AI is promising to change that. But behind the buzzwords lies a more nuanced shift: one that depends as much on data quality, team culture, and clinical execution as it does on algorithms.

For Najat Khan, chief R&D officer and chief commercial officer at Recursion, the answer starts with a shift in mindset. "We all agree that we have to do a better job of making medicines," she says. "In order to do that, you want to ensure that you're using AI and automation with a clear purpose in mind. It's about making differentiated medicines – and working backwards from there."

But even with cutting-edge models and robotic labs, the road to reinvention is far from frictionless. Beneath the surface lies a deeper set of challenges: data quality, cultural barriers, regulatory hurdles, and the fundamental unpredictability of cancer itself.

"The way you cover gaps is to have holistic data sets, not just bespoke snapshots, and that can only be done in an automated way."

But as Khan notes, it's not just about finding interesting biology – it's about finding biology that matters, what she calls "the unobvious edge". Khan points to an example in oncology: CDK12, a target long known to play a role in DNA damage response, but difficult to drug due to its similarity to CDK13. Using AI-generated biological maps, Recursion's team identified a previously overlooked target – RBM39, involved in splicing fidelity. Within 18 months, they'd designed a compound targeting it and advanced the programme into early clinical development. That's less than half the time such a process would typically take.

The approach is deliberately end-driven. By anchoring drug discovery around the patient outcome – whether that's a first-in-class target or a therapy for an unmet need – AI is not just changing how drugs are developed; it's forcing a re-evaluation of where drug discovery begins.

This loop enables faster, smarter decisions. "AI lets you continuously learn from your data – adjusting your models, refining your hypotheses, and narrowing your uncertainty with every cycle," says Khan. It's a shift from static experiments to dynamic learning systems, where insights evolve in real-time.

AI's impact on the clinical side of drug development may be even more immediate. According to Kerri Cali, senior director and head of clinical operations at Recursion, predictive modelling is already improving how trials are planned and executed. Algorithms can now forecast enrolment rates, identify optimal trial sites, and flag underrepresented patient populations – improving both efficiency and equity.

"Patient heterogeneity is a well-known phenomenon," says Hemanth Kanakamedala, Recursion's vice president of decision science and real-world evidence. "The ability to quantify this heterogeneity has been the biggest challenge for our industry. Given certain genomic and clinical features of a patient, what is the probability that they'd respond to an experimental therapy, or a combination therapy? That is the key question."

In this context, AI is being used not only to improve site selection and streamline recruitment logistics, but also to simulate treatment outcomes in silico, before any patient is enrolled.

Kanakamedala adds. "The ability to model the genomic attributes of patients with various tumours and their outcomes allows us to build a robust understanding of the biology of the patients. We're able to simulate whether modulating a certain target would result in a high magnitude of effect."

Still, while the promise of AI in oncology is exciting, the technical limitations are real. "One of the biggest gaps is between AI's potential to accelerate patient matching and optimise trial design versus the fragmented reality of clinical data systems. AI tools excel at pattern recognition when fed clean, structured datasets," Cali explains, "but clinical reality involves unstructured records – siloed across EMRs, labs, imaging systems, and genomic platforms that don't speak to each other, alongside inconsistent biomarker reporting, and lack of harmonized data formats across the healthcare landscape."

The positive news, as Cali explains, is that efforts are advancing to close these gaps, with emerging AI applications showing promise in targeted use cases, addressing the inherent complexity of real-world datasets.

"While progress is being made in wider strides and at an accelerated pace, the reality is that, until we solve these fundamental data infrastructure problems more holistically across the industry and create truly interoperable clinical data ecosystems, AI will continue making incremental improvements, rather than delivering the transformational acceleration needed to bring effective treatments to patients faster."

"People whose native language might be chemistry or computer science – but who understand enough of the other side to collaborate effectively."

She points to Boltz-2, a new biomolecular foundation model co-developed with MIT, as a key example of how this bilingual approach works in practice. Unlike earlier tools focused only on structure prediction, Khan explains that Boltz-2 adds a critical dimension: binding affinity. For medicinal chemists, that's a game-changer, adding one of the most critical elements of drug discovery and development to their virtual toolkit. And because the model includes a "controllability" feature, scientists can fine-tune parameters – guiding the model with their own knowledge of what makes a molecule viable.

That sense of co-ownership is central to shifting mindsets. "Instead of us versus them," Khan says. "It's about supercharging what we can do today to give it the best chance of success." But she's also clear-eyed about the operational realities. Creating usable datasets, shaping regulatory conversations, and scaling infrastructure all take time. "I would love to have a corpus of data that exists like PDB [Protein Data Bank], which helped, but there's not a lot of those clean, reproducible data sets that exist in the public domain. But it takes time to build the base before you can actually do the analysis, the insights, and the outcome," she explains.

For AI to truly accelerate oncology drug development, the entire ecosystem needs to evolve in parallel: not just algorithms and datasets, but people, processes, and regulatory frameworks. As more leaders across pharma begin to focus on usability and trust, the field may finally start to move beyond proof-of-concept – and toward broader adoption.

Still, for Khan, the reward is worth the complexity. As AI begins to reveal patterns in resistance, predict patient outcomes, and speed up cycle times, oncology drug development is inching toward something long overdue: a process that works not just faster, but smarter. Citing one of her favourite quotes from former US President Teddy Roosevelt, she concludes with a call to the industry: "There's no option for us not to try."