The Value of Prediction, and Cost of Success, in Drug Hunting
By Deborah Borfitz
March 26, 2026 | Across the pharmaceutical industry, generative artificial intelligence (AI) is now running an estimated 10 to 20 million predictions a day in a quest to accelerate drug discovery by exploring new molecules and chemical reactions at an unprecedented pace. Access to hardware determines AI capabilities, but the tech-driven process of making medicines is still “done by people for people,” with a human in the loop calling the shots and running the machinery, according to digital drug hunter Tudor Oprea, M.D., Ph.D., CEO of Expert Systems, Inc.
What no one seems to be asking is what a prediction’s worth, because the machine learning (ML) models being deployed aren’t free, he says. It costs a lot of money to collect the large amount of data necessary to produce accurate, reliable, and generalized results, and it takes skilled labor to properly run the experiments.
And then there’s the high cost of the GPUs, which limits creation of on-premises data centers to the big-pharma companies, allowing them to “jump ahead of the pack” like the lead cyclists ahead of the peloton in a road bike race, Oprea continues. “Clearly people who use AI are going to win, as opposed to those who don’t.”
AI in drug development nonetheless remains more an aspiration than a reality, he says, simply because of the complexity of human biology. Scientists rely heavily on data from cells that do not fully replicate the complex interactions of the immune system, for example. When looking at brain activity, for instance, they’ll often run cell tests in neurons without considering the surrounding matrix of white matter populated with astrocytes and glia that provide additional support and function and release molecules that influence neuronal activity.
Currently popular organoid models can help bridge that gap, says Oprea, but generally don’t fully recapitulate the complexity of the blood-brain barrier—the primary obstacle to treating central nervous system (CNS) disorders. Cardiovascular studies in mice and rats also have a blind spot related to the fact that their vagal response, the primary driver of the parasympathetic nervous system (rest and digest) is very different compared to humans and provides less counterbalance to the sympathetic nervous system (fight or flight).
“If I study heart drugs in mice ... it’s just not applicable to humans,” Oprea says. As the old joke goes, “[rodents] have been cured of all diseases you can think of and yet humans still have them.”
The possibilities and limitations of AI and ML in the world of drug discovery, including the rarely discussed economic aspects, will be the topic of a presentation Oprea will be making next month [CHANGE IF THIS RUNS IN APRIL] at the Drug Discovery Chemistry conference in San Diego. As he will share, active learning allows ML models to be more right than wrong over time while coupling them with agentic AI gives companies with the most GPUs the higher probability of success.
ML vs AI
It’s important to understand the key differences between ML and AI, says Oprea. Machine learning models are task-specific and limited by their training data, and the output of the large language models (LLMs) dominating the landscape hallucinate and can’t be fully trusted, although “reasoning” LLMs that specialize in solving complex, multi-step problems perform better.
Artificial intelligence is more broad-purpose and human-like and leverages a lot of human intelligence, he says, and newer generative AI models like Gemini 3 Pro are exceptionally good at extracting knowledge from patents. While AI struggles less than ML with basic chemistry tasks, it comes with a cognitive cost on humans in terms of their memory and critical thinking.
Humans face considerable risks unless they learn to grow and evolve in parallel with AI by employing their brainpower and learning new skills, says Oprea. “Intelligence is one of the key features of evolution that is probably less discussed, and probably not politically correct to highlight, but ... it is what distinguishes us from animals.”
For Oprea, over-reliance on AI at the cost of human intelligence conjures up the story world from the 2008 Pixar movie WALL-E where robots are doing everything for spaceship-dwelling humans. “That’s a lucky scenario,” he adds. “The unlucky scenario is either that the robots take over or that the planet is just covered in dust with no humans.”
This is not an inevitability. His advice to pharma executives is to invest equally in the technology and people who are willing to use it in a creative way.
Big-Pharma Advantage
The pharma companies with a massive uptick in AI usage, and the ability to extraordinarily accelerate drug discovery, are all well-known names in the industry, says Oprea. Three years ago, Sanofi declared itself an AI-first pharmaceutical company, an initiative that includes developing and deploying an agentic AI app called "plai" at scale.
Eli Lilly more recently announced the installation of an AI supercomputer named LillyPod that is powered by over 1,000 of NVIDIA’s latest-generation Blackwell Ultra GPUs. Eli Lilly and NVIDIA are also making a $1 billion investment to create a co-innovation lab focused on generative AI for drug discovery. This was followed by Roche’s announced expansion of its three-year strategic collaboration with NVIDIA to create a large-scale AI factory with a new investment of over 2,000 Blackwell GPUs.
“It is no longer a question of how many people can grow 10,000 mice or 1,000 primates to run their experiments,” he says. “It’s a question of how speedy is the trial that you conducted in silico, before you actually run an ex silico experiment, [and] the ability to run thousands of hypotheses at the speed of thought ... as opposed to waiting for an experiment to confirm or not confirm.”
This assumes that the companies deploying these models benefit from correct predictions, which is a safe bet given the investment trend, continues Oprea. With relative ease, a big-pharma company might use AI to design a new and improved drug (e.g., a statin) that works particularly well in a certain population (e.g., the elderly).
Smaller, cash-strapped companies potentially have something to lose here. Larger companies may want to fill gaps in their portfolio via acquisitions, provided they’re not competing on similar products. If they are, the lesser-resourced players “eventually will not thrive and basically go away,” Oprea predicts.
That risk poses the question about how much innovation is being fueled entirely by AI, and the philosophical discussion around whether AI systems can be named as inventors or patent holders on their own. Oprea’s view, which is backed by current legal and ethical frameworks, is that there should always be a human in the loop contributing to the idea-making to ensure accountability, safety, and oversight.
Nailing the Costs
In a forthcoming “What’s a Prediction Worth?” position paper, Oprea and his colleagues will be discussing the economics of ML. The exercise harkens back to the mid-1990s when he was a computational chemist at AstraZeneca, developing models and running predictions and getting it wrong wasn’t an option. “There was never any parity with how many shots on goal you get ... versus a medicinal chemist.”
A lot has changed since then, Oprea notes. Companies generally trust and act on predictions generated by predictive machine learning.
AstraZeneca has a program that Oprea helped set up in the 2000s, then called C-Lab for computational DMPK (drug metabolism and pharmacokinetics) and laboratory sciences and since rebranded PIP (Predictive Insights Platform). According to a 2024 paper, PIP was then making roughly a million predictions a day for the company (Drug Discovery Today, DOI: 10.1016/j.drudis.2024.103945).
Oprea says he is “trying to build the argument that any prediction that saves you running an experiment because you trust the prediction actually has an intrinsic economic value.” His position is that “a trustworthy prediction should be worth at least 5% of the actual experiment.”
Three types of experiments are being examined: physical chemistry, the example being water solubility; in vitro tests, using MBCK (Madin-Darby Canine Kidney) permeability; and an in vivo model, focused on the fraction of a compound in the brain using a modern metric for evaluating CNS drug penetration. The estimated per-compound cost of doing those experiments is, respectively, $60 to $100, $350 to $500, and $1,500 to $5,000. A direct economic value for a predictive model can be calculated based on the lower number of compounds that consequently require those resources—but only after factoring in the cost of deploying the ML in the first place.
This gets back to the major ongoing operating costs for data collection and skilled labor, and the significant capital expenditure for the GPUs. “In principle, you could ask a grad student to plug data into a cheminformatics prediction system and make a model ... but at the end it is equally an art, not just science,” says Oprea.
The best predictive models don’t just have the highest R-squared, meaning how well they explain the variability of the data, but also the lowest uncertainty based on metrics measuring differences between predicted and observed outcomes and the number of errors over time, he says. “So, as part of this value of prediction ... we bring in active learning.”
Commonly termed “design of experiments” in the 1980s and 1990s, active learning refers to machine learning bots identifying compounds strikingly different from those in the current portfolio to enrich the models and make them more stable, Oprea explains. This is how AstraZeneca, over a 30-year period, came to have stable models running a million trusted predictions a day. “It’s all about ... how can I use machine learning to take the next step to make decisions.”
Tools Aplenty
In terms of AI and ML, generative chemistry used for the design of novel molecules with specific desirable qualities is a bright spot, says Oprea. But in what may come as a surprise to many people, the concept has existed since 1993 when David Weininger obtained a patent for what was then called “genetic algorithms” for generating molecules in silico from scratch.
What’s happening now, he says, is that generative chemistry is utilizing “fitness functions” to guide AI models in designing molecules with the necessary properties like high binding affinity, synthetic accessibility, or optimal solubility. From all that collected knowledge rules have evolved enough that the models can provide what chemists find agreeable.
Human input has been crucial to weed out the nonsense, including compounds that are potentially toxic or not synthesizable. A computer program must be taught what is reasonable and technically possible by people who know. Only then does it open “a tremendous opportunity to explore additional new chemical space,” says Oprea.
In addition to Gemini 3 Pro, custom-made applications for generative chemistry are being cobbled together by combining LLMs, he adds. These are frequently designed to function as autonomous agents.
AstraZeneca has an open-source generative AI tool for molecule design called REINVENT that couples recurrent neural networks and reinforcement learning rather than a general-purpose LLM, and the latest version also incorporates transformer-based models, enabling the generation of complex, valid structures.
A growing number of commercial and open-source generative chemistry tools exist for new molecular entity discovery to, for example, find the most optimum brain-penetrant molecule with a half-life of 30 minutes so it might be developed into a drug useful in the operating theater. A big piece of the work for humans in the loop is to come up with the specific questions and parameters, based in part on existing knowledge about other drugs, to eventually get to something that could be taken to clinical trials.