Fig. 2: Potential association between kinase and AE as identified by ML modeling.

a Variable importance (VIMP) assessment of n = 444 predictive variables for five representative AEs. Bar length indicates the VIMP value of the variable, which represents the difference in the out-of-bag model prediction errors with and without this predictive variable being permuted⁴³. The identified top 25 predictive variables are listed for each AE. The blue-highlighted kinases are the representative experimentally well-established KI–AE pairs⁶. The yellow-highlighted pairs are validated by KI–AE pairs found in literature survey results for KI–AE pairs post the⁶ publication review (for references, see Supplementary Table 3). b Kaplan–Meier survival (no hypertension) probability curves stratified by VEGFR2 (top) and JAK2 (bottom) inhibitions based on individual-level drug exposures. Each point on the curve is the average of the ensemble survival function over all patients for a given time, with the error bands showing 95% pointwise confidence intervals for all patients together. The divergence between the two survival curves showed that the patients with “higher inhibition” (>median) on VEGFR2 had earlier hypertension onset than those with “lower inhibition” (<median), while the inhibition on JAK2 showed the opposite outcome.