PIVOT: The AI-based tool that can predict cancer-causing genes

Edited excerpts:

How does PIVOT work? 

It is a machine learning tool which uses different kinds of data, such as mutation and gene expression, to predict cancer-causing genes, called driver genes, in a patient.

What is generally used are tools that predict driver genes by analysing multiple samples from the same cancer type to identify genes responsible for the progression of tumours for that particular cancer type.

Such methods are not capable of identifying rare driver genes.

A personalised approach identifies driver genes in an individual and can identify driver genes mutated in as few as a single sample.

How can this tool help in devising personalised cancer treatment strategies?

PIVOT uses supervised model learns (ML) algorithms from known driver genes in an individual. Patient data and their corresponding personalised driver genes are not known. Our first challenge was labelling the data to train on.

We define multiple methods for labelling personalised driver genes and evaluate them.

Secondly, we build models on multiple data types. We consider mutation data, expression data and a combination of all data types, called multi-omic data.

This gives freedom to the user to use the model based on available data. We found that the multi-omic models perform the best.

The first step towards precision medicine is the identification of altered genes to devise a treatment plan.

Personalised methods, such as PIVOT, pave the way for the identification of targetable genes.

Also read: NTAGI's technical sub-committee recommends use of Corbevax, Covaxin for children between 5-12 years

Can this tool also predict the risk of cancer relapse? Are you planning to do research in this area as well?

No, this tool does not predict the risk of cancer relapse. It focuses on cancer-causing genes, using sequencing data from the tumour in the patient. If data on recurrence is available, we can study the correlation between the identified cancer-causing genes and probability of recurrence.

Currently, we are studying the relationship between personalised driver genes and drugs that can be used to improve survival.

Is this tool more accurate in predicting outcomes than traditional methods? 

The field of personalised medicine is very new. We compare our methods with existing tools, using a pan-cancer gene list, and find that PIVOT performs better for all three cancer types.

As we mentioned earlier, there is no gold-standard dataset available to compare the personalised predictions with. In such cases, we use a pan-cancer gene list for comparison. We are currently collaborating to validate our personalised results in an experimental setting.

Are you planning to roll out this tool for treatment surveillance of cancer patients at the national level?

The road to clinical implementation of personalised medicine is very long. We are looking forward to predicting personalised cancer-causing genes in Indian patients.

It is very interesting to study the variations observed in the Indian population when compared to the world.

Currently, we are working to identify personalised cancer-causing genes to predict the response to various drugs. We are developing a scoring mechanism to identify and rank drugs that can be used for treatment.

Also read: Healthcare associations write to FM, raise concerns over 5% GST on hospital room rent

There are other AI tools which too claim can predict cancer-causing genes in individuals. How different is PIVOT? 

Most of the established methods for predicting cancer-causing genes are unsupervised. We define ways to label personalised genes, which can be used by future methods to build supervised models.

The previous tools also integrate only mutation and expression data. PIVOT uses mutation, gene expressions, copy number variation of genes and network data to build models.

The mutation data by other methods do not consider the functional impact of the mutation. They only look at whether a gene is mutated or not. We consider the functional impact of the mutation on the gene by including data in models such as the type of mutation and multiple scores to define how damaging the mutation is.

This multi-omic approach captures maximum biological information to make predictions.

Last, and most importantly, we label the cancer-causing genes as tumour suppressor genes (TSG) and oncogene (OG) based on the functionality of the gene.

Other methods simply rank all the genes as cancer-causing genes. We label them as TSG, OG or neutral. This allows us to differentiate the role of genes that function differently in different individuals of the same cancer type.