Data was obtained from the UK Biobank under Application Number 47850. First, we downloaded the “l2r” files from the UK Biobank. Each chromosome has a separate “l2r” file. Each “l2r” file contained 488,377 columns and a variable number of rows. Each column represented a unique patient in the dataset, who is only identified by an encoded identification number. Each row represented a measurement at a different location in the genome. The values in the file represent the log (base 2) of the ratio of measured intensity measured in a microarray relative to the expected two copies at that location in the genome.
After downloading the “l2r” data from the UK Biobank, we computed the mean l2r value for a portion, we chose 25%, of the chromosome for each patient in the dataset. This process produced a dataset where each person was represented by a series of 88 numbers. Each number represents the length variation for 25% of the 22 non-sex chromosomes. A value of 0 (log2 ration) represents the nominal average length of that portion of the particular chromosome. We call this dataset the chromosomal-scale length variation (CSLV) dataset.
This CSLV dataset was matched with the UK Biobank COVID-19 dataset. The COVID-19 data were provided to UK Biobank by Public Health England. UK Biobank matched the person in the Public Health England data with UK Biobank’s internal records to produce the person’s encoded participant identification number. The dataset we have, provided by UK Biobank contains the participant ID, date the specimen was taken, laboratory that processed the sample, whether the patient was an inpatient when the sample was taken, and the result (positive/negative) of the test. The UK Biobank continues to update the data approximately biweekly.
The criteria for testing and interpretation of results in the UK Biobank COVID-19 data has evolved. A positive test in this dataset earlier than 27 April 2020 was a good indication that the person had severe disease. During this initial period of the pandemic, SARS-CoV-2 testing was only performed on symptomatic people and this particular dataset only includes people tested in a hospital. After 27 April 2020, NHS instructed hospitals to test all non-elective patients admitted, including asymptomatic patients. The UK Biobank dataset released after 27 May 2020 includes “pillar 2” positive test results. These “pillar 2” tests include people in hospitals for non-elective procedures and staff screening. These results can include asymptomatic patients.
Table 1
We segmented the dataset into three overlapping subsets. The first, which we called “1930” contained all UK Biobank participants born after 1930 who had a severe reaction to SARS-CoV-2 infection before 27 April 2020. The two subsets contained people born after 1940 and after 1950.
Dataset | Number |
1930 (< 90 years of age) | 981 |
1940 (< 80 years of age) | 880 |
1950 (< 70 years of age) | 468 |
Using the CSLV-COVID-19 dataset, we selected all people who tested positive before 27 April 2020 and labelled these as people having a severe reaction to COVID-19. We segmented these into three overlapping datasets, as shown in Table 1. We constructed an age-matched control group of the same size that had an identical age profile as those in the severe reaction group. The age-matched control group was selected from the entire UK Biobank dataset, excepting those few who had a severe reaction to COVID-19. Since only a small fraction of the people in the UK Biobank had a severe reaction to COVID-19, we could rerun the analysis with a different age-matched control group many times to build up statistics. We chose this method of selecting the control group based the finding that severe reactions to COVID-19 are both a strong function of age and uncommon (only about 20% of those infected with SARS-CoV-2 require ICU admission even among those in their 70 s)[4, 5].
We used the H2O machine learning package in R to create XGBoost[6] models that were trained to classify a person in the dataset, consisting of those who had a severe reaction and age-matched controls, based solely on their chromosomal-scale length variation data.