skip to content

Department of Physiology, Development and Neuroscience

 
Cancer prediction through artificial neural network algorithms

International team of researchers led by Paul Schofield developes a new algorithmic method capable of analyzing molecular data collected in studies and identifying genetic patterns associated to the development of cancer

A collaborative team lead by Paul Schofield from PDN, George Gkoutos from Birmingham and Robert Hoehndorf from KAUST's Computational Bioscience Research Center have recently shown that they could identify genes with a known causative role in cancer and pick out dozens of putative new ones for 20 different tumor types

To date, the search for genes with a causal role in cancer has been carried out by starting with DNA sequence data. By extensively cataloging tumor mutations shared among patients with a common type of cancer, the research community has documented hundreds of genes with a causal impact on tumor development. Experimental follow up is then used to functionally associate these genes with the hallmarks of cancer. This new method turns this approach on its head, essentially, the approach is knowledge-driven and only uses tumor sequencing data as validation. This is unlike most approaches, which are data-driven combined with interpretation of the findings with respect to established knowledge.

The method takes advantage of artificial neural network algorithms to sift through reams of molecular data collected from studies of cancer cell lines, mouse models and human patients, and teaching the artificial intelligence system to link tumor genome mutations to altered cell functionality.

An algorithm was created to recognize functional and phenotypic patterns that predispose a gene toward playing a role in driving tumor development. The model was then validated by using a publicly available database of some 27,000 different tumor variants as well as functional and sequence data--showing that it could accurately categorize known cancer-driving genes and detect more than 100 other likely culprits, many with specific roles in particular tumor types.

The algorithm was then further tested on molecular data gathered from two cohorts of cancer patients. The first was from King Abdulaziz University Hospital in Saudi Arabia, comprising 26 tumor samples from individuals with a rare type of head and neck cancer called nasopharyngeal carcinoma. The other cohort comprised 114 colorectal cancer samples from patients treated at the University of Birmingham Hospital in the United Kingdom. In both patient groups, the model singled out candidate driver genes that were frequently mutated and shared pathogenic features of other cancer-causing genes

The prediction method--described in Scientific Reports and freely available online--could help clinicians tailor medicines to the molecular subtypes of patients. It could also be used by drug companies in the hunt for new therapeutic targets.

Reference: Althubaiti, S., et al. (2019) Ontology-based prediction of cancer driver genes. Scientific Reports. doi.org/10.1038/s41598-019-53454-1.