Format

Send to

Choose Destination
Cell Mol Life Sci. 2019 Apr 13. doi: 10.1007/s00018-019-03097-2. [Epub ahead of print]

Computational algorithms for in silico profiling of activating mutations in cancer.

Author information

1
Graduate Group in Biochemistry and Molecular Biophysics, University of Pennsylvania, Philadelphia, PA, USA.
2
Department of Chemical and Biomolecular Engineering, University of Pennsylvania, Philadelphia, PA, USA.
3
Department of Bioengineering, University of Pennsylvania, Philadelphia, PA, USA.
4
Department of Pharmacology, Yale University, New Haven, CT, USA.
5
Cancer Biology Institute, Yale University, West Haven, CT, USA.
6
Department of Pediatrics, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA.
7
Children's Hospital of Philadelphia, Philadelphia, PA, USA.
8
Graduate Group in Biochemistry and Molecular Biophysics, University of Pennsylvania, Philadelphia, PA, USA. rradhak@seas.upenn.edu.
9
Department of Chemical and Biomolecular Engineering, University of Pennsylvania, Philadelphia, PA, USA. rradhak@seas.upenn.edu.
10
Department of Bioengineering, University of Pennsylvania, Philadelphia, PA, USA. rradhak@seas.upenn.edu.

Abstract

Methods to catalog and computationally assess the mutational landscape of proteins in human cancers are desirable. One approach is to adapt evolutionary or data-driven methods developed for predicting whether a single-nucleotide polymorphism (SNP) is deleterious to protein structure and function. In cases where understanding the mechanism of protein activation and regulation is desired, an alternative approach is to employ structure-based computational approaches to predict the effects of point mutations. Through a case study of mutations in kinase domains of three proteins, namely, the anaplastic lymphoma kinase (ALK) in pediatric neuroblastoma patients, serine/threonine-protein kinase B-Raf (BRAF) in melanoma patients, and erythroblastic oncogene B 2 (ErbB2 or HER2) in breast cancer patients, we compare the two approaches above. We find that the structure-based method is most appropriate for developing a binary classification of several different mutations, especially infrequently occurring ones, concerning the activation status of the given target protein. This approach is especially useful if the effects of mutations on the interactions of inhibitors with the target proteins are being sought. However, many patients will present with mutations spread across different target proteins, making structure-based models computationally demanding to implement and execute. In this situation, data-driven methods-including those based on machine learning techniques and evolutionary methods-are most appropriate for recognizing and illuminate mutational patterns. We show, however, that, in the present status of the field, the two methods have very different accuracies and confidence values, and hence, the optimal choice of their deployment is context-dependent.

KEYWORDS:

Driver mutations; Machine learning; Molecular dynamics; Passenger mutations; Structural bioinformatics

PMID:
30982079
DOI:
10.1007/s00018-019-03097-2

Supplemental Content

Full text links

Icon for Springer
Loading ...
Support Center