problem cannabis risk network (627 total edges; 328 positive, 299 negative; see Supplement for anatomy of edges present in 100% of CPM repetitions; masks of edges present in 75% and 100% of repetitions will be made publicly available at the time of publication at https://www.nitrc.org/projects/bioimagesuite/). Examining network anatomy based on overlap with macroscale brain regions (see Figure 1b), the highest degree nodes (i.e. nodes with the greatest number of edges) of the positive network were distributed between prefrontal, motor strip, parietal, temporal and occipital regions, as well as the insula and hippocampus, whereas the highest degree nodes of the negative network were localized in prefrontal, parietal, and cerebellar regions, as well as the striatum and hippocampus. Examining network features in terms of overlap with canonical neural networks, problem cannabis use was associated with increased connectivity among motor-sensory, medial frontal, frontoparietal, default mode, and visual networks, coupled with reduced connectivity between cerebellar, medial frontal, frontoparietal, motor-sensory, and subcortical networks (see Figure 1c).

Specificity and Generalizability

Specificity to cannabis versus alcohol

There was a significant association between problem cannabis risk network strength and alcohol use in the full BARCS sample (past 30-day alcohol use frequency at baseline: rho=.26, p<.001). Binary logistic regression assessed whether controlling for alcohol use would diminish the relationship between network strength and problem cannabis use. As anticipated, the model including network strength only was highly significant, χ²(1)=163.195, p<.001; when past 30-day alcohol use was included, the model remained significant (χ²(1)=164.205, p<.001), network strength remained a highly significant predictor of problem cannabis use (p<.001), whereas alcohol did not significantly add to the model (p=.335). Finally, in a subsample with no past-month cannabis use at baseline (n=138), there was no association between problem cannabis risk network strength and past-month alcohol use (rho=−.047, p=.582; see Figure 2a; see Supplement for additional sensitivity analyses addressing alcohol use and other clinical characteristics).

Figure 2. Specificity and Generalizability of the Problem Cannabis Risk Network.

Panel A illustrates the generalizability of the problem cannabis risk network for predicting cannabis use outcomes, as well as its specificity to cannabis versus alcohol use outcomes, across the BARCS, IMAGEN, and clinical replication datasets. The effect size for the primary CPM analysis predicting problem cannabis use in the BARCS dataset is plotted for reference with a dotted line. Effect sizes for analyses examining associations between problem cannabis risk network strength and cannabis use outcomes are depicted in green; effect sizes for associations between network strength and alcohol use outcomes are depicted in grey. Asterisks indicate significant results (p<.05). *The results shown here are derived from the analysis examining the association between problem cannabis risk network strength and alcohol use in the subsample of BARCS participants with no recent cannabis use at baseline (n=138). Panel B illustrates the anatomical specificity of the problem cannabis risk network relative to a recently identified alcohol use risk network (65). Cells shaded in green represent network connections that are more characteristic of the cannabis versus the alcohol use risk network and cells shaded in blue represent network connections that are more characteristic of the alcohol versus the problem cannabis risk network.

Generalizability to IMAGEN

Relative to the BARCS sample, rates of cannabis use were low in the IMAGEN sample, with only 14% reporting ≥40 lifetime cannabis use occasions at age 19. Nonetheless, problem cannabis risk network strength was significantly associated lifetime cannabis use (rho=.08, p=.003; Figure 2a). Similarly, network strength was significantly higher among the subset of participants reporting ≥40 lifetime cannabis use occasions (i.e., the highest response option on the ESPAD; n=186) relative to those with no lifetime cannabis use history (n=652; t=2.802, p=.005). In contrast, network strength was not significantly associated with lifetime alcohol use (n=1321, rho=.05, p=.068; Figure 2a), nor was there a significant difference between subgroups reporting no lifetime use versus those with 40 or more lifetime uses of alcohol (t=−.249, p=.803).

Clinical Extension

In the clinical sample, strength of the problem cannabis risk network was significantly associated with greater baseline addiction severity (rho=.378, p=0.03; Figure 2a), and poorer treatment outcomes (i.e., fewer cannabis-negative urine toxicology specimens during treatment; rho=−.376, p=.031). Conversely, network strength was not associated with alcohol problems at treatment entry (rho=.141, p=.434; Figure 2a).

Anatomical specificity of neural features predictive of problem cannabis versus alcohol use

Given the observed specificity of the identified network for predicting cannabis versus alcohol use outcomes across datasets, we compared the anatomy of the problem cannabis risk network to an alcohol use risk network that was recently identified in the IMAGEN dataset and found to generalize to the BARCS sample (65). As illustrated in Figure 2b and discussed in the Supplement, we find distinct patterns of FC predict cannabis versus alcohol use outcomes.

Discussion

The current study applied a whole-brain, data-driven, machine-learning approach to identify a neuromarker of risk for problem-level cannabis use in a non-clinical sample of college students. The neural features identified were found to be specifically related to cannabis versus alcohol use and to generalize across brain states (latter results/discussion in Supplement). Notably, we successfully replicated this finding in a geographically distinct sample of European adolescents, and further demonstrated that the same neural features were significantly related to clinical presentation in a third independent sample of treatment-seeking adults with CUD.

While the problem cannabis risk network was identified based on problem-level cannabis use, network strength was nonetheless predictive of lifetime use in the IMAGEN dataset. Therefore, to assess whether this network is specifically relevant to problematic patterns of use, we compared network strength among individuals with problem-level use, recreational use, and non-use within the BARCS sample and found no significant difference in network strength between non-use and recreational-use groups, whereas the problem-level-use group displayed significantly higher network strength than the recreational-use and non-use groups (see Supplement), supporting the idea that this network differentiates individuals that engage in problem-level cannabis use. Nonetheless, greater persistence of cannabis use was also associated with greater network strength across both recreational and problem-level-use groups (see Supplement). Further post-hoc analyses revealed that problem cannabis risk network strength was also significantly positively associated with several facets of impulsivity and sensation-seeking, whereas no relationship was observed with clinical symptomatology relating to ADHD, depression and anxiety (see Supplement). While exploratory, this pattern of results suggests a potential bidirectional relationship whereby strength of the problem cannabis risk network may reflect certain clinical characteristics (e.g., heightened impulsivity) that predispose individuals to develop problematic patterns of cannabis use and may also be sensitive to the effects of cannabis exposure. Future work is needed to carefully examine how network strength relates to clinical presentation and to investigate how cannabis exposure and/or discontinuation may impact network strength over time.

Examining the anatomy of the neural features identified based on overlap with canonical neural networks, our results suggest that risk for problem cannabis use is associated with increased integration among motor-sensory, medial frontal, frontoparietal and default mode networks, coupled with greater segregation between cerebellar, medial frontal, and frontoparietal, as well as motor-sensory and subcortical networks (see Figure 3). The key role of motor-sensory connectivity is consistent with recent work highlighting the importance of sensorimotor processing in the pathophysiology of addiction (43, 58, 74, 75) and related features (e.g., impulsivity; 76), and aligns with the observed association between network strength and motor impulsivity (see Supplement), as well as a growing body of work highlighting sensorimotor alterations among individuals with problem cannabis use (77–81). Notably, primary motor cortex (M1) was among the highest degree motor-sensory nodes (i.e., nodes with the greatest number of connections, see Figure 4) in the positive problem cannabis risk network. Whereas the function of this region was traditionally thought to be limited to the execution of motor movements, newer evidence suggests that subregions of M1 are strongly connected to medial frontal and parietal regions, supplementary motor area, insula, thalamus, putamen and cerebellum, collectively termed the somato-cognitive action network (SCAN), and thought to contribute to integrating interoceptive and somatosensory information to guide planning and execution of goal-directed behaviors (82) (See Supplement for expanded discussion of network anatomy).

Figure 3. Network Model of Problem Cannabis Risk Network.

Figure 3 presents a network model summarizing the dominant connections of the problem cannabis risk network. Stronger connectivity (red) among motor-sensory, medial frontal, frontoparietal and default mode networks, coupled with reduced connectivity (blue) between cerebellar, medial frontal, and frontoparietal, as well as motor-sensory and subcortical networks predicted problem-level cannabis use.

Figure 4. Highest Degree Nodes of the Problem Cannabis Risk Network.

Figure 4 displays the nodes of the problem cannabis risk network that have the highest number of significant edges (i.e., degree). The degree of positive network nodes is displayed as positive; degree of negative network nodes is displayed as negative. Nodes are color coded based on their canonical network membership.

This is the first study to apply a data-driven, whole-brain, machine-learning approach to examine neural substrates of risk for problem-level cannabis use in a non-clinical sample of emerging adults. Nonetheless, the neural features identified align with extant literature, including regional case-control studies of adolescents/emerging adults with cannabis use (24), as well as more recent whole-brain analyses in adult samples with chronic use (33, 77; see Supplement for detailed discussion). Collectively, these studies illustrate the utility of applying whole-brain machine-learning techniques to elucidate complex FC signatures of cannabis use behaviors and highlight key features that appear to be relevant to problematic cannabis use across diverse samples and different points along the course of development of problem-level use (i.e., early high frequency use vs. long-term chronic use).

Critically, the features identified in our problem cannabis risk network were found to generalize across multiple heterogenous samples. Problem cannabis risk network strength was associated with lifetime cannabis use in a large sample of European adolescents in the IMAGEN study (57), despite overall low rates of cannabis use. Network strength was also associated with clinical presentation in a third sample of treatment-seeking adults with CUD, including increased addiction severity and less successful treatment outcomes. These findings supports the clinical relevance of the problem cannabis risk network and suggest that targeting this network may hold promise for improving prevention and treatment approaches. They also point toward the potential utility of using larger non-clinical neuroimaging datasets (e.g., BARCS, IMAGEN) to inform analyses of smaller, more densely-phenotyped datasets, such as our clinical sample of individuals followed throughout treatment.

Cannabis and alcohol are two of the most widely used substances worldwide, and co-use is common. Notably, our network was found to be specific for predicting cannabis versus alcohol use across all three datasets. Whereas network strength was associated with recent alcohol use frequency in the full primary BARCS sample, analyses in a subsample without recent cannabis use indicated no association between problem cannabis risk network strength and alcohol use. Moreover, although there is a higher base rate of alcohol relative to cannabis use in the IMAGEN sample, network strength was specifically related to cannabis and not alcohol use in this replication sample. Similarly, although 63.6% of our clinical treatment-seeking sample had a lifetime history of alcohol use disorder (58), network strength was significantly associated with baseline cannabis addiction severity, whereas it did not correlate with baseline alcohol problems. These findings add to a growing literature highlighting substance-specific factors implicated in the pathophysiology of different substance use disorders, even among polysubstance-using individuals (43, 58, 83).

While the current study has notable strengths, including the use of a whole-brain, data-driven, machine-learning approach incorporating stringent internal cross-validation to identify neural mechanisms of problem cannabis use in a large sample of college students with external validation in two independent samples, there are also several limitations. First, our primary sample was underpowered to examine sex-specific neuromarkers of problem cannabis use. There is a wealth of data documenting sex differences in the pathophysiology of CUD (34, 84–97) and patterns of FC more broadly (98, 99), and our prior work has identified sex-specific neural substrates of risk for alcohol use (65). Therefore, it is essential that future work in larger samples examine sex differences in the neural mechanisms of problem cannabis use. Additionally, the current study used high-frequency cannabis use as a proxy for problem-level cannabis use. While our cutoff was defined based on evidence from a nationally representative sample of older adolescents as predictive of CUD (22), it is nonetheless a somewhat indirect measure of problem-level use and future work should use measures that specifically assess CUD symptoms and/or adverse consequences of use. Additionally, future work is needed to directly assess how the neural substrates identified map onto clinical features relevant to problematic use. Similarly, it will also be important to assess whether it is possible to target the problem cannabis risk network using behavioral, pharmacological, or neuromodulatory approaches, and whether this may have therapeutic potential. Therefore, the current results pave the way for future efforts to develop more targeted and effective prevention and treatment approaches for problem cannabis use.

KEY RESOURCES TABLE.

Resource Type	Specific Reagent or Resource	Source or Reference	Identifiers	Additional Information
Add additional rows as needed for each resource type	Include species and sex when applicable.	Include name of manufacturer, company, repository, individual, or research lab. Include PMID or DOI for references; use “this paper” if new.	Include catalog numbers, stock numbers, database IDs or accession numbers, and/or RRIDs. RRIDs are highly encouraged; search for RRIDs at https://scicrunch.org/resources.	Include any additional information or notes if necessary.
Software; Algorithm	CPM code	https://github.com/fve92/abcd_fv/tree/main/CPM_code
Software; Algorithm	Matlab v9.14	Mathworks
Software; Algorithm	SPSS v29	IBM