Immune clock of pregnancy: sub-challenge 1 - predicting gestational age

aya43@sfu.ca; raquel_aoki@sfu.ca

Invent the future ai4all is a 2 week summer enrichment program for grade 10 & 11 girls. This github repository contains materials for the computational biology project group.

Preterm birth (birth on or before 37 weeks of gestation) affects 15 million neonates per year and is the leading cause of infant morbidity and mortality. To understand whether or not a woman and her child is at risk of and de#terventions to prevent preterm birth, clinicians require two key information points: gestational age (GA) and the condition of the fetus in relation to its GA. These help to time care, schedule/interpret antepartum tests, evaluate fetal growth, and thus possibly prevent preterm birth.

GA is currently determined by timing a woman¡¯s last menstrual period or by ultrasound. The former is the most reliable metric but can be inaccurate and subjective based on how a patient self manages her pregnancy. The latter is objective but is costly and less accurate if done prior to 14 weeks of pregnancy. An objective, noninvasive and less costly method to determine GA is by analyzing maternal whole-blood transcriptomics.

Transcriptomics is a technology that studies an organism¡¯s transcriptome. The transcriptome encompasses all the RNA (ribonucleic acid) transcripts (more specifically messenger or mRNA¡¯s) created by replicating different genes in the genome. These RNA fragments are subsequently translated into proteins used to perform biological functions. In other words, transcriptomics is a study of how active each gene is, in contribution to a biological state, based on how many times the mRNA fragments in a sample map to each gene.

The clinical question here is: what maternal whole-blood mRNA genes/probe/isoforms can be used to accurately determine gestational age. This result can guide more practical and less expensive whole-blood transcriptomic tests that target specific genes and their expression profiles. Computationally, the questions are then to determine what model can best produce accurate results while maintaining interpretability of which genes or features contribute to those results.

This project is based on sub-challenge 1 of the preterm birth prediction (transcriptomics) challenge (link, webinar). It expands on a previous paper (GSE113966 DOI: 10.1126/science.aar3819 pdf, supplement) by utilizing a larger data set with heterogeneous patients (samples from women who 1. gave normal birth, 2. had early preeclampsia, and 3. spontaneous preterm delivery or rupture membrane; versus normal birth subjects only), genome-wide (versus targeted gene), whole-blood (versus cell-free mRNA) transcriptomics, etc.

The students will be familiarized with the transcriptomics data type and how this can be preprocessed and used to predict a continuous outcome GA. Students will be able to learn about how machine learning algorithms can be engineered into pipelines in application to real-world problems -- in this case to create equal and inclusive quality care for women of all birth condition whom we may not have GA information on.

Key Learning Opportunities

• Data preprocessing: feature selection, dimensionality reduction, de-noising
• Model construction: for predicting continuous value gestational age in weeks
• Model evaluation: CV (cross validation) experiments, RMSE (root mean square error) scoring
• Interpretation techniques: extract important features and visualize their inter-relations and relation with results

Deliverables

• Domain knowledge: demonstrate the ability to understand and communicate purpose, methods, results, implications, and possible extensions of the project
• Computational knowledge: recognize generalizability of each method used in pipeline and how they can be engineered and customized for specific data sets
• Project results: good feels

itinerary

to-do:

• download data to your "project directory"
  • 00_input/: input data (see below)
  • 01_features/: contains features transformed from 00_input/HTA20_RMA.RData
  • 02_models/: all models tested are saved here organized by feature used; these are loaded back into scripts to generate visualizations and tables
  • 03_results/: resulting visualizations and tables are saved here
  • cvinds.Rdata: indices used for cross validation in training, and testing
• download this github repository
• open and work on bio2019_script.Rmd: R notebook with data processing, feature extraction, model selection and prediction of final results.
  • input: HTA20_RMA.RData (32830 genes x 367+368 train+test samples) matrix
  • output: TeamX_SC1_prediction.csv (368 predicted GA result for test samples; GAs are continuous values 8-42 weeks rounded to 1 decimal); can submit max 5 results to leaderboard, and 1 final result as submission + code + write-up

friday:

• 10:30-11am pitch practice in the stadium
• 2-3pm pitch
• 3-4pm demo time
• 4-5pm closing ceremonies

links:

• data drive (see description below): https://drive.google.com/file/d/1-cnU4SbXujICW4OYbLlFrpupJd-cr8dj/view?usp=sharing
• challenge site (how to participate, leaderboard): https://www.synapse.org/#!Synapse:syn18380862
• ai4all sfu site: https://www.sfu.ca/computing/inventthefuture.html
• ai4all sfu 2019 schedule: booklet

scripts:

1. bio2019_all.R (RUN ME FIRST): script that generates all features and model results found in drive/data/ (except input 00_input); section on feature generation taken from bio2019_FeatureExtraction.R

• edit variables root (your "project directory"), and no_cores (number of cores you want to parrallelize on)
• models variable contains all the models used from the caret package; details can be found on https://topepo.github.io/caret/available-models.html
• pars variable contains all the parameters to test on these models, only the parameters with the best RMSE is kept

2. bio2019_loadtest.R: fill in the blank verion of bio2019_all.R that loads and displays results of each feature/model combination one by one as defined by variables xi (feature) and model

schedule

meetings:

• 2019-05-09 15:00 mentor's startup meeting (ppt, pdf, project ideas)
• 2019-05-17 11:00 project pitch (bio ideas)
• 2019-05-22 10:00 pregnancy project webinar (vod)
• 2019-05-28 15:30 refined project pitch
• 2019-06-28 project code ready
• 2019-07-03 test code on laptops
• 2019-07-10 install environment

challenge:

• 2019-05-04 sub-challenge 1 open
• 2019-05-22 sub-challenge 1 leaderboard (max 5 submissions)
• 2019-08-15 sub-challenge 1 deadline for submission; 2019-09-13 results announced, 2019-11-04 RSG with DREAM conference
• 2019-08-15 sub-challenge 2 open + leaderboard (max 5 submissions)
• 2019-12-05 sub-challenge 2 deadline for submission; 2020-01-03 results announced

program:

• 2019-07-19 project group pitch & preference survey
• 2019-07-22 day 1 explore the challenge, work on the data, and learn about feature extraction
• 2019-07-23 day 2 work on machine learning models and prediction
• 2019-07-25 day 3 work on the presentation of final results
• 2019-07-26 project group presentation

dataset: 00_input/

• HTA20_RMA.RData: maternal whole-blood preprocessed (32830/925032 gene/probeset x 367 train + 368 test samples) expression matrix
  • rownames: ENTREZ-gene(except for "_at" suffix)/probeset IDs
  • colnames: SampleID
• anoSC1_v11_nokey.csv: sample annotation file with the following columns
  • SampleID: unique identifier of the sample (matching the name of the .CEL file in HTA20 folder, except for extension .CEL);
  • GA: gestational age as determined by the last menstrual period and or ultrasound;
  • Batch: the batch identifier;
  • Set: name of the source dataset;
  • Train: 1 for samples to be used for training, 0 for samples to be used for test;
  • Platform: gene expression platform used to generate the cell files.
• TeamX_SC1_prediction.csv: result submission template
• preprocess_data_SC1.R: generates gene level and probeset (exon and exon-junction) level data HTA20_RMA.RData from raw (.CEL) files
  • obtained using the RMA implemented in the oligo package of Bioconductor
  • probsets based on a chip definition file generated from the pd.hta20.hs.entrezg_0.0.1.tar.gz annotation data from http://brainarray.mbni.med.umich.edu; annotation package pd.hta.2.0 (v 3.12.2) from Bioconductor can be used to map probesets to gene identifiers
  • processed separately for each of the 11 experimental batch profiling train + test samples being profiled in each batch
  • then combined across batches
  • quantile normalized
  • batch effect removal using the removeBatchEffects from the limma package of Bioconductor
• pd.hta20.hs.entrezg_0.0.1.tar.gz: annotation package used to generate gene level expression data.
• HT20 folder (not uploaded 47.36gb): 735 raw .CEL files from the HTA 2.0 platform