February 26th, 2019

Welcome to The Future Labs, a Podcast series talking to people building the future, today. We speak to experts around the globe about their role in changing the world for the better.

In this episode, we are exploring The Future of Personalized Cancer Treatment and how computer-based modelling can help identify the best ways to tackle cancer in individual patients or patient populations. We are speaking to Kristof Szalay, Founder and CTO of Turbine.ai, a company using computer-based modelling to reveal the potential of new and existing drugs and to help match these to the right patients.

“We have a dream that in a few decades it would be possible, at the bedside, to do these kinds of simulations [simulating patient cells], but there is still a long way to go…” - Kristof

What is the problem?

Although our understanding of cancers is improving, the underlying biology is complicated and varies significantly between patients. A drug or medical approach that may work well for one patient with cancer can have little or no effect in a different patient due to differences in cancer itself. These differences can arise from genetic mutations which alter the function or expression of a gene, changing the biology of cancer (e.g., how its metabolism functions or how it protects itself from drugs) as well as the underlying genetic differences between patients. Without an understanding of these differences between patients, developing and selecting the right medication (drug) is difficult and results in a reliance on broad-spectrum agents such as chemotherapy, which are a blunt tool to target fast-dividing cancer cells regardless of their nuances.

The management of cancer increasingly includes a better characterization of the tumor itself, as well as tailoring therapy accordingly. For example, for over a decade, patients with breast cancer have had biopsies of their tumors tested for the expression of HER2, a growth factor receptor often overexpressed on breast cancers, to determine whether they should receive targeted HER2 therapy such as trastuzumab (an antibody targeting HER2). This approach has proven to be hugely successful, and patients treated with trastuzumab have such good outcomes that it replacement standard of care for many breast cancer patients shortly after its approval. There are numerous other examples of drugs targeting specific point mutations, such as osimertinib for non-small cell lung cancer with EGFR-T790M substitutions, enasidenib for acute myeloid leukaemia with IDH2 mutation, and most recently larotrectinib was approved for treating TRK-fusion positive cancers regardless of tumor site.

As diagnostic technology and our understanding of cancer has improved, so has the volume of information that we can collect from patient samples. Today is not uncommon in developed countries for patients with solid tumors to have biopsies sequenced to identify a wide panel of specific mutations. In many cases, there are drugs developed for patients harboring these specific mutations, but these remain a tiny proportion of total possible genetic aberrations. The problem is even more complex as each molecule in a cell (e.g., genes, proteins) exist as part of large dynamic networks containing compensatory pathways and feedback loops that make it even harder to predict how cancers will respond to specific therapies. Another shift in recent years has been to process the reams of data we collect from tumors to understand these systems (“systems biology”) and the analysis of “omics” has become common practice (e.g., genomics, transcriptomics, proteomics).

What is Turbine.ai doing?

Turbine.ai is a startup combining this cancer-specific data with their computer-based models of cancer cells (simulated cells) to improve the way drugs are developed, and to improve our understanding of which approaches might work best in which patients. The idea is that if you can simulate cancer cells based on patient- and tumor-specific information (genomic, transcriptomic etc.), then you can predict potential weaknesses (e.g., over-reliance on a specific metabolic pathway) to take advantage of. Turbine.ai’s models characterize cancer cells as networks of interconnecting nodes that represent proteins interacting with each other. Their networks aim to recapitulate the dynamic nature of living cells adding an additional layer of complexity that goes beyond traditional approaches. The quantity of data involved requires the use of advanced data processing techniques and the use of machine learning to make sense of the patterns that emerge.

Turbine.ai has focused initially on modelling cancer cells. In their view, there is a high probability of success in this area, mainly due to the relatively “simple” outcome of assays in cancer – able to kill or not able to kill cancer cells. This binary outcome makes it easier for turbine.ai to test and refine the model. The underlying concept potentially applies to other diseases, and they expect to expand to immunology (esp. in relation to cancer) and other disease areas once they have refined their system in cancer.

Their technology is currently being used to model cancer cell lines, which are “immortalized” cells from patient biopsies that are now used globally for scientific research, and training their simulations based on actual lab work. Turbine.ai has tested 50+ drugs in their simulations using these cell lines in a laboratory, and have found that their models can accurately predict how well these drugs kill cancer cells in around half to two-thirds of experiments.

This level of accuracy is sufficient for their models to be helpful in drug development, and they expect to test novel compounds in their model to help guide which patient populations and types of cancer would be vulnerable to the new drug. There are clear applications for such an approach to tailoring the clinical development of a drug to the patients most likely to respond to it. As the accuracy and versatility of the simulation platform improve, they also hope to find a role in active patient management; i.e., that physicians could feed in patient-specific data and then make treatment decisions based on the output.

This space is evolving rapidly, with significant investment and R&D effort being focused on personalized medicine, the collection of more and more patient and tumor data, as well as increasing use of machine learning approaches to make sense of this data.