Listen to this story
Indian Institute of Technology Madras (IIT Madras) researchers have developed an Artificial Intelligence-based tool, ‘PIVOT’, that can predict cancer-causing genes in an individual. This tool will help in devising personalised cancer treatment strategies The findings of the research have been published in a peer-reviewed journal called Frontiers.
The research was led by Professor Raghunathan Rengaswamy, Dean (Global Engagement), IIT Madras, and Professor, Department of Chemical Engineering, IIT Madras, Dr. Karthik Raman, Associate Professor, Bhupat and Jyoti Mehta School of Biosciences, IIT Madras and a Core Member, Robert Bosch Centre for Data Science and Artificial Intelligence (RBCDSAI), IIT Madras, and Malvika Sudhakar, a Research Scholar, IIT Madras.
Analytics India Magazine interacted with the researchers to find out more about this innovation and its future prospects.
Sign up for your weekly dose of what's up in emerging technology.
“We had previously developed a tool called cTaG to identify cancer-causing genes across multiple cancers. While this method helps to identify pan-cancer-causing genes, it doesn’t tell us about the genes causing cancer in a patient. For patients with no mutation in known cancer-causing genes, it is difficult to identify targeted therapies,” said Dr Karthik Raman. This got the team interested in developing a tool for the identification of cancer-causing genes, which led to the idea of PIVOT.
The researchers used multi-omics data, which means it includes mutation, gene expression, and copy number variation data from patients of a given cancer type. They labelled genes of an individual as tumour suppressor gene, oncogene or neutral gene. As the number of cancer-causing genes is far fewer than neutral genes, they used ML algorithms, which take care of the imbalance. Finally, they used different metrics to evaluate the models and identify the best among them.
Download our Mobile App
Labelling cancer-causing genes for training
In terms of challenges, Malvika feels that the major challenge was labelling cancer-causing genes in individuals to use for training. “Previous methods use unsupervised strategies to identify personalised cancer-causing genes. Second, we needed to integrate different data types such as mutation, gene expression, and copy number variation to best capture the information across data modalities. Lastly, as mentioned earlier, the number of cancer-causing genes is far fewer than neutral genes, and hence care needs to be taken while model building,” she adds.
In order to overcome these challenges, the researchers formulated a supervised problem by defining four strategies for labelling data. Then, they evaluated and identified the best method for labelling and training personalised cancer-causing genes.
Imbalance-based algorithms to build better models
“Notably, this provides strategies for labelling personalised driver genes for future research. We define features that are used by the models for prediction to integrate data across different data types, which help in capturing the information and understanding the system as a whole. To account for the imbalance of cancer-causing and neutral genes, we use imbalance-based algorithms to build better models, states Malvika.
With PIVOT, it is possible to label cancer-causing genes as tumour suppressor genes and oncogenes, unlike the previous tools. Identifying genes for an individual helps to understand the differences observed within the same cancer type.
“Identification of personalised cancer-causing genes is the first step toward personalised medicine. Methods like PIVOT are required to push the boundaries of personalised medicine. We identify cancer-causing genes that are mutated in even a single tumour, which allows for the identification of rare cancer-causing genes that are very difficult to identify by using existing tools,” adds Malvika.
Customising PIVOT for Indian genomic data
The team plans to expand PIVOT to other cancer types, further adding that the tool has not been exposed a lot to Indian cancer genomic data. That will be the immediate focus. Experimental validation and reiteration to include new data types can only help improve models.
“We are currently working on including the predictions made by PIVOT to predict the response to drugs and rank drugs for personalised treatment. While we do not have any plans to commercialise at the moment, we are looking to explore and analyse personalised cancer-causing genes in an Indian context,” concludes Dr Karthik.