An infinite digital space for human ingenuity

When OpenAI shared its chatbot with the public in late 2022, it took the world by storm, triggering an unprecedented AI boom that has since been firing from all engines, with seemingly no end in sight. It was not the first AI breakthrough that unleashed a massive boom cycle. Already in 2015, when Google’s artificial intelligence unit DeepMind won its first game of Go against a human player, interest in the technology spiked globally. Similar cycles had happened before, with AI springs leading to AI winters every time investors felt the technology was not ready for prime time or when exuberant statements led to general skepticism about what AI can do. This time looks different. At least for now. What began in the 1960s, when research was largely academic and funding reached just a few million US dollars, has since led to annual capital investments of more than 250 billion US dollars – a figure that continues to rise. High hopes, hard work But overpromising is also a hallmark of the current boom. Some even fear that today’s AI models could become so powerful that, if they fell into the wrong hands or went rogue, they could spell the end of humanity. Both hopes – and fears – are high that AI will fundamentally change every aspect of our lives, including the pharmaceutical industry, with AI enthusiasts believing that cancer and other devastating diseases could soon be cured.

Research meeting dedicated to AI developments at the Novartis Campus in Cambridge.

Such high expectations are nothing new to Christian Diehl, Chief Data & Digital Officer at Novartis Biomedical Research. He has gone through many of the ups and downs of the AI industry himself and has learned to balance potential and capacity. When it comes to the pharmaceutical industry, he is fully aware of the promises and limits of AI. The previous boom cycle, which started 10 years ago, was also full of similar bold expectations, many of which are still distant dreams. But the momentum generated by the breakthroughs of Demis Hassabis, who later won the Nobel Prize in chemistry for his protein-modelling work at Google’s DeepMind culminating in the release of Alphafold, led to serious research across the board, eclipsing everything that came before. Novartis, which previously focused on telemedicine projects and digital medicine, started to build massive bespoke systems, including Sense Bridge, which allows monitoring of the company’s clinical trials, or data42, one of the industry's most ambitious digital R&D projects to date. But none of these systems were set up lightly or without hiccups, which serves as a reminder to Diehl that many exaggerated claims of what AI can do gloss too easily over the difficult problems the pharmaceutical industry is actually trying to solve.

“Today, some market players seek to reduce drug discovery and development to an engineering problem that we can solve through computational power, essentially saying: Give me a target and I will find a compound that will bind to that target,” Diehl said. A universe of possibilities Still, finding molecules with the help of AI can make a big difference since AI vastly increases the solution space within which scientists can operate – essentially allowing them to tinker with infinity. The new field carved out by AI is in fact so vast that Diehl is using a metaphor to describe its size: “Today’s science operates in a space that looks like a glass of water compared to the universe, which is the space within which AI can operate,” he said. Besides adding new possibilities, AI is also set to speed up the discovery process and free up researcher capacity. “If chemists don’t need two but only one year to do a lead optimization project, that adds 50 percent capacity in the system. So, we can have more shots on target,” he said. But while there is no match to computation when it comes to AI, making sense of the hypotheses requires – first and foremost – human judgement. “AI can generate millions of scientific hypotheses, but which of them makes sense?” Diehl asked. “We need the human judgment. And we can only have this if we build up decades of scientific understanding. Otherwise, it’s worthless, just a hypothesis without any link to a biomechanistic insight or human pathology,” he added.

The Pocket Crafter has provided researchers with new possibilities.

Human factor Diehl also points to the fact that the millions of compounds that are known and synthesized today have only led to the development of some 3,000 innovative drugs, with an average of approx. 40 to 50 drug approvals per year, meaning that only a small portion of the possible molecular structures may have medical value. “The challenge is not only to create a molecule as a drug candidate. Researchers also need to ask themselves whether they have the right target or whether the compound really affects a disease. Do we know enough about disease biology to claim that AI is going to solve it?” he asked. Consequently, the work of biologists, pathologists, chemists, biochemists, physicists along with that of data scientists and many others will become increasingly important as AI itself will only be useful if its proposals are understood in its complex biochemical context. Also, even having an AI-generated molecule is not the end of the story. “Discovering a molecule is only one small piece of the entire R&D value chain,” Diehl said. Making a successful medicine includes myriads of steps, most of which cannot be accurately modelled by AI yet. The untapped potential and empowerment For Diehl, the potential of AI is summarized in his question: “Can we organize all of our data? And can we pair it with decades of biomedical knowledge and scientific understanding? If we achieve this, we are one step closer to unlock AI in the complex drug development value chain,” he said. Efforts to reach this ambitious goal have been going on for years and are continuing with increased speed. Besides existing big systems such as data42 or Sense Bridge, AI has been implemented in almost all R&D steps and is delivering incremental improvements across the board. This has not happened overnight. Novartis has been steadily investing into machine learning and other top-notch digital technologies for more than 15 years and has always sought to be at the forefront of digital progress. Today, Diehl is working with a team of around 450 data and digital associates, who work closely together with both external partners and around 400 data scientists across the rest of the organization in Biomedical Research.

Their work is palpable across the board. In early drug discovery, researchers have identified promising new targets in areas such as neuroscience. Other efforts include increasingly complex collaborations between functions that did not readily interact in the past. In one such collaboration, data scientists reached out to the Novartis cell therapy facilities to gather manufacturing parameters with the goal to build a forecasting algorithm that can predict patient response. “The project, called Predi-CART, worked surprisingly well. No one in the past would have been able to combine manufacturing data with clinical outcome data, and build an algorithm on top to predict it. But that’s what systems like data42 now make possible,” Diehl said. Furthermore, AI is also empowering researchers from different fields to explore new avenues, as testified by a recent hackathon, where a team of non-specialists came up with a novel idea for disease identification.

Analyzing cell samples has been massively accelerated through AI.

At a recent hackathon aimed at finding new hypotheses for multiple sclerosis biomarkers, a wide number of associates from across Novartis participated in upskilling themselves with AI. “The group that eventually won the hackathon was a finance data science team which, by thinking out of the box, pointed us to new hypotheses that we are now testing,” Diehl said, adding that “when you pair AI and human ingenuity you can truly unearth hidden biological relationships that so far have escaped us.” Next wins Compared to projects such as “chip on a pill,” which were pursued more than a decade ago and reached a dead end, today’s digital systems are not only much more powerful but also have massive potential – even if they are unlikely to provide a panacea for all medical challenges. The real test, for Diehl, will not be the models themselves as a stand-alone component, but the interactions with humans. Likewise, Diehl says that remit and goal setting will be instrumental in creating a strong scientific basis before tackling the next steps. As a result, rather than going for a moonshot, his initial target may sound less glamorous: “My dream would be for AI to help us to eliminate failures caused by safety issues in our development pipelines,” he said. According to Diehl, more lofty goals may be still too far out as human biology is not understood in all its complexity. Basic research will have to unveil many more secrets before those findings can be modelled into an AI system. Thus, despite all advances, which are accelerating across the board, human ingenuity will remain of paramount importance. What has expanded is the space in which human ingenuity can express itself. For Diehl, AI is a tool. But a tool that brings Novartis even closer to its purpose: to improve and extend people’s lives and change the practice of medicine.

History of Novartis New horizons. An infinite digital space for human ingenuity Text Goran Mijuk. Photos Adriano Biondo. Once a niche sector in which telemedicine and ‘chips on a pill’ were considered the pinnacle of digital medicine, data and digital tools have taken over the healthcare industry by storm in less than a decade. But all recent advances will only bear fruit if artificial intelligence and other top-notch technologies are paired with scientific insights from researchers, says Christian Diehl, who leads the data and digital efforts of Novartis in Biomedical Research.

“We need the human judgment. And we can only have this if we build up decades of scientific understanding.” Christian Diehl

Data scientists in Basel.