From Pixels to Prescriptions: How AI Can Revive Drug Discovery & Development

For hundreds of years, the cold chain industry relied on little more than big blocks of ice to get food from one place to another without spoiling. That made speed and distance the measures of success in the marketplace. With the invention of refrigerated transport, however, producers suddenly had the ability to ship goods around the globe. What looked to consumers like a more convenient shopping experience—bananas on the shelves in January—occasioned a full-scale reshaping of the geopolitical landscape, as multinational conglomerates like the United Fruit Company emerged and controlled markets with an iron fist, inspiring the term “banana republic.”

A similar “curve jump” may be on the horizon for drug discovery. With its basic ability to find patterns in vast amounts of data and predict outcomes, artificial intelligence promises to transform the long and arduous process of discovering and developing new drugs. Much like the humble refrigerator, AI could transform how entire industries function, fulfilling the dream of personalized medicine while also raising difficult ethical questions.

In my 20+ years in this industry, with the support of incredibly talented team members, I’ve had the opportunity to deploy AI-based approaches across the Research & Development value chain (starting with that original idea, completing the preclinical validation work, regulatory requirements for testing in humans, onwards to clinical testing in patient populations, and beyond). I’ve worked on nine approved medicines that are currently helping hundreds of thousands of people around the globe, so I’ve seen firsthand what it takes to move a drug from laboratory idea to regulatory approval. AI stands to optimize the entire process. The risks, to be sure, are legion, but how do they compare to the benefits? And what is it about biopharmaceuticals that makes AI such ripe fruit for the plucking?

Is Drug Discovery Flatlining?

The journey for new medicines is long and circuitous. Over 90 percent fail in clinical trials, and those that survive may take up to 15 years before they hit pharmacies, costing roughly $1 billion along the way. Lack of reproducible efficacy, insufficient safety and/or tolerability, and process stagnation are common reasons for failure. There’s also the sheer sophistication of modern drug manufacturing: While a single molecule of aspirin contains just 21 atoms, a modern biopharmaceutical antibody could contain tens of thousands. The level of technical sophistication is astounding and requires hundreds if not thousands of experts to move the needle from idea to medicine. 

Given these contingencies, it’s easy to see why industry efficiency rates have flatlined. Speed to market, clinical success rates, and overall productivity have all languished in recent decades, even while R&D spending has grown by some measures more than 600% over 2 decades from $40 billion in 2001 to over $276 billion in 2021. R&D efficiency and efficacy has been the subject of much analysis and concerns over the years.  So much of the world’s health is impacted by medicines.  A recent Deloitte report blamed diminishing returns on a mix of factors: changing regulations, patent expiries, inflation, the complexity of R&D protocols, advancement in the standard of care requiring increasingly sophisticated solutions to outdo yesterday’s advances.  A hybrid set of solutions, some with sector-wide appeal, but also some specifically tailored to any one organization, or even business unit are in order.  In many ways, these challenges boil down to a simple mantra: The more we know, the more we realize how much we don’t know. 

From Data to Medicine

If ever a curve jump were needed, it is here. AI could be the heavy machinery needed to optimize and accelerate the discovery process. Deep learning tools, in particular, can parse huge datasets and accelerate actions at each phase—from target identification to lead discovery—while lowering overall computational costs.

Failure of a new drug candidate can occur anywhere along the drug discovery path, so for a technology to deliver a “curve jump” like the one that disrupted the cold chain industry, it needs to deliver on all fronts. It needs to generate higher returns in shorter timelines, increased throughput of treatments to patients, more optimized trial designs, and shorter overall innovation cycles.

That’s a tall order, but it’s worth noting that the very thing which underpins the challenges in drug discovery also underpins the tool most likely to solve them: data. The entire field of life science matured alongside the big data revolution. This is an industry defined by rich datasets—high dimensionality, multimodal, complex, non-linear. With such vast amounts of information comes the opportunity to uncover novel disease cellular and molecular endotypes, and to optimize precision medicine approaches by matching the right patient to the right treatment—the dream of personalized medicine.

Pain Points

As transformative as AI could be, it’s not a panacea. For example, multi-omics data—like those generated from the genome, transcriptome, or proteome—are often incomplete, not widely accessible, or sequestered in either the preclinical or clinical world. It’s also quite noisy; the training sets of biological activity data are often drawn from the literature and not always quality-controlled. Data collection is always going to be limited by the powers of observation, which are themselves limited by the tiny scales of biological systems. Moreover, there are challenges on the training and interpretation of AI models, and how they can be applied to different disease models or experimental predictions.

There are ethical concerns, too. AI’s very powers of prediction are also what make it so controversial. Certainly, patient information needs to be protected, and there are legitimate concerns around the sourcing of training data in an era of compromised privacy. AI algorithms are also notoriously unexplainable; they cannot tell you how they produced a given output and may conceal serious operational biases.

What’s Next?

AI may well do to the biopharmaceutical industry what the refrigerator did to the cold chain industry, but it’s worth remembering: The refrigerator didn’t replace ice. Ice is still used and relied upon in too many ways to count, and its obsolescence was never really a serious prospect. Can the same be said of humans in the age of AI? We sure think so.

Through these articles, we’ll explore this space to break through and reveal the many pain points of the R&D value chain. We may take detours into philosophy, questioning if nature is really written in numbers, if we predict the future, how much we know about life, and how much we need to know to discover novel medicines against all diseases.