Genotype prediction of 336,463 samples from public expression data.

Tens of thousands of RNA-sequencing experiments comprising hundreds of thousands of individual samples have now been performed. These data represent a broad range of experimental conditions, sequencing technologies, and hypotheses under study. The Recount project has aggregated and uniformly processed hundreds of thousands of publicly available RNA-seq samples. Most of these samples only include RNA expression measurements; genotype data for these same samples would enable a wide range of analyses including variant prioritization, eQTL analysis, and studies of allele specific expression. Here, we developed a statistical model based on the existing reference and alternative read counts from the RNA-seq experiments available through Recount3 to predict genotypes at autosomal biallelic loci in coding regions. We demonstrate the accuracy of our model using large-scale studies that measured both gene expression and genotype genome-wide. We show that our predictive model is highly accurate with 99.5% overall accuracy, 99.6% major allele accuracy, and 90.4% minor allele accuracy. Our model is robust to tissue and study effects, provided the coverage is high enough. We applied this model to genotype all the samples in Recount 3 and provide the largest ready-to-use expression repository containing genotype information. We illustrate that the predicted genotype from RNA-seq data is sufficient to unravel the underlying population structure of samples in Recount3 using Principal Component Analysis.

Postprediction Inference for Clinical Characteristics Extracted With Machine Learning on Electronic Health Records.

PURPOSE: Real-world data (RWD) derived from electronic health records (EHRs) are often used to understand population-level relationships between patient characteristics and cancer outcomes. Machine learning (ML) methods enable researchers to extract characteristics from unstructured clinical notes, and represent a more cost-effective and scalable approach than manual expert abstraction. These extracted data are then used in epidemiologic or statistical models as if they were abstracted observations. Analytical results derived from extracted data in this way may differ from those given by abstracted data, and the magnitude of this difference is not directly informed by standard ML performance metrics. METHODS: In this paper, we define the task of postprediction inference, which is to recover similar estimation and inference from an ML-extracted variable that would be obtained from abstracting the variable. We consider fitting a Cox proportional hazards model that uses a binary ML-extracted variable as a covariate and evaluate four approaches for postprediction inference in this setting. The first two approaches only require the ML-predicted probability, while the latter two additionally require a labeled (human abstracted) validation data set. RESULTS: Our results for both simulated data and EHR-derived RWD from a national cohort demonstrate that we can improve inference from ML-extracted variables by leveraging a limited amount of labeled data. CONCLUSION: We describe and evaluate methods for fitting statistical models using ML-extracted variables subject to model error. We show that estimation and inference is generally valid when using extracted data from high-performing ML models. More complex methods that incorporate auxiliary labeled data provide further improvements.

Methods for correcting inference based on outcomes predicted by machine learning.

Many modern problems in medicine and public health leverage machine-learning methods to predict outcomes based on observable covariates. In a wide array of settings, predicted outcomes are used in subsequent statistical analysis, often without accounting for the distinction between observed and predicted outcomes. We call inference with predicted outcomes postprediction inference. In this paper, we develop methods for correcting statistical inference using outcomes predicted with arbitrarily complicated machine-learning models including random forests and deep neural nets. Rather than trying to derive the correction from first principles for each machine-learning algorithm, we observe that there is typically a low-dimensional and easily modeled representation of the relationship between the observed and predicted outcomes. We build an approach for postprediction inference that naturally fits into the standard machine-learning framework where the data are divided into training, testing, and validation sets. We train the prediction model in the training set, estimate the relationship between the observed and predicted outcomes in the testing set, and use that relationship to correct subsequent inference in the validation set. We show our postprediction inference (postpi) approach can correct bias and improve variance estimation and subsequent statistical inference with predicted outcomes. To show the broad range of applicability of our approach, we show postpi can improve inference in two distinct fields: modeling predicted phenotypes in repurposed gene expression data and modeling predicted causes of death in verbal autopsy data. Our method is available through an open-source R package: https://github.com/leekgroup/postpi.

Human splicing diversity and the extent of unannotated splice junctions across human RNA-seq samples on the Sequence Read Archive.

BACKGROUND: Gene annotations, such as those in GENCODE, are derived primarily from alignments of spliced cDNA sequences and protein sequences. The impact of RNA-seq data on annotation has been confined to major projects like ENCODE and Illumina Body Map 2.0. RESULTS: We aligned 21,504 Illumina-sequenced human RNA-seq samples from the Sequence Read Archive (SRA) to the human genome and compared detected exon-exon junctions with junctions in several recent gene annotations. We found 56,861 junctions (18.6%) in at least 1000 samples that were not annotated, and their expression associated with tissue type. Junctions well expressed in individual samples tended to be annotated. Newer samples contributed few novel well-supported junctions, with the vast majority of detected junctions present in samples before 2013. We compiled junction data into a resource called intropolis available at http://intropolis.rail.bio . We used this resource to search for a recently validated isoform of the ALK gene and characterized the potential functional implications of unannotated junctions with publicly available TRAP-seq data. CONCLUSIONS: Considering only the variation contained in annotation may suffice if an investigator is interested only in well-expressed transcript isoforms. However, genes that are not generally well expressed and nonetheless present in a small but significant number of samples in the SRA are likelier to be incompletely annotated. The rate at which evidence for novel junctions has been added to the SRA has tapered dramatically, even to the point of an asymptote. Now is perhaps an appropriate time to update incomplete annotations to include splicing present in the now-stable snapshot provided by the SRA.