Extended Data Fig. 4: Robustness assessment.

a. We used bootstrapping to resample the training data (with replacement) 1000 times, each containing 10% to 150% of the training dataset in 10% increments. Box and whiskers graphs showed the correlation coefficient r value in each resampling. The bounds of box demonstrated data extend from the 25th to 75th percentiles. The centre line in the box was plotted at the median. The whiskers went down to the smallest and up to the largest value. Stability in correlation coefficient after about 40% of the sample size were observed. b. To estimate the stability of the findings across subsamples, we resampled the same proportion 90% of original sample size as train dataset for 1000 times, reran the sCCA algorithm and calculated the correlation between the resulting feature in the remaining 10% test dataset.