(A-16) Structure-based Prediction of Kinase Activation Amidst a Varied Mutational Landscape Using Privileged Learning

Introduction:: Post-translational modifications such as phosphorylation catalyzed by kinases are essential for cell signaling. The activation of mutated kinase in a cancer cell can profoundly impact disease progression and drug efficacy. However, numerous clinical mutations in the human kinome impose challenges in defining the quantitative structure-activity relationship. Previous work (Patil K. et al. PNAS 2021, 118(10), e2019132118.) shows that perturbation of structural properties such as hydrogen-bonding occupancy in the αC-helix and the activation loop computed from molecular dynamics (MD) is a good indicator of the activation status of anaplastic lymphoma kinase (ALK) mutants with 2/3rd of the mutants utilizing this mechanism of activation.

Materials and Methods:: Here we develop a neural network-based privileged learning algorithm that utilizes structural features computed from MD simulation for a fraction of kinase mutants as privileged information to improve the binary classification of kinase activation. 1. Molecular dynamics (MD) simulations are performed to extract structural and energetic descriptors of inactive and active conformations of 221 mutants across five kinases: ALK, HER2, EGFR, MEK1, and BRAF using the GROMACS package. Solvated kinase mutants are modeled using the CHARMM27 and TIP3P force fields. Isothermal-isobaric simulations were conducted at room temperature and ambient pressure for 100 ns. Structural features, including root mean square deviation, hydrogen bond occupancy, solvent accessible surface area, and residue-residue interaction energy of αC-helix, P-loop, catalytic loop, and activation loop domains of both active and inactive kinase conformations are analyzed using GROMACS analysis tools. 2. For privileged learning algorithm, for 221 mutants with privileged information, the weights W and bias factors B of output hidden nodes are obtained from the neural network (Net1) using all 135 features. The values F of output hidden nodes are obtained from Net2 using only 59 mutational features. The adjusted hidden features W*F+B and the activation status are used as inputs and outputs for Net3, respectively. For the rest of mutants without privileged information, W′ and B′ are estimated from a weighted linear combination of W and B obtained from Net1, where the similarity coefficients are calculated using the radial basis function kernel.

Results, Conclusions, and Discussions:: We show that privileged learning methods achieve 72.4% balanced accuracy against ~1000 mutations across ~100 kinases which outperforms other tree-based ML methods which are trained without privileged information.
Also, we show that privileged learning predictions outperform ML algorithms that use K nearest neighbor (KNN) imputation method for estimating the missing structural features for the rest 690 mutants. A possible explanation for the superior performance of privilege learning is that the weakly nonlinear hidden features: output relationship, in contrast to a strongly nonlinear mapping of input features: output, thereby increasing the reliability of prediction for mutants lacking privileged information. Our approach paves way for interpretable prediction of cancer-driving kinase mutations and guiding precision medicine treatment.


Acknowledgements (Optional): : We acknowledge funding from DOE grant (583468 AMND 1/DE-SC0022240) and NIH grants CA250044 and CA 244660. Computational resources were provided in part by XSEDE resources under grant MCB200101.

References (Optional): :