Paul Pavlidis

The physical substrate of memory in the brain is still a subject of debate. Itis thought that strongly connected networks of neurons, called engrams, canreproduce the original activity pattern from partial cues. Synapses betweenengram neurons are thus the most likely candidate for memory storage;but questions remain over their long-term stability. Some groups have proposedactivity related transcription, while normally considered transient,could be the beginning of a transition to a more permanent cell type thatstores long term memory. Persistent chromatin conformation changes andresulting transcription changes, triggered by reactivation, could be a stablelong-lasting storage mechanism which enables remembering at remote timepoints. Some groups have reported transcription predicted by this model.We trained classifiers using single-cell RNA (scRNA-seq) sequencing fromthe earlier transient signature (Jaeger et al., 2018; Lacar et al., 2016) andthe long term engram neuron signature (Chen et al., 2020). The transientearly signature was readily identifiable but when trying to identify engramcells exhibiting the long-term memory signature we found a significant decreasein classifier performance. The important features of the classifier werenot genes reported as deferentially expressed in the original publication. Reproducing the original author’s results using their data proved challenging,suggesting the persistent long-term memory signature was not detected ordoes not exist. Reactivation did not appear to elicit a strong transcriptionalresponse either, which contradicts models of transcriptomic engramcell formation. Unfortunately, the design of the original experiment does notallow for the falsification of the supposed persistent transcriptional programinduced by reactivation. My research suggests future directions to take inevaluating transcription’s contribution to synaptic plasticity and memory.

View record

Establishing the molecular diversity of cell types is crucial for the study of the nervous system. I compiled a cross-laboratory database of mouse brain cell type-specific transcriptomes from 36 major cell types from across the mammalian brain using rigorously curated published data from pooled cell type microarray and single-cell RNA-sequencing (RNA-seq) studies. I used these data to identify cell type-specific marker genes, discovering a substantial number of novel markers, many of which we validated using computational and experimental approaches. By examining datasets with known cell type proportion differences, I further demonstrate that summarized expression of marker gene sets (MGSs) in bulk tissue data can be used to estimate the relative cell type abundance across samples. Using this approach, I show that majority of genes previously reported as differentially expressed in Parkinson’s disease can be attributed to the reduction in dopaminergic cell number rather than regulatory events. To facilitate use of this expanding resource, I provide a user-friendly web interface at www.neuroexpresso.org.

View record

Coexpression analysis has been widely used for gene function prediction, based on the principleof guilt by association. Most studies use transcriptomic data obtained from bulk tissues, wherethe expression level of genes reflects the contribution of multiple cell types. Previous work hasalready documented how variability of cellular composition impacts coexpression analysis.However, the connection between the predictability of gene functions, coexpression networksand cell type profiles has not been studied. I hypothesized that one reason bulk-data-derivedcoexpression networks contain signals relevant to function prediction is that it containsinformation about genes’ expression profiles across cell types. Focusing on human braindatasets, I applied several approaches to test this hypothesis, including creating simulated bulkdatasets from single-nucleus data and bulk data deconvolution. I find that much predictive powercan be attributed to cell type proportion variation. Consequently, a more explicit andinterpretable function prediction can be made directly using expression patterns across cell types,which not only yields similar results but also clearly reveals the association between thefunctional terms and specific cell types. These findings have important implications forcoexpression analysis and function prediction.

View record

A persistent challenge in genetics and genomics is the interpretation of “hit lists” of genes, leading to the development of, and almost universal application of methods such as Gene Ontology (GO) enrichment analysis. While these methods have been of unquestionable utility, GO enrichment and similar approaches based on gene annotations leave much to be desired and they are often used as a “sanity check” rather than a way to make discoveries. To offer a complementary perspective with the potential to remedy some existing challenges, I developed and evaluated an algorithm that helps put hit lists of genes into biological context by performing large-scale mining on patterns of differential expression (DE). In this work, I present the development and evaluation of my algorithm which mines over 10,000 transcriptomic datasets in a process we term “condition enrichment”. The output of the algorithm is a list of biological condition comparisons (drug treatments, diseases, etc.) scored according to their relatedness (in terms of DE) to the query genes. I show that performing searches on gene sets of a priori interest enables my algorithm to effectively identify known gene-condition relationships in real and simulated data, providing a useful summary of the condition comparisons most highly associated with the differential expression of the gene set. Finally, I present a powerful open-source web application to provide researchers access to Gemma DE, in the hope that it will aid future research.

View record

Quantitative analysis of large single-cell measures acquired by phospho flow cytometry typically involves establishing inclusion gate thresholds and combining measures from accepted cells into a single median metric. Though this analysis method is simple, it overlooks the heterogeneity of cell populations and there could be information missing from the single-cell level. Here, we have formulated approaches that can recognize the heterogeneity and extract additional information involving dose-response and interactions between multiple molecules from phospho flow cytometry datasets. Using phospho flow multiplexed sampling of cell physical features, and primary antibodies against protein markers, including GAPDH as a protein expression control, HA tag as an exogenous gene/variant transfection measurement, and 8 antibodies detecting the activation (phosphorylation) states of 8 proteins within conserved molecular pathways, two panels of phospho-specific antibodies were used simultaneously for multiplexed measures in the same cells. Our approach involves single-cell standardization, fitting loess regression, identifying linear domains in dose-response plots, building linear mixed-effects models, and multi-dimensional analyses to detect interactions between phosphorylated protein markers. We demonstrate the utility of this approach by expressing wild-type and 5 variants (4A, D268E, Y138L, P38H, G129E) of PTEN on 8 markers of molecular pathways downstream of PTEN, and we also expressed RHEB WT testing its impact on markers in the shared associated pathways. We succeeded in differentiating subtypes of PTEN loss-of-function variants and were able to predict that PTEN P38H is a loss-of-lipid-phosphatase-function variant. We were also able to infer that pAKT, p4EBP1, pS6, and pCREB are all downstream targets of PTEN regulation while pAKT is between PTEN and p4EBP1, pS6, or pCREB. In conclusion, our results demonstrate dose response and molecular pathway interactions unavailable from reducing population data to single values, and our approach manifests strong promise in variant function measurement and molecular signaling pathway inference.

View record

Autism spectrum disorder (ASD) is a neurodevelopmental disorder characterized by impairments in social interaction and communication, and restrictive repetitive behaviours or interests, with extreme phenotypic and genetic heterogeneity. Currently, genetic association studies have identified 90 risk genes with high confidence out of an estimated 1000. Researchers have begun to use machine learning methods leveraging heterogeneous biological network data in attempts to aid in discovery of ASD risk genes. However, the real-world utility of these studies is questionable: network-based machine learners are often biased towards well studied genes because they operate on a principle called “guilty by association.” In this thesis, I evaluate and compare genetic and computation approaches to ASD risk gene prioritization. I demonstrate that network-based computational approaches are adding little additional useful information compared to genetic approaches for prioritization. Furthermore, I demonstrate that gene expression profiles, and generic measures of disease gene likelihood may provide less biased contextual information that can be used to supplement genetic association data to prioritize ASD risk genes. Lastly, I discuss how data quality and data dependence impacts evaluation of machine learning algorithms and genetic association studies.

View record

An accurate phylogeny of a cancer tumour has the potential to shed light on numerous phenomena, such as key oncogenetic events, relationships between clones, and evolutionary responses to treatment. Most work in cancer phylogenetics to-date relies on bulk tissue data, which can resolve only a few genotypes unambiguously. Meanwhile, single-cell technologies have considerably improved our ability to resolve intra-tumour heterogeneity. Furthermore, most cancer phylogenetic methods use classical approaches, such as Neighbor-Joining, which put all extant species on the leaves of the phylogenetic tree. But in cancer, ancestral genotypes may be present in extant populations. There is a need for scalable methods that can capture this phenomenon.We have made progress on this front by developing the Genotype Tree representation of cancer phylogenies, implementing three methods for reconstructing Genotype Trees from binary single-nucleotide variant profiles, and evaluating these methods under a variety of conditions. Additionally, we have developed a tool that simulates the evolution of cancer cell populations, allowing us to systematically vary evolutionary conditions and observe the effects on tree properties and reconstruction accuracy.Of the methods we tested, Recursive Grouping and Chow-Liu Grouping appear to be well-suited to the task of learning phylogenies over hundreds to thousands of cancer genotypes. Of the two, Recursive Grouping has the strongest and most stable overall performance, while Chow-Liu Grouping has a superior asymptotic runtime that is competitive with Neighbor-Joining.

View record

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder affecting roughly 1% of the human population. Genomics research to date has discovered only a fraction of the variants causative for ASD. To this end, we whole-genome sequenced a cohort of 119 ASD individuals in order to find likely pathogenic variation. After quality and frequency filters, we prioritized variants as likely causal according to rarity and predicted damage scores (CADD and Snap2). Here, we report five de novo damaging variants and seven likely damaging variants of unknown inheritance. Since much of the variation reported in ASD cases is uncertain both in function and in significance in ASD, we aimed to functionally characterize missense variants from the ASD literature in PTEN and SYNGAP1, two well-characterized ASD genes. We curated missense variants of unknown significance from the ASD literature and assayed their functional effect in yeast using a Synthetic Genetic Array. We chose previously biochemically validated variants, population variants, and other variants in the genes of interest to gain insight into the functional diversity of PTEN and SYNGAP1 variation. We established functional effect of the ASD variants of unknown significance in PTEN and showed that computational predictors of damage are reasonable predictors of variants’ functional effects in yeast. We found that agreement of computational metrics breaks down when predicting damage in certain genes, such as SYNGAP1. Functionalizing variants in this way contributes to our understanding of the range of functional effects of ASD variants.

View record

Expression patterns across tissues are a primary indicator of gene function. High-throughput technology created many cross-tissue data sets on a transcriptomic level (tissue panel data sets). However, the existence of multiple tissue panel data sets creates a challenge for the scientific community to decide if these data sets are equally valid or decide which data set to choose. To date, the multiple tissue panel data sets have not been well compared, nor fully evaluated. In my Master’s thesis, I collected a large number of public-available tissue panel data sets, harmonized them, integrated the data sets into a tissue expression atlas including human data and mouse data, compared and contrasted the data sets across the atlas, evaluated each data set preliminarily with a gene-specific disagreement index that I developed. I found in general, these data sets had a good agreement. However, in certain data sets the amount of disagreement was high, which indicated the qualities of these data sets were suspect.Applying the disagreement index, I was able to offer a summarized expression pattern in the tissue expression atlas with either consensus or disagreements outlined. I also developed a web-based prototype to access to this atlas.Furthermore, I explored the range of changes in gene expression patterns that may be caused by experimental conditions, such as diseases or drug treatments. I found most of the changes could not be as dramatic as a change from unexpressed to highly expressed, even though these changes were reported as statistically significant in literatures. Only a couple of conditions such as cancer or inflammation could cause an unexpressed-to-highly-expressed change, because tissue composition in those conditions were changed substantially.

View record

Recently, there has been a major effort by neuroscientists to systematically organize and integrate vast quantities of brain data. However, electrophysiological properties have been shown to be sensitive to experimental conditions, thus directly comparing them between experiments could lead to inconsistent results. Here, I characterize the general effects of experimental solution composition differences on the reported ephys measurements. For that purpose, I employ text-mining, supplemented with manual curation to gather experimental solution information from published neurophysiological articles. I integrate the extracted information into the existing NeuroElectro database, which contains the electrophysiology, neuron type and experimental conditions information (temperature, electrode type, animal age, etc.) from the above neuroscientific literature. Exploring commonly used experimental solution recipes, I found the effect of solution compositions of explaining variance in electrophysiological properties to be small, relative to the amount of the existing ephys variability. Then, I created models for predicting the variability of ephys properties commonly reported by neurophysiologists, using the available experimental conditions information. These models can be used to remove a portion of the ephys variance when comparing results from different experiments, generally making such comparisons more reliable. To validate their performance, I adjusted a portion of NeuroElectro data to experimental conditions used by Allen Institute for Brain Science and compared the respective ephys properties before and after the adjustment.

View record

There is intense interest in understanding the molecular mechanisms that contribute to neurodegenerative disorders (NDs), which involve complex interplays of genetic and environmental factors. To catch early events involved in disease initiation requires investigation on pre-symptomatic brain samples. It is difficult to capture early molecular events using post-mortem human brain samples since these samples represent the late phase of the disorder with progressive brain damage and neurodegeneration. Disease mouse models are developed to study disease progression and pathophysiology. Here, I focus on two of the most studied NDs: Alzheimer’s disease (AD) and Huntington’s disease (HD). Mouse models developed for the disease (AD or HD) often share similar phenotypes mimicking human disease symptoms, which suggest potential common underlying mechanisms of disease initiation and progression across mouse models of the same disease. Investigation of gene expression profiles of pre-symptomatic animals from different mouse models may shed light on the mechanisms occurred in the early disease phase. Gene expression profiling analyses have been performed on mouse models and some of the studies investigate the molecular changes in pre-symptomatic phase of AD and HD respectively. However, their findings have not reached a clear consensus. To identify shared molecular changes across mouse models, I conducted a systematic meta-analysis of gene expression in mouse models of AD and HD, consisted of 369 gene expression profiles from 23 independent studies. The goal of this project is to identify transcriptional alterations shared among different mouse models of each disease respectively, especially changes during early disease phase that may link to disease-causing mechanisms, and potential common cross-disease changes. For both of the disorders, the results showed subtle but biologically interpretable changes shared across mouse models in the early disease phase that may contribute to the early disease progression: dysregulation of genes involved in cholesterol biosynthesis and complement system in AD mouse models and genes encoding mitochondrial respiratory chain complexes in HD mouse models. Cross-disease similarities in the late phase suggested that different brain regions may share mechanisms in response to neuronal loss and toxic protein aggregates.

View record

An accurate phylogeny of a cancer tumour has the potential to shed light on numerous phenomena, such as key oncogenetic events, relationships between clones, and evolutionary responses to treatment. Most work in cancer phylogenetics to-date relies on bulk tissue data, which can resolve only a few genotypes unambiguously. Meanwhile, single-cell technologies have considerably improved our ability to resolve intra-tumour heterogeneity. Furthermore, most cancer phylogenetic methods use classical approaches, such as Neighbor-Joining, which put all extant species on the leaves of the phylogenetic tree. But in cancer, ancestral genotypes may be present in extant populations. There is a need for scalable methods that can capture this phenomenon.We have made progress on this front by developing the Genotype Tree representation of cancer phylogenies, implementing three methods for reconstructing Genotype Trees from binary single-nucleotide variant profiles, and evaluating these methods under a variety of conditions. Additionally, we have developed a tool that simulates the evolution of cancer cell populations, allowing us to systematically vary evolutionary conditions and observe the effects on tree properties and reconstruction accuracy.Of the methods we tested, Recursive Grouping and Chow-Liu Grouping appear to be well-suited to the task of learning phylogenies over hundreds to thousands of cancer genotypes. Of the two, Recursive Grouping has the strongest and most stable overall performance, while Chow-Liu Grouping has a superior asymptotic runtime that is competitive with Neighbor-Joining.

View record

Proteins are macromolecules responsible for a wide range of activities in the structure and function of cells. Their activities have been described in different contexts as a mean to elucidate their ``function". These descriptions have been captured across biological databases in a standardized format called Gene Ontology Annotations (GOA), to disseminate the knowledge and extrapolate the information to other proteins whose function is still unknown. Furthermore, the annotations are used to analyse and interpret data from high-throughput studies and also as a benchmark for the assessment of protein function prediction algorithms. Constant changes occur in GOA that can potentially impact such usages, but only limited effort has been put into exploring their instability, or to assess the impact that these changes have on reproducibility or interpretation of previous analyses. In the present work, I performed the most comprehensive analysis of the annotation instability for 14 representative model organisms (E.coli, fruit fly, Mouse, etc.). The results showed important instability patterns that were species-specific. As such information would be of use to the community to trace the instability of annotations of their interest, a web-based visualization tool was built to track these changes on a protein, functional term and species specific basis. Additionally, we identified artifacts on the annotation data that can be attributed to curation patterns. We propose such artifacts to be considered for a more accurate assessment of function prediction algorithms. Furthermore, the impact that changes in the annotations have on common settings like gene set enrichment analyses was also explored. In particular, 2,000 datasets were used to assess the robustness of enrichment results over time. On average, the results would display a 60% similarity after only 2 years. However, cases were found were the similarity will drop 80% within the same year, demonstrating the impact that the instability has on such applications. In conclusion, the results of this work will prove useful for those who use the annotations to interpret their studies to assess their reliability on a case-by-case scenario.

View record

DNA methylation is thought to play an important role in the regulation of mammalian gene expression. Part of the evidence for this role is the observation that lack of CpG island methylation in gene promoters is associated with high transcriptional activity. However, CpG island methylation level only accounts for a fraction of the variance in gene expression, and methylation in other domains is hypothesized to play a role (e.g., island shores and shelves). We set out to improve understanding of the human methylome through a meta-analysis approach, using 1737 samples from 30 publicly available studies. An initial screen identified 15224 CpGs that are “ultra-stable” in their state, being always fully methylated or unmethylated across diverse tissues, cell types and developmental stages (974 always methylated; 14250 always unmethylated). A further analysis of ultra-stable CpGs led us to identify a novel class of CpG islands, “ravines”, that exhibit a markedly consistent pattern of low methylation with highly methylated flanking shores and shelves. Our findings were validated using independent and heterogeneous datasets assayed on the same and different technologies. Building on additional existing data types such as gene expression microarrays, DNase hypersensitive sites, and histone modifications, we found that ravines are associated with higher gene expression, compared to typical unmethylated CpG islands. This finding suggests a novel role for methylation in promoters, markedly different from the traditional view that active promoters need to be unmethylated. We propose ravines are a new class of CpG islands, established early in development and maintained through differentiation, that mark universally active genes and provide new evidence that methylation beyond the CpG island could play a role in gene expression.

View record

The first chapter of this thesis explored the dominant gene expression pattern in the adult human brain. We discovered that the largest source of variation can be explained by cell type marker expression. Across brain regions, expression of neuron cell type markers are anti-correlated with the expression of oligodendrocyte cell type markers. Next, we explored gene function convergence and divergence in the adult mouse brain. Our contributions are as follows. First, we provide candidate cell type markers for investigating specific cell type populations. Second, we highlight orthologous genes that show functional divergence between human and mouse brains.In the second chapter, we present our preliminary work on the effects of tissue types and experimental conditions on human microarray studies. First, we measured the expression and differential expression levels of tissue-enriched genes. Next, we identified modules with similar expression levels and differential expression p-values. Our results show that expression levels reflect tissue type variation. In contrast, differential expression levels are more complex, owing to the large diversity of experimental conditions in the data. In summary, our work provides a different perspective on the functional roles of genes in human microarray studies.

View record

Declines in Pacific salmon stocks in recent decades have spurred much research into their physiology and survivorship, but comparatively little into their genomics. Sockeye salmon in particular are experiencing high levels of mortality during their migration upriver, and the numbers of returning sockeye have fluxuated wildly with respect to predictions in recent years. The goal of my project is to gain insight into the basic genomics of Pacific salmon stocks, including the sockeye, through bioinformatic approaches to gene expression profiling. Using microarray technology, I have conducted a large-scale analysis of over 1,000 samples from multiple tissues, stocks, and species of salmon. I identified tissue-specific and housekeeping genes and compared them to orthologs in mouse and human, respectively. I have also classified a number of microarray samples with a support vector machine (SVM) using qPCR data showing the presence of several common pathogens affecting Pacific salmon populations. Using identified housekeeping genes as normalizing factors, I modeled in silico a qPCR assay designed to identify salmon as infected or uninfected with a particular pathogen. With these data I hope to increase basic knowledge of the genomics of the Pacific salmon.

View record

Autism spectrum disorders (ASD) are clinically heterogeneous and biologically complex.State of the art genetics research has unveiled a large number of variants linked to ASD. Butin general it remains unclear, what biological factors lead to changes in the brains of autisticindividuals. We build on the premise that these heterogeneous genetic or genomic aberrationswill converge towards a common impact downstream, which might be reflected in thetranscriptomes of individuals with ASD. Similarly, a considerable number of transcriptomeanalyses have been performed in attempts to address this question, but their findings lack aclear consensus. As a result, each of these individual studies has not led to any significantadvance in understanding the autistic phenotype as a whole. The goal of this research is tocomprehensively re-evaluate these expression profiling studies by conducting a systematicmeta-analysis. Here, we report a meta-analysis of over 1000 microarrays across twelveindependent studies on expression changes in ASD compared to unaffected individuals,in blood and brain. We identified a number of genes that are consistently differentiallyexpressed across studies of the brain, suggestive of effects on mitochondrial function. Inblood, consistent changes were more difficult to identify, despite individual studies tendingto exhibit larger effects than the brain studies. Our results are the strongest evidence to dateof a common transcriptome signature in the brains of individuals with ASD.

View record

The first chapter of this thesis addresses a common problem in genomics experiments: interpreting a resulting "hit list" of interesting genes. We present work on an approach for summarizing and exploring "hit lists" that makes use of the large amount of gene expression data in public repositories such as the Gene Expression Omnibus. We compare the query list with datasets that we have analyzed for differential expression of genes. Studies that have similarities to the given hit list yield additional insights, help contextualize studies, and serve as a basis for future meta-analysis. A conceptually similar problem that we addressed is the classification or clustering of datasets based on patterns of differential expression. Both problems required a method for determining distances between datasets based on rankings of genes. We tested and benchmarked several methods using manually annotated datasets. The method that performed best according to our evaluation process is based on Kendall's Tau top-k distance. We investigated potential sources of confounds, finding that the largest challenge may be posed by the high prevalence of certain gene expression patterns. These highly prevalent patterns tended to dominate search results. Nonetheless, we demonstrated the effectiveness of this approach in a case study. In the second chapter, we investigated the role of microRNAs in the context of major depression and suicide. We profiled microRNA and messenger RNA levels in post-mortem prefrontal cortex and hippocampus brain tissue of depressed suicides, suicides, and controls. In the prefrontal cortex, we found miR-1202 to be down-regulated in suicides versus controls, and LCT (lactase enzyme) was up-regulated in suicides or depressed suicides compared to controls. The former result was independently confirmed using quantitative PCR. While further study is needed, our results have the potential to provide insight into molecular changes in the brains of depressed and suicidal individuals.

View record

Microarray expression data sets vary in size, data quality and other features, but most methods for selecting coexpressed gene pairs use a ‘one size fits all’ approach. There have been many different procedures for selecting coexpressed gene pairs of high functional similarity from an expression dataset. However, it is not clear which procedure performs best as there are few studies reporting comparisons of these approaches. The goal of this thesis is to develop a set of “best practices” in order to select coexpression links of high functional similarity from an expression dataset, along which methods for identifying datasets likely to yield poor information. With these goals, we hope to improve the quality of gene function predictions produced by coexpression analysis.Using 80 human expression datasets we examined the impact of different thresholds, correlation metrics, expression data filtering and transformation procedures on performance in functional prediction. We also investigated the relationship between data quality and other features of expression datasets and their performance in functional prediction. We used the annotations of the Gene Ontology as a primary metric to measure similarity in gene function, and employ additional functional metrics for validation. Our results show that several dataset features have a greater influence on the performance in functional prediction than others. Expression datasets which produce coexpressed gene pairs of poor functional quality can be identified by a similar set of data features. Some procedures used in coexpression analysis have a negligible effect on the quality of functional predictions while others are essential to achieving the best performance in the algorithm. We also find that some procedures interact greatly with features of expression datasets and that these interactions increase the number of high quality coexpressed gene pairs retrieved through coexpression analysis. This thesis uncovers important information on the many intrinsic and extrinsic factors that influence the performance in functional prediction of coexpression analysis. The information summarized here will help guide future studies using coexpression analysis and improve the quality of gene function predictions.

View record