In oncology drug development, clinical trials fail at the final stage of development (PhIII) about half of the time. Our client’s pipeline was no exception to this. Like many organisations across the industry, it was spending an outsized share of its total drug development costs (roughly 45%) in the final stage of development with only a 45% success rate.
In an attempt to cut wasteful investments in the final stages of development, our client looked to improve decision making on what trials to progress from PhII to PhIII of drug development. Could data or AI play a role in helping our client make better decisions and allocating their drug development budgets more wisely?
Making better decisions
By backtesting our approach against sets of historical data, we evaluated that our machine learning solution had a 70% chance of predicting success in Phase III of drug development. It would go on to help our client choose the right portfolio studies to invest in and continues to help our them make better decisions on how to allocate budgets.
Creating new ways of predicting success
Before progressing a trial from PhII to PhIII, many research organisations gauge the probability of success through qualitative judgement. This is due to the perceived complexity of capturing aspects of data and using it to inform decisions. Yet, we’d go on to gather and use data in new and strategic ways.
We developed quantitative scores to capture two key drivers that we felt could help us predict the probability of success more effectively:
- The quality and strength of Phase II results – We considered several trial specific characteristics that could have potentially influenced the quality of PhII results. These included study design, targeted endpoints and the effectiveness of the patient population used.
- The experience of the sponsor at the time of study initiation – Here, by backtesting historical records, we took into account the strength of the organisation conducting the research, the sponsor involved and the experience of those running the study.
We took these two scores and integrated them into a decision making model that used machine learning to identify projects with high probabilities of success. This created a powerful and systematic way of using and presenting data in a way that could help our client make decisions.
What made it successful?
Taking inspiration from Moore’s law and not Eroom’s law…
Less than 50 years ago, an investment of one billion USD brought an average of more than ten new therapies to patients. Today, that number has dropped to less than one.
Eroom’s law describes an exponential drop in R&D productivity over time. Yet Moore’s law, which predicts an exponential increase in computational power over time, is something we should draw inspiration from.
We’re using new technology in pioneering ways. So, as Eroom’s law suggests, let’s acknowledge that traditional methods of innovation are on a downturn. What would it take to turn Eroom around?