Search

For information from sensory organs to be processed by the brain, it is usually passed to appropriate areas of the cerebral cortex. Almost all of this information passes through the thalamus, a relay structure that reciprocally connects to the vast majority of the cortex. The thalamus facilitates this information transfer through a set of thalamocortical connections that vary in cellular structure, molecular profiles, innervation patterns, and firing rates. Additionally, corticothalamic connections allow for intracortical information transfer through the thalamus. These efferent and afferent connections between the thalamus and cortex have been the focus of many studies, and the importance of cortical connectivity in defining thalamus anatomy is demonstrated by multiple studies that parcellate the thalamus based on cortical connectivity profiles.Here, we examine correlated morphological variation between the thalamus and cortex, or thalamocortical structural covariance. For each voxel in the thalamus as a seed, we construct a cortical structural covariance map that represents correlated cortical volume variation, and examine whether high structural covariance is observed in cortical areas that are functionally relevant to the seed. Then, using these cortical structural covariance maps as features, we subdivide the thalamus into six non-overlapping regions (clusters of voxels), and assess whether cortical structural covariance is associated with cortical connectivity that specifically originates from these regions.We show that cortical structural covariance is high in areas of the cortex that are functionally related to the seed voxel, cortical structural covariance varies along cortical depth, and sharp transitions in cortical structural covariance profiles are observed when varying seed locations in the thalamus. Subdividing the thalamus based on structural covariance, we additionally demonstrate that the six thalamic clusters of voxels stratify cortical structural covariance along the dorsal-ventral, medial-lateral, and anterior-posterior axes. These cluster-associated structural covariance patterns are prominently detected in cortical regions innervated by fibers projecting out of their related thalamic subdivisions.Together, these results advance our understanding of how the thalamus and the cortex couple in their volumes. Our results indicate that these volume correlations reflect functional organization and structural connectivity, and further provides a novel segmentation of the mouse thalamus that can be used to examine thalamic structural variation and thalamocortical structural covariation in disease models.

Autism spectrum disorder (ASD) is a neurodevelopmental disorder that affects about 1 in 55 children worldwide, imposing enormous economic and socioemotional burden on families and communities. Genetic studies of ASD have identified de novo copy number variants (CNVs) and point mutations that contribute significantly to the genetic architecture of ASD, but the majority of these studies were conducted in outbred populations, which are not ideal for detecting autosomal recessive (AR) inheritance. However, several studies of ASD in consanguineous populations point towards AR as an under-appreciated source of ASD variants. Here, we used trio whole exome sequencing to look for rare variants for ASD in 115 proband- mother-father trios from populations with high rates of consanguinity, namely Pakistan, Iran, and Saudi Arabia. We report 87 candidate sequence variants, with 57% biallelic, 21% autosomal dominant/de novo, and the rest X-linked. 52% of the variants were loss of function (LoF) or putative LoF (splice site, stop loss), and 47% nonsynonymous. Our analysis indicates an enrichment of biallelic variants. These include variants in genes previously reported for AR ASD and/or intellectual disability (ID) (AGA, ASL, ASPA, BTN3A2, CC2D1A, DEAF1, HTRA2, KIF16B, LINS1, MADD, MED25, MTHFR, RSRC1, TECPR2, VPS13B, ZNF335), and 32 previously unreported candidates, including 15 LoF or splice variants, in genes such as DAGLA, EFCAB8, ENPP6, FAXDC2, ILDR2, PKD1L1, SCN10A, and SLC36A1. We also identified several candidate biallelic exonic loss CNVs. The enrichment shown here for biallelic variants confirms that the genetic architecture for ASD among consanguineous populations is different to non-consanguineous populations. However, as has been shown in many recessive disorders, the genes reported here may also be relevant to outbred populations.

Cholinergic dysfunction has been implicated in the pathophysiology of psychiatric disorders such as schizophrenia, depression, and bipolar disorder. The basal forebrain (BF) cholinergic nuclei, defined as cholinergic cell groups Ch1-3 and Ch4 (Nucleus Basalis of Meynert; NBM), provide extensive cholinergic projections to the rest of the brain. Here, we examined microstructural neuroimaging measures of the cholinergic nuclei in patients with untreated psychosis (∼ 31 weeks of psychosis

Reduced brain-derived neurotrophic factor (BDNF) and gamma-aminobutyric acid (GABA) neurotransmission co-occur in brain conditions (depression, schizophrenia and age-related disorders) and are associated with symptomatology. Rodent studies show they are causally linked, suggesting the presence of biological pathways mediating this link. Here we first show that reduced BDNF and GABA also co-occur with attenuated autophagy in human depression. Using mice, we then show that reducing Bdnf levels (Bdnf(+/−)) leads to upregulated sequestosome-1/p62, a key autophagy-associated adaptor protein, whose levels are inversely correlated with autophagic activity. Reduced Bdnf levels also caused reduced surface presentation of α5 subunit-containing GABA(A) receptor (α5-GABA(A)R) in prefrontal cortex (PFC) pyramidal neurons. Reducing p62 gene dosage restored α5-GABA(A)R surface expression and rescued PFC-relevant behavioral deficits of Bdnf(+/−) mice, including cognitive inflexibility and reduced sensorimotor gating. Increasing p62 levels was sufficient to recreate the molecular and behavioral profiles of Bdnf(+/−) mice. Collectively, the data reveal a novel mechanism by which deficient BDNF leads to targeted reduced GABAergic signaling through autophagic dysregulation of p62, potentially underlying cognitive impairment across brain conditions.

Anorexia nervosa (AN) is a highly heritable psychiatric disorder characterized by starvation and emaciation and associated with changes in brain structure. The precise nature of these changes remains unclear, as does their developmental time course and capacity for reversal with weight restoration. In this exploratory neuroimaging study, we sought to characterize changes in white matter microstructure in women with acute and remitted AN. Diffusion-weighted MRI data was collected from underweight women with a current diagnosis of AN (acAN: n = 23), weight-recovered women with a past diagnosis of AN (recAN: n = 23), and age-matched healthy control women (HC: n = 24). Image processing and analysis were performed with Tract-Based Spatial Statistics, part of FSL, and group differences in voxelwise, brain-wide fractional anisotropy (FA) and mean diffusivity (MD), indices of white matter microstructure, were tested with nonparametric permutation and threshold-free cluster enhancement. No significant main effect of group on FA was identified. A significant main effect of group on MD was observed in a large cluster covering 9.2% of white matter and including substantial portions of the corpus callosum, corona radiata, internal capsule, and superior longitudinal fasciculus, and post hoc analyses revealed similar effects of group on axial diffusivity (AD) and radial diffusivity (RD). Clusterwise MD was significantly higher in acAN participants (+3.8%) and recAN participants (+2.9%) than healthy controls, and the same was true for clusterwise AD and RD. Trait-based increases in diffusivity, changes in which have been associated with atypical myelination and impaired axon integrity, suggest a link between altered white matter microstructure and vulnerability to AN, and evidence of reduced oligodendrocyte density in AN provides further support for this hypothesis. Potential mechanisms of action include atypical neurodevelopment and systemic inflammation.

Application of RNA sequencing has enabled the characterization of genome-wide gene expression in the human brain, including distinct layers of the neocortex. Neuroanatomically, the molecular patterns that underlie the laminar organization of the neocortex can help link structure to circuitry and function. To advance our understanding of cortical architecture, we created LaminaRGeneVis, a web application that displays across-layer cortical gene expression from multiple datasets. These datasets were collected using bulk, single-nucleus, and spatial RNA sequencing methodologies and these data were harmonized to facilitate comparisons between datasets. The online resource facilitates single- and multi-gene analyses by providing figures and statistics for user-friendly assessment of laminar gene expression patterns in the adult human neocortex.Availability and implementationLaminaRGeneVis is available at https://ethanhkim.shinyapps.io/laminargenevis. The source code and data is accessible at https://github.com/ethanhkim/laminargenevis.

High resolutionin situhybridization (ISH) images of the brain capture spatial gene expression at cellular resolution. These spatial profiles are key to understanding brain organization at the molecular level. Previously, manual qualitative scoring and informatics pipelines have been applied to ISH images to determine expression intensity and pattern. To better capture the complex patterns of gene expression in the human cerebral cortex, we applied a machine learning approach. We propose gene re-identification as a contrastive learning task to compute representations of ISH images. We train our model on a ISH dataset of ~1,000 genes obtained from postmortem samples from 42 individuals. This model reaches a gene re-identification rate of 38.3%, a 13x improvement over random chance. We find that the learned embeddings predict expression intensity and pattern. To test generalization, we generated embeddings in a second dataset that assayed the expression of 78 genes in 53 individuals. In this set of images, 60.2% of genes are re-identified, suggesting the model is robust. Importantly, this dataset assayed expression in individuals diagnosed with schizophrenia. Gene and donor-specific embeddings from the model predict schizophrenia diagnosis at levels similar to that reached with demographic information. Mutations in the most discriminative gene,SCN4B, may help understand cardiovascular associations with schizophrenia and its treatment. We have publicly released our source code, embeddings, and models to spur further application to spatial transcriptomics. In summary, we propose and evaluate gene re-identification as a machine learning task to represent ISH gene expression images.

Porphyromonas gingivalis, a bacterium associated with periodontal disease, is a suspected cause of Alzheimer’s disease. This bacterium is reliant on gingipain proteases, which cleave host proteins after arginine and lysine residues. To characterize gingipain susceptibility, we performed enrichment analyses of arginine and lysine proportion proteome-wide. Genes differentially expressed in brain samples with detected P. gingivalis reads were also examined. Genes from these analyses were tested for functional enrichment and specific neuroanatomical expression patterns. Proteins in the SRP-dependent cotranslational protein targeting to membrane pathway were enriched for these residues and previously associated with periodontal and Alzheimer’s disease. These ribosomal genes are up-regulated in prefrontal cortex samples with detected P. gingivalis sequences. Other differentially expressed genes have been previously associated with dementia (ITM2B, MAPT, ZNF267, and DHX37). For an anatomical perspective, we characterized the expression of the P. gingivalis associated genes in the mouse and human brain. This analysis highlighted the hypothalamus, cholinergic neurons, and the basal forebrain. Our results suggest markers of neural P. gingivalis infection and link the cholinergic and gingipain hypotheses of Alzheimer’s disease.

Porphyromonas gingivalis, a keystone species in the development of periodontal disease, is a suspected cause of Alzheimer’s disease. This bacterium is reliant on gingipain proteases, which cleave host proteins after arginine and lysine residues. To characterize gingipain susceptibility, we performed enrichment analyses of arginine and lysine proportion proteome-wide. Proteins in the SRP-dependent cotranslational protein targeting to membrane pathway were enriched for these residues and previously associated with periodontal and Alzheimer’s disease. These ribosomal genes are up-regulated in prefrontal cortex samples with detected P. gingivalis sequences. Other differentially expressed genes have been previously associated with dementia (ITM2B, MAPI, ZNF267, and DHX37). For an anatomical perspective, we characterized the expression of the P. gingivalis associated genes in the mouse and human brain. This analysis highlighted the hypothalamus, cholinergic neurons, and the basal forebrain. Our results suggest markers of neural P. gingivalis infection and link the gingipain and cholinergic hypotheses of Alzheimer’s disease.

Anorexia nervosa is a complex eating disorder with genetic, metabolic, and psychosocial underpinnings. Using unbiased genome-wide methods, recent studies have associated a variety of genes with the disorder. We characterized these genes by projecting them into aggregated gene expression data from reference transcriptomic atlases of the prenatal and adult human brain. We found that genes from an induced stem cell study of anorexia nervosa are expressed at higher levels in the lateral parabrachial and the ventral tegmental areas. The adult expression enrichment of the lateral parabrachial is confirmed with genes from two independent genetic studies. In the fetal brain, enrichment of the ventral tegmental area is also observed for the six genes near the only common variant associated with the disorder (rs4622308). We also observed signals in the adult and fetal pontine raphe, but they were not observed when using the genes from the genetic studies. In addition to signals related to calcitonin gene-related peptide neurons and the tachykinin, we found more than the expected number of microglia marker genes within the gene sets. Using mouse transcriptomic data, we identified several anorexia nervosa associated genes that are differentially expressed during food deprivation. While these genes that respond to fasting are not enriched in the gene sets, we highlightRPS26which is proximal to rs4622308. We did not observe expression enrichment in the cingulate cortex or hypothalamus suggesting other targets for deep brain stimulation should be considered for severe cases. This work improves our understanding of the neurobiological causes of anorexia nervosa by suggesting disturbances in subcortical appetitive circuits.

By engaging angiotensin-converting enzyme 2 (ACE2 or Ace2), the novel pathogenic severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) invades host cells and affects many organs, including the brain. However, the distribution of ACE2 in the brain is still obscure. Here, we investigated the ACE2 expression in the brain by analyzing data from publicly available brain transcriptome databases. According to our spatial distribution analysis, ACE2 was relatively highly expressed in some brain locations, such as the choroid plexus and paraventricular nuclei of the thalamus. According to cell-type distribution analysis, nuclear expression of ACE2 was found in many neurons (both excitatory and inhibitory neurons) and some non-neuron cells (mainly astrocytes, oligodendrocytes, and endothelial cells) in the human middle temporal gyrus and posterior cingulate cortex. A few ACE2-expressing nuclei were found in a hippocampal dataset, and none were detected in the prefrontal cortex. Except for the additional high expression of Ace2 in the olfactory bulb areas for spatial distribution as well as in the pericytes and endothelial cells for cell-type distribution, the distribution of Ace2 in the mouse brain was similar to that in the human brain. Thus, our results reveal an outline of ACE2/Ace2 distribution in the human and mouse brains, which indicates that the brain infection of SARS-CoV-2 may be capable of inducing central nervous system symptoms in coronavirus disease 2019 (COVID-19) patients. Potential species differences should be considered when using mouse models to study the neurological effects of SARS-CoV-2 infection.

Visual Abstract, Alzheimer’s disease (AD) starts decades before clinical symptoms appear. Low-glucose utilization in regions of the cerebral cortex marks early AD. To identify these regions, we conducted a voxel-wise meta-analysis of previous studies conducted with positron emission tomography that compared AD patients with healthy controls. The resulting map marks hypometabolism in the posterior cingulate, middle frontal, angular gyrus, and middle and inferior temporal regions. Using the Allen Human Brain Atlas, we identified genes that show spatial correlation across the cerebral cortex between their expression and this hypometabolism. Of the six brains in the Atlas, one demonstrated a strong spatial correlation between gene expression and hypometabolism. Previous neuropathological assessment of this brain from a 39-year-old male noted a neurofibrillary tangle in the entorhinal cortex. Using the transcriptomic data, we estimate lower proportions of neurons and more microglia in the hypometabolic regions when comparing this donor’s brain with the other five donors. Within this single brain, signal recognition particle (SRP)-dependent cotranslational protein targeting genes, which encode primarily cytosolic ribosome proteins, are highly expressed in the hypometabolic regions. Analyses of human and mouse data show that expression of these genes increases progressively across AD-associated states of microglial activation. In addition, genes involved in cell killing, chronic inflammation, ubiquitination, tRNA aminoacylation, and vacuole sorting are associated with the hypometabolism map. These genes suggest disruption of the protein life cycle and neuroimmune activation. Taken together, our molecular characterization reveals a link to AD-associated hypometabolism that may be relevant to preclinical stages of AD.

Ethnicity impacts Alzheimer’s disease risk, especially among Caribbean-Hispanics. We report the first RNA-sequencing analysis of brain tissue from 45 Alzheimer’s disease and control Caribbean-Hispanics. Data were compared with two independent samples of non-Hispanic Caucasians (total n=729). By identifying and characterizing those genes with ancestry- and region-specific expression patterns in Alzheimer’s disease, we reveal molecular insights that may help explain epidemiological disparities in this understudied aging population.

By engaging angiotensin-converting enzyme 2 (ACE2 or Ace2), the novel pathogenic SARS-coronavirus 2 (SARS-CoV-2) may invade host cells in many organs, including the brain. However, the distribution of ACE2 in the brain is still obscure. Here we investigated the ACE2 expression in the brain by analyzing data from publicly available brain transcriptome databases. According to our spatial distribution analysis, ACE2 was relatively highly expressed in some brain locations, such as the choroid plexus and paraventricular nuclei of the thalamus. According to cell-type distribution analysis, nuclear expression of ACE2 was found in many neurons (both excitatory and inhibitory neurons) and some non-neuron cells (mainly astrocytes, oligodendrocytes, and endothelial cells) in human middle temporal gyrus and posterior cingulate cortex. A few ACE2-expressing nuclei were found in a hippocampal dataset, and none were detected in the prefrontal cortex. Except for the additional high expression of Ace2 in the olfactory bulb areas for spatial distribution as well as in the pericytes and endothelial cells for cell-type distribution, the distribution of Ace2 in mouse brain was similar to that in the human brain. Thus, our results reveal an outline of ACE2/Ace2 distribution in the human and mouse brain, which indicates the brain infection of SARS-CoV-2 may be capable of inducing central nervous system symptoms in coronavirus disease 2019 (COVID-19) patients. Potential species differences should be considered when using mouse models to study the neurological effects of SARS-CoV-2 infection.

Major depressive disorder (MDD) is the most prevalent psychiatric disorder worldwide and affects individuals of all ages. It causes significant psychosocial impairments and is a major cause of disability. A recent consortium study identified 102 genetic variants and 269 genes associated with depression. To provide targets for future depression research, we prioritized these recently identified genes using expression data. We examined the differential expression of these genes in three studies that profiled gene expression of MDD cases and controls across multiple brain regions. In addition, we integrated anatomical expression information to determine which brain regions and transcriptomic cell types highly express the candidate genes. We highlight 12 of the 269 genes with the most consistent differential expression: MANEA, UBE2M, CKB, ITPR3, SPRY2, SAMD5, TMEM106B, ZC3H7B, LST1, ASXL3, ZNF184 and HSPA1A. The majority of these top genes were found to have sex-specific differential expression. We place greater emphasis on ZNF184 as it is the top gene in a more conservative analysis of the 269. Specifically, the differential expression of ZNF184 was strongest in subcortical regions in males and females. Anatomically, our results suggest the importance of the dorsal lateral geniculate nucleus, cholinergic, monoaminergic and enteric neurons. These findings provide a guide for targeted experiments to advance our understanding of the genetic underpinnings of depression.

Background Major depressive disorder (MDD) affects women approximately twice as often as men. Women are three times as likely to have atypical depression, with hypersomnia and weight gain. This suggests that the molecular mechanisms of MDD may differ by sex. Methods To test this hypothesis, we performed a large-scale gene expression meta-analysis across three corticolimbic brain regions: the dorsolateral prefrontal cortex, subgenual anterior cingulate cortex, and basolateral amygdala (26 men, 24 women with MDD and sex-matched control subjects). Results were further analyzed using a threshold-free approach, Gene Ontology, and cell type–specific analyses. A separate dataset was used for independent validation (13 MDD subjects/sex and 22 control subjects [13 men, 9 women]). Results Of the 706 genes differentially expressed in men with MDD and 882 genes differentially expressed in women with MDD, only 21 were changed in the same direction in both sexes. Notably, 52 genes displayed expression changes in opposite directions between men and women with MDD. Similar results were obtained using a threshold-free approach, in which the overall transcriptional profile of MDD was opposite in men and women. Gene Ontology indicated that men with MDD had decreases in synapse-related genes, whereas women with MDD exhibited transcriptional increases in this pathway. Cell type–specific analysis indicated that men with MDD exhibited increases in oligodendrocyte- and microglia-related genes, while women with MDD had decreases in markers of these cell types. Conclusions The brain transcriptional profile of MDD differs greatly by sex, with multiple transcriptional changes in opposite directions between men and women with MDD.

Neurobiological underpinnings of cortical thickness in the human brain are largely unknown. Here we use cell-type-specific gene markers to evaluate the contribution of 9 neural cell-types in explaining inter-regional variations in cortical thickness and age-related cortical thinning in the adolescent brain. Gene-expression data were derived from the Allen Human Brain Atlas (and validated using the BrainSpan Atlas). Values of cortical thickness/thinning were obtained with magnetic resonance imaging in a sample of 987 adolescents. We show that inter-regional profiles in cortical thickness relate to those in the expression of genes marking CA1 pyramidal cells, astrocytes, and microglia; taken together, the 3 cell types explain 70% of regional variation in cortical thickness. We also show that inter-regional profiles in cortical thinning relate to those in the expression of genes marking CA1 and S1 pyramidal cells, astrocytes and microglia. Using Gene Ontology analysis, we demonstrate that the difference in the contribution of CA1 and S1 pyramidal cells may relate to biological processes such as neuronal plasticity and potassium channel activity, respectively. This "virtual histology" approach (scripts provided) can be used to examine neurobiological underpinnings of cortical profiles associated with development, aging, and various disorders.

Exposures to life stressors accumulate across the lifespan, with possible impact on brain health. Little is known, however, about the mechanisms mediating age-related changes in brain structure. We use a lifespan sample of participants (n = 21 251; 4–97 years) to investigate the relationship between the thickness of cerebral cortex and the expression of the glucocorticoid- and the mineralocorticoid-receptor genes (NR3C1 and NR3C2, respectively), obtained from the Allen Human Brain Atlas. In all participants, cortical thickness correlated negatively with the expression of both NR3C1 and NR3C2 across 34 cortical regions. The magnitude of this correlation varied across the lifespan. From childhood through early adulthood, the profile similarity (between NR3C1/NR3C2 expression and thickness) increased with age. Conversely, both profile similarities decreased with age in late life. These variations do not reflect age-related changes in NR3C1 and NR3C2 expression, as observed in 5 databases of gene expression in the human cerebral cortex (502 donors). Based on the co-expression of NR3C1 (and NR3C2) with genes specific to neural cell types, we determine the potential involvement of microglia, astrocytes, and CA1 pyramidal cells in mediating the relationship between corticosteroid exposure and cortical thickness. Therefore, corticosteroids may influence brain structure to a variable degree throughout life.

Alzheimer’s disease (AD) starts decades before clinical symptoms appear. Low glucose utilization in regions of the cerebral cortex marks early AD and is clinically useful. To identify these regions, we conducted a voxel-wise meta-analysis of positron emission tomography studies that compared AD patients with healthy controls. This meta-analysis included 27 studies that assayed glucose utilization in 915 AD patients and 715 healthy controls. The resulting map marks hypometabolism in the posterior cingulate, middle frontal, angular gyrus, middle and inferior temporal regions. Using the Allen Human Brain Atlas, we identified genes with expression patterns associated with this hypometabolism pattern in the cerebral cortex. Of the six brains in the Atlas, one demonstrated a strong spatial association with the hypometabolism pattern. Previous neuropathological assessment of this brain from a 39-year-old male noted a neurofibrillary tangle in the entorhinal cortex. Using the transcriptomic data, we estimate lower proportions of neurons and more microglia in the hypometabolic regions when compared with the other five brains. Within this single brain, signal recognition particle (SRP)-dependent cotranslational protein targeting genes, which primarily encode cytosolic ribosome proteins, are highly expressed in the hypometabolic regions. Analyses of human and mouse data show that expression of these genes progressively increases across AD-associated states of microglial activation. In addition, genes involved in cell killing, chronic inflammation, ubiquitination, tRNA aminoacylation, and vacuole sorting are associated with the hypometabolism map. These genes suggest disruption of the protein life cycle and neuroimmune activation. Taken together, our molecular characterization of cortical hypometabolism reveals a molecular link to AD associated hypometabolism that may be relevant to preclinical stages.

BACKGROUND Mental illness affects a significant portion of the worldwide population. Online mental health forums can provide a supportive environment for those afflicted and also generate a large amount of data that can be mined to predict mental health states using machine learning methods. OBJECTIVE This study aimed to benchmark multiple methods of text feature representation for social media posts and compare their downstream use with automated machine learning (AutoML) tools. We tested on datasets that contain posts labeled for perceived suicide risk or moderator attention in the context of self-harm. Specifically, we assessed the ability of the methods to prioritize posts that a moderator would identify for immediate response. METHODS We used 1588 labeled posts from the Computational Linguistics and Clinical Psychology (CLPsych) 2017 shared task collected from the Reachout.com forum. Posts were represented using lexicon-based tools, including Valence Aware Dictionary and sEntiment Reasoner, Empath, and Linguistic Inquiry and Word Count, and also using pretrained artificial neural network models, including DeepMoji, Universal Sentence Encoder, and Generative Pretrained Transformer-1 (GPT-1). We used Tree-based Optimization Tool and Auto-Sklearn as AutoML tools to generate classifiers to triage the posts. RESULTS The top-performing system used features derived from the GPT-1 model, which was fine-tuned on over 150,000 unlabeled posts from Reachout.com. Our top system had a macroaveraged F1 score of 0.572, providing a new state-of-the-art result on the CLPsych 2017 task. This was achieved without additional information from metadata or preceding posts. Error analyses revealed that this top system often misses expressions of hopelessness. In addition, we have presented visualizations that aid in the understanding of the learned classifiers. CONCLUSIONS In this study, we found that transfer learning is an effective strategy for predicting risk with relatively little labeled data and noted that fine-tuning of pretrained language models provides further gains when large amounts of unlabeled text are available.