Some biopharmaceutical enterprises are taking a leap of faith by reviving failed investigational therapies and aiming them towards new indications and freshly reformulated study designs. With the help of artificial intelligence (AI), new clinical trials are targeting optimized patient selection criteria, novel endpoints and improved safety and efficacy signals. In this interview, Krishnan Nandabalan, CEO of InveniAI, will discuss how AI is shifting the clinical trial paradigm in oncology and central nervous system (CNS) studies.
Krishnan Nadabalan:We are all just sitting on fifty to sixty years of clinical trial data. What used to be reported as meta-analysis is now possible at a much grander scale, because of AI. The prominent place to start with AI-enabled meta-analysis is to identify the right patient segment for the trial. The success of the trial is going to depend on whether the right kind of patient is included, if there is enough power to detect clinical signals and whether the patients can be recruited in a timely manner.
The hope is to gain insight from the data that allow for the better design of these trials to avoid adverse events and to identify super-responders. This becomes very important in those diseases where patient recruitment is hard, such as second line or third line therapies in oncology trials. The use of AI to improve the success rate in these instances is prudent. Additionally, AI-powered analysis of clinical trial data can very quickly identify signals that point to the best clinical sites – not necessarily at low cost, but in an efficient manner that will achieve the end goal of the trial most efficiently and successfully. The most expensive trial is not the one that runs at a high cost but the one that fails.
KN:The digitization of medical records has surged in recent years, and has provided a valuable bank of data that can be tracked over time. While these electronic medical records anonymize patient names, a meta-analysis can be used to identify a region and maybe even particular sites that have the best chance of recruiting the ‘right type’ of patient for a trial. The challenge in running the analysis is stitching together the different data sources. Inconsistency in recording diagnoses is common, as physicians and other treating stakeholders have their own style of recording a diagnosis and treatment regimen. For example, some will say the patient is ‘relapsed’, and others might say the patient is ‘refractory’ or ‘not responding’. Nonetheless, an AI-powered analysis is capable of objectively evaluating patients and determining a more consistent reporting of diagnosis, which can be helpful during a clinical trial prescreening process. Moreover, manually mining this kind of data set can be a complex task and is resource intensive for site staff, but using AI will allow sites to quickly find patients that meet the inclusion criteria – increasing the speed and efficiency of the prescreening and the patient identification process.
KN:I think the biggest challenge that AI can help us overcome is to clean, curate, and connect the data. Clinical trial data is highly variable, and AI can be used to harmonize all the data, analyze it in an unbiased manner and have consistent standards applied so that when experts view the analysis, they are comparing apples to apples. And as simple as it may sound, it’s actually a very complex problem.
But creating a hammer and finding a nail is the wrong way to do it, it’s finding the nail and then creating the hammer or whatever tool you need is more optimal. And even with AI, whatever the question may be, it must start with a fundamental business unit question. Typically, tech companies have built the hammer without knowing the nail, so that is the disconnect in the industry. By taking the question from a pharma and biotech’s perspective, we can understand the business case better and devise technologies and workflows to help address those business questions.
KN: In a collaboration of ours, we analyzed numerous drugs that had not met the endpoint because the wrong patient segment was targeted or did meet the endpoint, but the trial was launched overseas. Our challenge was to leverage all of that clinical trial data and see if we could find the right disease indication and the right patient segment for these drugs. We narrowed the drug candidates down to their final concepts, so that our collaborator to could potentially launch the trial again. With our help, a new patient segment and even a new disease altogether could be paired with an ‘old’ drug.
Our sister company, BioXcel Therapeutics, is developing an oncology drug that had failed clinical trials in the early 2000s. But the patient selection criteria the previous developer was using was not optimal, and they were applying the drug to the wrong tumors. By analyzing the whole field of immunology, we drew a different conclusion and found that if we applied the drug to the right tumor type, it might be efficacious. That trial has already started.
Most of our success comes in identifying the right patient group based on the mechanism of action of the drug. It’s even harder to identify patients who may have a sporadic adverse event. Ultimately, we will reach a point, maybe thirty or fifty years from now, when we will have enough data where artificial intelligence and machine learning will come in handy, so that we can predict those types of cases. But true success is when the drug clears the registration trial and is available for patients.
KN:Currently, one of the projects is working with Centrexion, a pharma company developing non-opioid pain therapeutics. They came to us with three pre-clinical early stage programs that hadn’t yet entered clinical trials and wanted to know where the best signals of efficacy were and in which patient segment they occurred. Even in pain therapeutics, there is an immense amount of data that just can’t be sorted through by humans, which is where AI steps in. We came up with about a dozen patient segments for three programs where one of the programs has progressed to a Phase 2 trial.
Based on a strong efficacy signal we detected, the program was licensed. The power of AI is in its ability to leverage all this disparate and diverse data to align the mechanism of the drug to the most optimal patient segment so that the chances of success in a trial are enhanced. Not just in building efficiencies and building from automation or faster recruitment, but in finding the right patient segment for developing the drug.