Standardizing workflows in imaging transcriptomics with the abagen toolbox.

Markello RD, Arnatkeviciute A, Poline JB, Fulcher BD, Fornito A, Misic B. · 2021

Read Standardizing workflows in imaging transcriptomics with the abagen toolbox. by Markello RD, Arnatkeviciute A, Poline JB, Fulcher BD, Fornito A, Misic B. (2021) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc8660024.txt.

Introduction

Markello and co-authors introduce abagen, a software toolbox that standardizes work with brain gene expression data. Gene activity shapes the structure and function of the human brain. Open atlases let teams study this link in living people. Yet each group processed the data its own way, so results varied.

The team tested 750,000 distinct processing pipelines on one atlas. Pipeline choice had a large effect on findings. Some choices swung result correlations by a full point or more. Steps that set gene normalization mattered most. The toolbox gives the field one shared, open route. The authors argue this base will make imaging transcriptomics more sound.

Catalog record

CREATORMarkello RD, Arnatkeviciute A, Poline JB, Fulcher BD, Fornito A, Misic B.
YEAR2021
COLLECTIONPAPERS
FORMATTXT
FILE SIZE108 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC8660024

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Standardizing workflows in imaging transcriptomics with the abagen toolbox By Ross D Markello, Aurina Arnatkeviciute, Jean-Baptiste Poline, Ben D Fulcher, Alex Fornito, Bratislav Misic, Saad Jbabdi, Tamar R Makin Published in: eLife DOI: 10.7554/eLife.72129 ABSTRACT Gene expression fundamentally shapes the structural and functional architecture of the human brain. Open-access transcriptomic datasets like the Allen Human Brain Atlas provide an unprecedented ability to examine these mechanisms in vivo; however, a lack of standardization across research groups has given rise to myriad processing pipelines for using these data. Here, we develop the abagen toolbox, an open-access software package for working with transcriptomic data, and use it to examine how methodological variability influences the outcomes of research using the Allen Human Brain Atlas. Applying three prototypical analyses to the outputs of 750,000 unique processing pipelines, we find that choice of pipeline has a large impact on research findings, with parameters commonly varied in the literature influencing correlations between derived gene expression and other imaging phenotypes by as much as ρ ≥ 1.0. Our results further reveal an ordering of parameter importance, with processing steps that influence gene normalization yielding the greatest impact on downstream statistical inferences and conclusions. The presented work and the development of the abagen toolbox lay the foundation for more standardized and systematic research in imaging transcriptomics, and will help to advance future understanding of the influence of gene expression in the human brain. INTRODUCTION Technologies like magnetic resonance imaging (MRI) provide unique insights into macroscopic brain structure and function in vivo. Modern research increasingly emphasizes how microscale attributes, such as gene expression, influence these imaging-derived phenotypes (Fornito et al., 2019; Arnatkeviciute et al., 2019; Arnatkevičiūtė et al., 2021). Gene expression is particularly useful as it is a fundamental molecular phenotype that can be plausibly linked to the function of biological pathways (Whitaker et al., 2016; Seidlitz et al., 2018), protein synthesis (Zheng et al., 2019), receptor distributions (Beliveau et al., 2017; Nørgaard et al., 2021; Shine et al., 2019; Deco et al., 2020; Preller et al., 2018), and cell types (Hansen et al., 2021; Anderson et al., 2020b; Anderson et al., 2018; Seidlitz et al., 2020; Gao et al., 2020). However, researchers looking to bridge these macro- and microscopic phenotypes must overcome multiple challenges. Although there are numerous technical and analytic considerations, one foundational issue is that acquiring high-quality transcriptomic data from the human brain is both costly and highly invasive, requiring budgets far greater than most typical neuroimaging studies and restrictive access to tissue from post-mortem donors or cranial surgical patients. As such, researchers must often rely on freely available repositories of gene expression data. There exist multiple open-access repositories for gene expression in the human brain, including BrainSpan (Miller et al., 2014; Kang et al., 2011) and PsychENCODE (Gandal et al., 2018; Li et al., 2018; Wang et al., 2018; among others: Sousa et al., 2017; Darmanis et al., 2015; Lake et al., 2016); however, these datasets generally provide relatively sparse anatomical coverage, limiting the types of analyses that can be performed. Thus, researchers who aim to compare transcriptomic expression with whole-brain imaging-derived phenotypes have primarily relied on the Allen Human Brain Atlas (AHBA; Hawrylycz et al., 2012; Hawrylycz et al., 2015). Initially released in 2010, the AHBA remains the most spatially comprehensive dataset of its kind. Derived from bulk microarray analysis of tissue samples obtained from six donors, the AHBA provides expression data for more than 20,000 genes across 3702 brain areas in MRI-derived stereotactic space. With its superior resolution, the AHBA has significantly contributed to the emergence of the field of imaging transcriptomics (Fornito et al., 2019), enabling dozens of studies over the past decade examining relationships between gene expression and an array of macroscale imaging attributes, including cortical thickness (Shin et al., 2018), myelination (Burt et al., 2018), developmental brain maturation (Whitaker et al., 2016; Kirsch and Chechik, 2016), structural brain networks (Seidlitz et al., 2018; Romero-Garcia et al., 2018; Arnatkevičiūtė et al., 2020), functional brain networks (Richiardi et al., 2015; Krienen et al., 2016; Vértes et al., 2016), and human cognition (Fox et al., 2014; Hansen et al., 2021). The AHBA has also highlighted the importance of whole-brain gene expression in neurological and psychiatric diseases, where it has become increasingly clear that transcriptional pathways play a critical role in shaping the broader dynamics of disease progression and emergent symptomatology (Zheng et al., 2019; Shafiei et al., 2021; Henderson et al., 2019; Vogel et al., 2020; Rittman et al., 2016; Anderson et al., 2020a; Romme et al., 2017; McColgan et al., 2018; Morgan et al., 2019). Since its release, several software toolboxes have been developed to help researchers use transcriptional data from the AHBA (French and Paus, 2015; Gorgolewski et al., 2015; Rittman et al., 2017; Rizzo et al., 2016); however, these tools often focus primarily on facilitating integration of the AHBA with neuroimaging data, offering limited if any functionality for modifying how the data are processed prior to analysis. Instead, a recent comprehensive review revealed that many research groups have opted to develop their own processing pipelines for the AHBA (Arnatkeviciute et al., 2019). Unfortunately, as there are no field-accepted standards for processing imaging transcriptomic data, the generated pipelines vary substantially across groups. The extent to which such processing variability affects analytic outcomes from the AHBA remains unknown. Indeed, over the past decade neuroimaging research has shown that methodological variability can have broad influences on analyses using structural MRI (Bhagwat et al., 2021; Kharabian Masouleh et al., 2020), diffusion MRI (Oldham et al., 2020; Maier-Hein et al., 2017; Schilling et al., 2019), task fMRI (Carp, 2012; Botvinik-Nezer et al., 2020), and resting-state fMRI (Parkes et al., 2018; Ciric et al., 2017). Although researchers are beginning to grapple with the consequences of this variability, the lack of baseline gene expression datasets against which to compare new results impedes the development of standardized practices. In these situations, some researchers have proposed performing ‘multiverse’ analyses (Steegen et al., 2016; Dragicevic et al., 2019), wherein all possible permutations of data processing are analyzed and the full range of analytic results reported. Although such analyses can be computationally intensive, they offer a path to understand how processing choices impact statistical inferences and conclusions, and provide a mechanism by which to help researchers converge on an optimal pipeline. Here, we comprehensively investigate how different processing choices influence the results of analyses using the AHBA. First, we develop an open-source Python toolbox, abagen, that collates all possible processing parameters into a set of turn-key workflows, optimized for flexibility and ease-of-use. We then use the toolbox to process the AHBA through approximately 750,000 unique pipelines. Across three prototypical imaging transcriptomic analyses, we examine whether and how these different processing options modify derived statistical estimates and quantify the relative importance of each option. Next, we replicate a curated set of processing pipelines from the literature to assess how previously reported findings compare to the full range of potential outcomes observed across all examined pipelines. Finally, we end with a set of recommendations, integrated directly into the developed abagen toolbox, to promote standardized use of the AHBA in future work. RESULTS We introduce the abagen toolbox, an open-access software package designed to streamline

Read the complete text