Predictive modelling of brain disorders with magnetic resonance imaging: A systematic review of modelling practices, transparency, and interpretability in the use of convolutional neural networks.

Brain disorders comprise several psychiatric and neurological disorders which can be characterized by impaired cognition, mood alteration, psychosis, depressive episodes, and neurodegeneration. Clinical diagnoses primarily rely on a combination of life history information and questionnaires, with a distinct lack of discriminative biomarkers in use for psychiatric disorders. Symptoms across brain conditions are associated with functional alterations of cognitive and emotional processes, which can correlate with anatomical variation; structural magnetic resonance imaging (MRI) data of the brain are therefore an important focus of research, particularly for predictive modelling. With the advent of large MRI data consortia (such as the Alzheimer's Disease Neuroimaging Initiative) facilitating a greater number of MRI-based classification studies, convolutional neural networks (CNNs)-deep learning models well suited to image processing tasks-have become increasingly popular for research into brain conditions. This has resulted in a myriad of studies reporting impressive predictive performances, demonstrating the potential clinical value of deep learning systems. However, methodologies can vary widely across studies, making them difficult to compare and/or reproduce, potentially limiting their clinical application. Here, we conduct a qualitative systematic literature review of 55 studies carrying out CNN-based predictive modelling of brain disorders using MRI data and evaluate them based on three principles-modelling practices, transparency, and interpretability. We propose several recommendations to enhance the potential for the integration of CNNs into clinical care.

Computational identification of variables in neonatal vocalizations predictive for postpubertal social behaviors in a mouse model of 16p11.2 deletion.

Autism spectrum disorder (ASD) is often signaled by atypical cries during infancy. Copy number variants (CNVs) provide genetically identifiable cases of ASD, but how early atypical cries predict a later onset of ASD among CNV carriers is not understood in humans. Genetic mouse models of CNVs have provided a reliable tool to experimentally isolate the impact of CNVs and identify early predictors for later abnormalities in behaviors relevant to ASD. However, many technical issues have confounded the phenotypic characterization of such mouse models, including systematically biased genetic backgrounds and weak or absent behavioral phenotypes. To address these issues, we developed a coisogenic mouse model of human proximal 16p11.2 hemizygous deletion and applied computational approaches to identify hidden variables within neonatal vocalizations that have predictive power for postpubertal dimensions relevant to ASD. After variables of neonatal vocalizations were selected by least absolute shrinkage and selection operator (Lasso), random forest, and Markov model, regression models were constructed to predict postpubertal dimensions relevant to ASD. While the average scores of many standard behavioral assays designed to model dimensions did not differentiate a model of 16p11.2 hemizygous deletion and wild-type littermates, specific call types and call sequences of neonatal vocalizations predicted individual variability of postpubertal reciprocal social interaction and olfactory responses to a social cue in a genotype-specific manner. Deep-phenotyping and computational analyses identified hidden variables within neonatal social communication that are predictive of postpubertal behaviors.

Altered expression of ACOX2 in non-small cell lung cancer.

Peroxisomes are organelles that play essential roles in many metabolic processes, but also play roles in innate immunity, signal transduction, aging and cancer. One of the main functions of peroxisomes is the processing of very-long chain fatty acids into metabolites that can be directed to the mitochondria. One key family of enzymes in this process are the peroxisomal acyl-CoA oxidases (ACOX1, ACOX2 and ACOX3), the expression of which has been shown to be dysregulated in some cancers. Very little is however known about the expression of this family of oxidases in non-small cell lung cancer (NSCLC). ACOX2 has however been suggested to be elevated at the mRNA level in over 10% of NSCLC, and in the present study using both standard and bioinformatics approaches we show that expression of ACOX2 is significantly altered in NSCLC. ACOX2 mRNA expression is linked to a number of mutated genes, and associations between ACOX2 expression and tumour mutational burden and immune cell infiltration were explored. Links between ACOX2 expression and candidate therapies for oncogenic driver mutations such as KRAS were also identified. Furthermore, levels of acyl-CoA oxidases and other associated peroxisomal genes were explored to identify further links between the peroxisomal pathway and NSCLC. The results of this biomarker driven study suggest that ACOX2 may have potential clinical utility in the diagnosis, prognosis and stratification of patients into various therapeutically targetable options.

Genes regulated by BCL11B during T-cell development are enriched for de novo mutations found in schizophrenia patients.

While abnormal neurodevelopment contributes to schizophrenia (SCZ) risk, there is also evidence to support a role for immune dysfunction in SCZ. BCL11B, associated with SCZ in genome-wide association study (GWAS), is a transcription factor that regulates the differentiation and development of cells in the central nervous and immune systems. Here, we use functional genomics data from studies of BCL11B to investigate the contribution of neuronal and immune processes to SCZ pathophysiology. We identified the gene targets of BCL11B in brain striatal cells (n = 223 genes), double negative 4 (DN4) developing T cells (n = 114 genes) and double positive (DP) developing T cells (n = 518 genes) using an integrated analysis of RNA-seq and ChIP-seq data. No gene-set was enriched for genes containing common variants associated with SCZ but the DP gene-set was enriched for genes containing missense de novo mutations (DNMs; p = .001) using data from 3,447 SCZ trios. Post hoc analysis revealed the enrichment to be stronger for DP genes negatively regulated by BCL11B. Biological processes enriched for genes negatively regulated by BCL11B in DP gene-set included immune system development and cytokine signaling. These analyses, leveraging a GWAS-identified SCZ risk gene and data on gene expression and transcription factor binding, indicate that DNMs in immune pathways contribute to SCZ risk.

Clinical and molecular significance of microvascular inflammation in transplant kidney biopsies.

The diagnostic criteria for antibody-mediated rejection (AMR) are continuously evolving. Here we investigated the clinical and molecular significance of different Banff microvascular inflammation (MVI) scores in transplant kidney biopsies. A total of 356 patients with clinically indicated kidney transplant biopsies were classified into three groups based on MVI scores of 0, 1, 2, or more for Groups 1-3, respectively. Gene expression profiles were assessed using arrays on a representative subset of 93 patients. The incidence of donor-specific anti-HLA antibodies was increased from 25% in Group 1 to 36% in Group 2 and to 54% in Group 3. Acute and chronic AMR were significantly more frequent in Group 3 (15% and 35%) compared with the Group 2 (3% and 15%) and Group 1 (0% and 5%), respectively. Gene expression profiles showed increased interferon-γ and rejection-induced, cytotoxic and regulatory T-cell, natural killer cell-associated and donor-specific antibody (DSA)-selective transcripts in Group 3 compared with Groups 1 and 2. There was no significant difference in gene expression profiles between the Groups 1 and 2. Increased intragraft expression of DSA-selective transcripts was found in the biopsies of C4d- Group 3 patients. Thus, an MVI score of 2 or more was significantly associated with a histological diagnosis of acute and chronic antibody-mediated rejection. Hence, increased intragraft DSA-selective gene transcripts may be used as molecular markers for AMR, especially in C4d- biopsies.

Alignment-free clustering of transcription factor binding motifs using a genetic-k-medoids approach.

BACKGROUND: Familial binding profiles (FBPs) represent the average binding specificity for a group of structurally related DNA-binding proteins. The construction of such profiles allows the classification of novel motifs based on similarity to known families, can help to reduce redundancy in motif databases and de novo prediction algorithms, and can provide valuable insights into the evolution of binding sites. Many current approaches to automated motif clustering rely on progressive tree-based techniques, and can suffer from so-called frozen sub-alignments, where motifs which are clustered early on in the process remain 'locked' in place despite the potential for better placement at a later stage. In order to avoid this scenario, we have developed a genetic-k-medoids approach which allows motifs to move freely between clusters at any point in the clustering process. RESULTS: We demonstrate the performance of our algorithm, GMACS, on multiple benchmark motif datasets, comparing results obtained with current leading approaches. The first dataset includes 355 position weight matrices from the TRANSFAC database and indicates that the k-mer frequency vector approach used in GMACS outperforms other motif comparison techniques. We then cluster a set of 79 motifs from the JASPAR database previously used in several motif clustering studies and demonstrate that GMACS can produce a higher number of structurally homogeneous clusters than other methods without the need for a large number of singletons. Finally, we show the robustness of our algorithm to noise on multiple synthetic datasets consisting of known motifs convolved with varying degrees of noise. CONCLUSIONS: Our proposed algorithm is generally applicable to any DNA or protein motifs, can produce highly stable and biologically meaningful clusters, and, by avoiding the problem of frozen sub-alignments, can provide improved results when compared with existing techniques on benchmark datasets.

Increased intragraft rejection-associated gene transcripts in patients with donor-specific antibodies and normal biopsies.

We investigated why some donor-specific antibody-positive patients do not develop antibody-mediated rejection. Of 71 donor-specific antibody-positive patients, 46 had diagnosis of antibody-mediated rejection and 25 had normal biopsies. Fifty donor-specific antibody-negative patients with normal biopsies were used as a control group. A subgroup of 61 patients with available biopsy and 64 with blood samples were analyzed by microarrays. Both donor-specific antibody-positive/antibody-mediated rejection-positive and negative biopsies showed increased expression of gene transcripts associated with cytotoxic T cells, natural killer cells, macrophages, interferon-gamma, and rejection compared to donor-specific antibody-negative biopsies. Regulatory T-cell transcripts were upregulated in donor-specific antibody-positive/antibody-mediated rejection-positive and B-cell transcripts in donor-specific antibody-positive/antibody-mediated rejection-negative biopsies. Whole-blood gene expression analysis showed increased immune activity in only donor-specific antibody-positive/antibody-mediated rejection-positive but not negative patients. During a median follow-up of 36 months, 4 donor-specific antibody-positive/antibody-mediated rejection-negative patients developed antibody-mediated rejection, 12 continued to have donor-specific antibody, but 9 lost their donor-specific antibody. Gene expression profiles did not predict the development of antibody-mediated rejection or the persistence of donor-specific antibody. Thus, donor-specific antibody-positive/antibody-mediated rejection-negative patients had increased rejection-associated gene transcripts in their allografts despite no histologic findings of rejection but not in their blood. This was found in both biopsy and blood samples of donor-specific antibody-positive/antibody-mediated rejection-positive patients.

The Wasp System: an open source environment for managing and analyzing genomic data.

The challenges associated with the management, analysis and interpretation of assays based on massively-parallel sequencing (MPS) are both individually complex and numerous. We describe what we believe to be the appropriate solution, one that represents a departure from traditional computational biology approaches. The Wasp System is an open source, distributed package written in Spring/J2EE that creates a foundation for development of an end-to-end solution for MPS-based experiments or clinical tests. Recognizing that one group will be unable to solve these challenges in isolation, we describe a nurtured open source development model that will allow the software to be collectively used, shared and developed. The ultimate goal is to emulate resources such as the Virtual Observatory of the astrophysics community, enabling computationally-inexpert scientists and clinicians to explore and interpret their MPS data. Here we describe the current implementation and development of the Wasp System and the roadmap for its community development.

Performance and Advancement of the Kidney Solid Organ Response Test.

BACKGROUND: The kidney solid organ response test (kSORT) has been investigated for the prediction of acute rejection in kidney transplant recipients with conflicting results. We aimed to investigate if the kSORT assay score is associated with rejection or immune quiescence. METHODS: The blinded association between rejection and kSORT > 9 were investigated. Optimization of kSORT prediction was evaluated after unblinding to determine the optimal prediction cutoff value of kSORT score. Additionally, the predictive capability of the kSORT gene set was assessed using blinded normalized gene expression data from microarray (Affymetrix) and qPCR assays. RESULTS: Of the 95 blood samples analyzed, 18 patients had blood samples before transplant, 77 patients after transplant and 71 had clinically indicated biopsies of which 15 biopsies showed acute rejection and 16 showed chronic active antibody-mediated rejection. When 31 patients with rejection were compared to the remaining 64 patients, positive predictive value (PPV) was 54.29% and negative predictive value (NPV) was 75% when stratified using a kSORT score > 9, and PPV was 57.89% and NPV was 78.95% when stratified using a kSORT score > 5. Using the kSORT assay for detection of rejection showed an area under the curve value of 0.71. Microarray data improved prediction accuracy with PPV of 53% and NPV of 84% compared to qPCR results (PPV and NPV were 36% and 66%), respectively. CONCLUSIONS: The kSORT assay has the potential to be used as a predictive tool for active rejection and/or immune quiescence, but additional studies will be useful in improving and refining the kSORT assay, in particular the prediction algorithm.

Loss of MMR and TGFBR2 Increases the Susceptibility to Microbiota-Dependent Inflammation-Associated Colon Cancer.

BACKGROUND AND AIMS: Mutations in DNA mismatch repair (MMR) genes are causative in Lynch syndrome and a significant proportion of sporadic colorectal cancers (CRCs). MMR-deficient (dMMR) CRCs display increased mutation rates, with mutations frequently accumulating at short repetitive DNA sequences throughout the genome (microsatellite instability). The TGFBR2 gene is one of the most frequently mutated genes in dMMR CRCs. Therefore, we generated an animal model to study how the loss of both TGFBR2 signaling impacts dMMR-driven intestinal tumorigenesis in vivo and explore the impact of the gut microbiota. METHODS: We generated VCMsh2/Tgfbr2 mice in which Msh2loxP and Tgfbr2loxP alleles are inactivated by Villin-Cre recombinase in the intestinal epithelium. VCMsh2/Tgfbr2 mice were analyzed for their rate of intestinal cancer development and for the mutational spectra and gene expression profiles of tumors. In addition, we assessed the impact of chemically induced chronic inflammation and gut microbiota composition on colorectal tumorigenesis. RESULTS: VCMsh2/Tgfbr2 mice developed small intestinal adenocarcinomas and CRCs with histopathological features highly similar to CRCs in Lynch syndrome patients. The CRCs in VCMsh2/Tgfbr2 mice were associated with the presence of colitis and displayed genetic and histological features that resembled inflammation-associated CRCs in human patients. The development of CRCs in VCMsh2/Tgfbr2 mice was strongly modulated by the gut microbiota composition, which in turn was impacted by the TGFBR2 status of the tumors. CONCLUSIONS: Our results demonstrate a synergistic interaction between MMR and TGFBR2 inactivation in inflammation-associated colon tumorigenesis and highlight the crucial impact of the gut microbiota on modulating the incidence of inflammation-associated CRCs.

Optimal targeting of BCL-family proteins in head and neck squamous cell carcinoma requires inhibition of both BCL-xL and MCL-1.

Mechanisms of treatment resistance in head and neck squamous cell carcinoma (HNSCC) are not well characterized. In this study, HNSCC tumors from a cohort of prospectively enrolled subjects on an ongoing tissue banking study were divided into those that persisted or recurred locoregionally (n=23) and those that responded without recurrence (n=35). Gene expression was evaluated using llumina HumanHT-12-v3 Expression BeadChip microarrays. Sparse Partial Least Squares - Discriminant Analysis (sPLS-DA) identified 135 genes discriminating treatment-resistant from treatment-sensitive tumors. BCL-xL was identified among 23% of canonical pathways derived from this set of genes using Ingenuity Pathway analysis. The BCL-xL protein was expressed in 8 HNSCC cell lines examined. Cells were treated with the BCL-xL inhibitor, ABT-263 (navitoclax): the average half maximal inhibitory concentration (IC50) was 8.9µM (range 6.6µM - 13.9µM). Combining ABT-263 did not significantly increase responses to 2 Gy radiation or cisplatin in the majority of cell lines. MCL-1, a potential mediator of resistance to ABT-263, was expressed in all cell lines and HNSCC patient tumors, in addition to BCL-xL. Treatment with the MCL-1 inhibitor, A-1210477, in HNSCC cell lines showed an average IC50 of 10.7µM (range, 8.8µM to 12.7µM). Adding A-1210477 to ABT-263 (navitoclax) treatment resulted in an average 7-fold reduction in the required lethal dose of ABT-263 (navitoclax) when measured across all 8 cell lines. Synergistic activity was confirmed in PCI15B, Detroit 562, MDA686LN, and HN30 based on Bliss Independence analysis. This study demonstrates that targeting both BCL-xL and MCL-1 is required to optimally inhibit BCL-family pro-survival molecules in HNSCC, and co-inhibition is synergistic in HNSCC cancer cells.

LPA receptor activity is basal specific and coincident with early pregnancy and involution during mammary gland postnatal development.

During pregnancy, luminal and basal epithelial cells of the adult mammary gland proliferate and differentiate resulting in remodeling of the adult gland. While pathways that control this process have been characterized in the gland as a whole, the contribution of specific cell subtypes, in particular the basal compartment, remains largely unknown. Basal cells provide structural and contractile support, however they also orchestrate the communication between the stroma and the luminal compartment at all developmental stages. Using RNA-seq, we show that basal cells are extraordinarily transcriptionally dynamic throughout pregnancy when compared to luminal cells. We identified gene expression changes that define specific basal functions acquired during development that led to the identification of novel markers. Enrichment analysis of gene sets from 24 mouse models for breast cancer pinpoint to a potential new function for insulin-like growth factor 1 (Igf1r) in the basal epithelium during lactogenesis. We establish that ß-catenin signaling is activated in basal cells during early pregnancy, and demonstrate that this activity is mediated by lysophosphatidic acid receptor 3 (Lpar3). These findings identify novel pathways active during functional maturation of the adult mammary gland.

Integrative Metabolic Signatures for Hepatic Radiation Injury.

BACKGROUND: Radiation-induced liver disease (RILD) is a dose-limiting factor in curative radiation therapy (RT) for liver cancers, making early detection of radiation-associated liver injury absolutely essential for medical intervention. A metabolomic approach was used to determine metabolic signatures that could serve as biomarkers for early detection of RILD in mice. METHODS: Anesthetized C57BL/6 mice received 0, 10 or 50 Gy Whole Liver Irradiation (WLI) and were contrasted to mice, which received 10 Gy whole body irradiation (WBI). Liver and plasma samples were collected at 24 hours after irradiation. The samples were processed using Gas Chromatography/Mass Spectrometry and Liquid Chromatography/Mass Spectrometry. RESULTS: Twenty four hours after WLI, 407 metabolites were detected in liver samples while 347 metabolites were detected in plasma. Plasma metabolites associated with 50 Gy WLI included several amino acids, purine and pyrimidine metabolites, microbial metabolites, and most prominently bradykinin and 3-indoxyl-sulfate. Liver metabolites associated with 50 Gy WLI included pentose phosphate, purine, and pyrimidine metabolites in liver. Plasma biomarkers in common between WLI and WBI were enriched in microbial metabolites such as 3 indoxyl sulfate, indole-3-lactic acid, phenyllactic acid, pipecolic acid, hippuric acid, and markers of DNA damage such as 2-deoxyuridine. Metabolites associated with tryptophan and indoles may reflect radiation-induced gut microbiome effects. Predominant liver biomarkers in common between WBI and WLI were amino acids, sugars, TCA metabolites (fumarate), fatty acids (lineolate, n-hexadecanoic acid) and DNA damage markers (uridine). CONCLUSIONS: We identified a set of metabolomic markers that may prove useful as plasma biomarkers of RILD and WBI. Pathway analysis also suggested that the unique metabolic changes observed after liver irradiation was an integrative response of the intestine, liver and kidney.

Clinical, Histological, and Molecular Markers Associated With Allograft Loss in Transplant Glomerulopathy Patients.

BACKGROUND: We aimed to investigate the clinical, histopathological, and molecular factors associated with allograft loss in transplant glomerulopathy (TGP) patients. METHODS: Of the 525 patients who underwent clinically indicated kidney biopsies, 52 (10%) had diagnosis of TGP. Gene expression profiles of 28 TGP and 11 normal transplant kidney biopsy samples were analyzed by Affymetrix HuGene 1.0 ST expression arrays. RESULTS: Over a median follow up of 23 months (1-46 months) after the diagnosis of TGP by biopsy, 17 patients (32%) lost their allografts at a median of 16 months (1-44 months). There was no difference between the 2 groups in terms of any demographic variables, serum creatinine, panel reactive antibody levels, donor-specific antibody frequency, or mean fluorescence intensity values. Patients who lost their allograft had a significantly higher median spot protein to creatinine ratio 2.81 (1.20-6.00) compared to no graft loss patients 1.16 (0.15-2.53), (P < 0.01), and a trends toward a higher mean chronic glomerulopathy (cg) score (1.65 ± 0.93 vs 1.11 ± 0.93) (P = 0.05). There was also no difference in microvascular inflammation or any other Banff injury scores between the 2 groups. Although 117 gene transcripts were upregulated in both groups, 86 and 57 were upregulated in graft loss and functioning allograft groups, respectively. There were significantly increased levels of intragraft endothelial cell-associated transcripts, gene transcripts associated with complement cascade, interleukins and their receptors and granulysin in graft loss patients compared to patients with a functioning allograft. CONCLUSION: Our results demonstrate differential intragraft gene expression profiles in TGP patients with allograft loss.

Allele-specific transcriptional elongation regulates monoallelic expression of the IGF2BP1 gene.

BACKGROUND: Random monoallelic expression contributes to phenotypic variation of cells and organisms. However, the epigenetic mechanisms by which individual alleles are randomly selected for expression are not known. Taking cues from chromatin signatures at imprinted gene loci such as the insulin-like growth factor 2 gene 2 (IGF2), we evaluated the contribution of CTCF, a zinc finger protein required for parent-of-origin-specific expression of the IGF2 gene, as well as a role for allele-specific association with DNA methylation, histone modification and RNA polymerase II. RESULTS: Using array-based chromatin immunoprecipitation, we identified 293 genomic loci that are associated with both CTCF and histone H3 trimethylated at lysine 9 (H3K9me3). A comparison of their genomic positions with those of previously published monoallelically expressed genes revealed no significant overlap between allele-specifically expressed genes and colocalized CTCF/H3K9me3. To analyze the contributions of CTCF and H3K9me3 to gene regulation in more detail, we focused on the monoallelically expressed IGF2BP1 gene. In vitro binding assays using the CTCF target motif at the IGF2BP1 gene, as well as allele-specific analysis of cytosine methylation and CTCF binding, revealed that CTCF does not regulate mono- or biallelic IGF2BP1 expression. Surprisingly, we found that RNA polymerase II is detected on both the maternal and paternal alleles in B lymphoblasts that express IGF2BP1 primarily from one allele. Thus, allele-specific control of RNA polymerase II elongation regulates the allelic bias of IGF2BP1 gene expression. CONCLUSIONS: Colocalization of CTCF and H3K9me3 does not represent a reliable chromatin signature indicative of monoallelic expression. Moreover, association of individual alleles with both active (H3K4me3) and silent (H3K27me3) chromatin modifications (allelic bivalent chromatin) or with RNA polymerase II also fails to identify monoallelically expressed gene loci. The selection of individual alleles for expression occurs in part during transcription elongation.

The Einstein Center for Epigenomics: studying the role of epigenomic dysregulation in human disease.

There is increasing interest in the role of epigenetic and transcriptional dysregulation in the pathogenesis of a range of human diseases, not just in the best-studied example of cancer. It is, however, quite difficult for an individual investigator to perform these studies, as they involve genome-wide molecular assays combined with sophisticated computational analytical approaches of very large datasets that may be generated from various resources and technologies. In 2008, the Albert Einstein College of Medicine in New York, USA established a Center for Epigenomics to facilitate the research programs of its investigators, providing shared resources for genome-wide assays and for data analysis. As a result, several avenues of research are now expanding, with cancer epigenomics being complemented by studies of the epigenomics of infectious disease and a neuroepigenomics program.