AI in Drug Discovery: A Dystopian Dream and an Immediate Reality

Advances in computer science have had a profound, largely positive impact on our personal lives – helping us to sustain friendships, cultivate new ideas and more expertly explore our world. More recently, advances in data science such as artificial intelligence (AI) have augmented society’s new digital reality, empowering decision-making in nearly every facet of consumer life – except healthcare.

We find ourselves in a pregnant moment, then, where the anticipation of material impact of AI on human health is accentuated by its certain reality but its uncertain ultimate manifestations. While it’s increasingly clear that disruptive technologies from Silicon Valley will radically change the way that we make and test medicines, biopharma and healthcare need to level up in digital.

Complicating matters, the acute distortion of the AI innovation-time continuum has created impressive, distracting hype. We are contacted at least weekly by curious investors and aspirational innovators boldly claiming to have unlocked secrets of disease biology with machine learning algorithms and public-domain datasets. I do not doubt that this can be true, though diligence seldom identifies exemplary, transformative use cases. Public datasets derived from published literature have limited prototypical forays largely to hypothesis generation, often with only a suggestion of causality.

At NIBR, we are interested in any insight leading to an atomic resolution of disease, but find ourselves not lacking for hypotheses as we are thirsty for new solutions to historical challenges in ligand discovery and radical advances in the efficiency of drug development. 

So how do we break the hype cycle and make AI relevant for drug discovery and development?

Augmented reality

Having grown up on “Fahrenheit 451” and “Terminator,” I admittedly love the idea of leading a dystopian research laboratory through VR gloves/goggles, commanding a legion of cyborg chemists through definitive lead optimization. Until this AI judgement day, we imagine a near-term augmented reality for our scientists where the decades-long experience of iterative Novartis chemistry is at every fume hood, guiding a more efficient path to development candidates.

Like their predecessors rifling through print copies of “Cell” for first insights years ago, today’s disease biologists are powerfully equipped with a readily accessible and digitized cortex of scholarship (literature), invention (patents) and data (GEO, et al.). So we work towards the assembly of a knowledge graph to augment target discovery and validation. We’re hoping to achieve a seamless integration of advanced analytics with bench science.

Indeed, the 21st century drug hunter visualizes disease biology not through the narrow lens of a microscope, but through multi-parametric descriptors integrated and analyzed today on the command line and tomorrow invisibly behind facile user interfaces. 

Catalysis

Another prevalent fantasy of AI in drug discovery imagines a Tolkien-esque ideal where One Ring rules them all, perhaps inspired by the market dominance of singular companies in personal computing or social networking. Perhaps someday Sauron will forge the definitive computational connectivity between molecule, target and disease state (the holy trinity of drug-hunting) as a uniform search bar or software package. But we are skeptical, having seen so many diverse, enabling, even disruptive technologies fertilize the entirety of the healthcare ecosystem. Rather, we see AI as a new power tool in the drug hunter’s toolbox, alongside monoclonal antibodies, CADD, AAV gene therapy, and, more recently, CRISPR. Like computational palladium, digital can be catalytic – accelerating the rate of innovation at each step in the convergent synthesis of new therapeutic entities.

We are therefore organizing at NIBR around the local innovation of AI catalysts that lower activation energy and improve the precision of technology transitions. From the more unambiguous measurement of emergent clinical signs of mobility, to a computational framework that matches chemical identity to gene-based descriptors of mechanism, to distributive clinical trials that bring trials to patients (not patients to trials), there are so many opportunities to expedite the delivery of definitive medicines to patients, as they rightly expect.

Mutation

Fundamental innovation in AI is flourishing in academia and the private sector, fueled by significant consumer, security and financial applications. The lowest-hanging fruit in biomedicine is to download – to identify and implement algorithms of relevance to the science of therapeutics.

There is indeed a ton of low-hanging fruit. We can leverage, for example, deep neural networks to learn from vast stores of image-based data. And the same real-time, interpretive analysis of pedestrians used to guide Tesla trajectories can help us understand the progression of acute macular degeneration, quantify fatigue attributable to multiple sclerosis, and measure the behavioral manifestations of genetic autism disorders in preclinical models. Who would have thought that an app used to find a cat on the internet would help us find first medicines for severe neurologic diseases?

But biopharma should not be so complacent. For years, biopharma exemplified this narrative of applied science in the natural sciences, where innovation in the academy prompted derivative applied research in industry.

Today, I see real parity among leading institutions in academia and industry with respect to fundamental discovery research, enabling technology and experimental methodology. We should therefore not repeat this history or download for too long. Much of drug hunting is not processive or image-based, but rather artisanal and guided by quantum mechanics. So we are beginning to craft the bespoke computational framework befitting the science of therapeutics. 

A human resource

Science is not a molecules, genes and data business; it is ever more a people business. As not all of our associates will have prior coding experience, we have recruited and cultivated a generation of data scientists to work shoulder-to-shoulder with project teams at all stages of discovery and early drug development.

More than 250 data scientists have joined our ranks. (We once humbly wondered if we could attract top computational talent away from the dot-com world of unicorn hunting.) Indeed, most arrive at NIBR following domain-agnostic training or productive careers in the retail and marketing world, now compelled to contribute their computational gifts to humanity through the innovation of breakthrough medicines for life-threatening diseases. Unsurprisingly, their download speed on human biology and cheminformatics is ripping fast, but they have also uploaded to our organization a firm cultural adaptation to the new digital reality – a sense of curiosity and wonder.

Humility

Having led a laboratory effort that leaned heavily on large, genome-wide datasets to understand drugs that affect cellular memory in cancer (integrative epigenomic analysis of BET bromodomain inhibitors), I have seen the power of data science in service to drug discovery. Measures of genome structure in cancer models and patient samples led to previously unrecognized mechanistic insights into cancer biology and tumor dependencies. These ultimately guided the path for drug development.

I have also seen how hard it is to create, debug, deploy, share, service and sustain the suite of novel applications required to integrate disparate datasets and answer specific questions of human biology. Even from my role here at NIBR, one of the largest and most impactful biomedical research institutes in the world, it is humbling to witness the pace of AI and quantum computing in other industries.

So we are approaching this work with total humility and openness to collaboration, and we think we are a pretty great partner for innovators in this space. We have a critical mission for humanity, an industry-leading record of productive shoulder-to-shoulder innovation, science at the clinical interface where secondary endpoints are easy to add, and tons of data. Indeed, our data scientists presently leverage more than 50 petabytes of discovery data, more than 1 billion unique compounds, and more than 1 million highly-structured patient records from prospective studies. Please challenge us with your important ideas.

I recently took the podium at BIO 2018 to talk about these and other opportunities to advance data science in therapeutics science. There is so much opportunity. My best advice is to learn to code or find a friend who can code, to hack drug discovery with cutting edge AI algorithms, and to organize around the most pressing challenges facing patients today – all the while reimagining medicine in this brave new digital world.