Variable prioritization in nonlinear black box methods: A genetic association case study

Crawford, Lorin; Flaxman, Seth; Runcie, Daniel E.; West, Mike

doi:10.1214/18-aoas1222

Cited by 30 publications

(48 citation statements)

References 88 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…To estimate the model in Equation (4), we use an elliptical slice sampling Markov chain Monte Carlo (MCMC) algorithm (Supplementary Note Section 1.1). This allows samples from the approximate posterior distribution of f (given the data), and also allows for the computation of an effect size analog for each topological summary statistic [53–55] where ( X Τ X ) † is the generalized inverse of ( X Τ X ).…”

Section: Methodsmentioning

confidence: 99%

“…SINATRA uses these weights and assigns a measure of relative centrality to each summary statistic (first panel Fig. 1(c)) [55]. This criterion evaluates how much information in classifying each shape is lost when a particular topological feature is removed from the model.…”

Section: Methodsmentioning

confidence: 99%

“…Second, γ is bounded on the unit interval [0, 1] with the natural interpretation of providing relative evidence of association for shape features; higher values suggest greater importance. For this metric, the null hypothesis assumes that every feature equally contributes to the total variance between shape classes, while the alternative proposes that some features are more central than others [55]. As we show in the Supplementary Note, when the null assumption is met, SINATRA displays association results that appear uniformly distributed and effectively indistinguishable.…”

Section: Methodsmentioning

confidence: 99%

“…Code for implementing the SINATRA pipeline is freely available at https://github.com/lcrawlab/SINATRA, and is written in (version 3.5.3). As part of this procedure: (i) inference for the Gaussian process classification (GPC) model using elliptical slice sampling was carried out using the package (version 1.2) [58] and (ii) the computation of effect sizes and association measures for the Euler characteristic curves was done with the “RelATive cEntrality (RATE)” source code in (version 1.0.0; https://github.com/lorinanthony/RATE) [55].…”

Section: Methodsmentioning

confidence: 99%

See 3 more Smart Citations

A Statistical Pipeline for Identifying Physical Features that Differentiate Classes of 3D Shapes

Wang

Sudijono

Kirveslahti

et al. 2019

Preprint

Self Cite

View full text Add to dashboard Cite

22The recent curation of large-scale databases with 3D surface scans of shapes has motivated the devel-23 opment of tools that better detect global-patterns in morphological variation. Studies which focus on 24 identifying differences between shapes have been limited to simple pairwise comparisons and rely on 25 pre-specified landmarks (that are often known). We present SINATRA: the first statistical pipeline for 26 analyzing collections of shapes without requiring any correspondences. Our novel algorithm takes in two 27 classes of shapes and highlights the physical features that best describe the variation between them. We 28 use a rigorous simulation framework to assess our approach. Lastly, as a case study, we use SINATRA to 29 analyze mandibular molars from four different suborders of primates and demonstrate its ability recover 30 known morphometric variation across phylogenies. 31 Introduction 32 Sub-image analysis is an important open problem in both medical imaging studies and geometric mor-33 phometric applications. The problem asks which physical features of shapes are most important for 34 differentiating between two classes of 3D images or shapes such as computed tomography (CT) scans of 35 bones or magnetic resonance images (MRI) of different tissues. More generally, the sub-image analysis 36problem can be framed as a regression-based task: given a collection of shapes, find the properties that 37 explain the greatest variation in some response variable (continuous or binary). One example is identify-38 ing the structures of glioblastoma tumors that best indicate signs of potential relapse and other clinical 39 outcomes [1]. From a statistical perspective, the sub-image selection problem is directly related to the 40 variable selection problem -given high-dimensional covariates and a univariate outcome, we want to 41 infer which variables are most relevant in explaining or predicting variation in the observed response. 42 2Framing sub-image analysis as a regression presents several challenges. The first challenge centers 43 around representing a 3D object as a (square integrable) covariate or feature vector. The transformation 44 should lose a minimum amount of geometric information and apply to a wide range of shape and imaging 45 datasets. In this paper, we use a tool from integral geometry and differential topology called the Eu-46 ler characteristic (EC) transform [1][2][3][4], which maps shapes into vectors without requiring pre-specified 47 landmark points or pairwise correspondences. This property is central to our innovations. 48After finding a vector representation of the shape, our second challenge is quantifying which topological 49 features are most relevant in explaining variation in a continuous outcome or binary class label. We 50 address this classic take on variable selection by using a Bayesian regression model and an information 51 theoretic metric to measure the relevance of each topological feature. Our Bayesian method allows us 52 to perform variable selection for nonlinear func...

show abstract

Section: Methodsmentioning

confidence: 99%

Section: Methodsmentioning

confidence: 99%

Section: Methodsmentioning

confidence: 99%

Section: Methodsmentioning

confidence: 99%

See 2 more Smart Citations

A Statistical Pipeline for Identifying Physical Features that Differentiate Classes of 3D Shapes

Wang

Sudijono

Kirveslahti

et al. 2019

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…Unfortunately, the classic statistical idea of variable selection and hypothesis testing is lost within machine learning methods since they do not naturally produce interpretable significance measures (e.g., P -values or PIPs) like traditional LMMs [33,35].…”

Section: Introductionmentioning

confidence: 99%

Multi-scale Inference of Genetic Trait Architecture using Biologically Annotated Neural Networks

Demetçi

Cheng

Darnell

et al. 2020

Preprint

Self Cite

View full text Add to dashboard Cite

AbstractWith the emergence of large-scale genomic datasets, there is a unique opportunity to integrate machine learning approaches as standard tools within genome-wide association (GWA) studies. Unfortunately, while machine learning methods have been shown to account for nonlinear data structures and exhibit greater predictive power over classic linear models, these same algorithms have also become criticized as “black box” techniques. Here, we present biologically annotated neural networks (BANNs), a novel probabilistic framework that makes machine learning fully amenable for GWA applications. BANNs are feedforward models with partially connected architectures that are based on biological annotations. This setup yields a fully interpretable neural network where the input layer encodes SNP-level effects, and the hidden layer models the aggregated effects among SNP-sets. Part of our key innovation is to treat the weights and connections of the network as random variables with prior distributions that reflect how genetic effects are manifested at different genomic scales. The BANN software uses scalable variational inference to provide fully interpretable posterior summaries which allow researchers to simultaneously perform (i) fine-mapping with SNPs and (ii) enrichment analyses with SNP-sets on complex traits. Through simulations and real GWA data applications, we show that our method improves upon state-of-the-art approaches in both settings across a wide range of genetic architectures.

show abstract

A spectrum of explainable and interpretable machine learning approaches for genomic studies

Conard

DenAdel

Crawford

2023

WIREs Computational Stats

Self Cite

View full text Add to dashboard Cite

The advancement of high‐throughput genomic assays has led to enormous growth in the availability of large‐scale biological datasets. Over the last two decades, these increasingly complex data have required statistical approaches that are more sophisticated than traditional linear models. Machine learning methodologies such as neural networks have yielded state‐of‐the‐art performance for prediction‐based tasks in many biomedical applications. However, a notable downside of these machine learning models is that they typically do not reveal how or why accurate predictions are made. In many areas of biomedicine, this “black box” property can be less than desirable—particularly when there is a need to perform in silico hypothesis testing about a biological system, in addition to justifying model findings for downstream decision‐making, such as determining the best next experiment or treatment strategy. Explainable and interpretable machine learning approaches have emerged to overcome this issue. While explainable methods attempt to derive post hoc understanding of what a model has learned, interpretable models are designed to inherently provide an intelligible definition of their parameters and architecture. Here, we review the model transparency spectrum moving from black box and explainable, to interpretable machine learning methodology. Motivated by applications in genomics, we provide background on the advances across this spectrum, detailing specific approaches in both supervised and unsupervised learning. Importantly, we focus on the promise of incorporating existing biological knowledge when constructing interpretable machine learning methods for biomedical applications. We then close with considerations and opportunities for new development in this space.This article is categorized under: Statistical Models > Nonlinear Models Applications of Computational Statistics > Genomics/Proteomics/Genetics Applications of Computational Statistics > Computational and Molecular Biology

show abstract

Variable prioritization in nonlinear black box methods: A genetic association case study

Cited by 30 publications

References 88 publications

A Statistical Pipeline for Identifying Physical Features that Differentiate Classes of 3D Shapes

A Statistical Pipeline for Identifying Physical Features that Differentiate Classes of 3D Shapes

Multi-scale Inference of Genetic Trait Architecture using Biologically Annotated Neural Networks

A spectrum of explainable and interpretable machine learning approaches for genomic studies

Contact Info

Product

Resources

About