AI Is Already Transforming Oncology Drug Design and Clinical Trials, AstraZeneca's Reis-Filho Tells AACR
核心洞察
AI transformation in oncology drug design is no longer theoretical but actively reshaping target discovery, molecular design, biomarker development, and clinical trials.
Fit-for-purpose evaluations, benchmarks, and intentional inclusion of negative data are critical to avoid "AI hypothesis slop" and ensure model quality.
AI can help reduce "human slop" — cognitive shortcuts and fragmented reasoning — raising the standard of scientific thought and decision-making.
The debate over whether artificial intelligence will reshape oncology drug development is over — the transformation is already underway. That was the central message delivered by Jorge Reis-Filho, Chief of AI for Science Innovation in the Enterprise AI Unit at AstraZeneca, during the American Association for Cancer (搜索) Research (AACR) plenary session "AI in Oncology: Facts and Hopes."
Speaking alongside Benjamin Haibe-Kains and in a session chaired by Sarah Skerratt, Reis-Filho crystallized five conclusions that reflect both the current state and future trajectory of AI in oncology R&D.
AI's Footprint Across the Drug Development Continuum
Reis-Filho emphasized that AI is no longer a speculative tool but an active force across the entire drug development pipeline. "We have moved beyond debating whether AI will transform how medicines are designed and developed. The transformation is happening now," he stated, citing firsthand evidence across target discovery, molecular design, biomarker development, patient selection, and clinical trials.
This integration spans from early discovery through late-stage development, with AI models increasingly informing decisions that were once the exclusive domain of human expert judgment.
The Imperative of Rigorous Evaluation
A key caution raised by Reis-Filho concerns the quality of AI-driven insights. He argued that the success of this transformation hinges on "fit-for-purpose evals and benchmarks" and the intentional generation and inclusion of negative data — experiments designed to teach models where hypotheses fail.
Without this discipline, he warned, "greater scale may simply produce more plausible-sounding, seemingly sophisticated but weakly grounded 'AI hypothesis slop.'" This call for methodological rigor underscores the risk that poorly validated models could generate compelling but ultimately incorrect biological hypotheses.
Confronting 'Human Slop'
In a complementary insight attributed to Benjamin Haibe-Kains during the Q&A, Reis-Filho noted that AI can also help reduce what was termed "human slop" — the cognitive shortcuts, fragmented synthesis of evidence, inherited assumptions, and inconsistencies that constrain scientific reasoning. "Used well, AI should raise the standard of thought and expose its weaknesses before they harden into decision-making errors," Reis-Filho said.
Educating the Next Generation
Reis-Filho challenged the scientific community to reimagine how future scientists are trained. He posed the question: "What if we taught them to think through AI — how to interrogate models, challenge outputs, design discriminating experiments, and combine deep domain expertise with computational reasoning while understanding the limits of current models and agents?"
He noted that the models and agents available today "will be the worst ones we will ever have," making the case that building AI literacy now is essential preparation for an increasingly AI-enabled scientific future.
Data as the Decisive Advantage
The fifth conclusion centered on data access. "Access to comprehensive and deep translational data will determine the next wave of biological and clinical insights," Reis-Filho said, specifying that this includes positive and negative data, longitudinal and multimodal data, and experimental feedback loops that allow models to learn from reality.
When these elements align, he argued, AI will not only accelerate insight generation but enable entirely novel discoveries — helping scientists "ask better questions, design more discriminating experiments, and deliver more meaningful advances for patients."
Reimagining Clinical Trials Through Patient Intelligence
In a separate discussion with Hannah Amies, Co-founder and CEO of Luvida (搜索), Reis-Filho explored how AI could fundamentally reimagine clinical trials by incorporating behavioral, lifestyle, socioeconomic, and environmental factors into trial design.
"Trial design and operational decisions have traditionally centered on biomedical and clinical data," he noted. "What if behavioral, lifestyle, socioeconomic and environmental factors could be incorporated in a safe and responsible manner to improve recruitment, consent, retention and adherence?"
He suggested that AI can help clinical teams anticipate recruitment, retention, adherence, and protocol risks before a study begins, "when the opportunity to mitigate them is greatest." This proactive approach could address long-standing barriers that have restricted trial participation for decades.
Reis-Filho also highlighted the potential for AI to confront biases in trial access, creating "opportunities for broader and more representative patient populations to participate in, and potentially benefit from, clinical research." Realizing this promise, he cautioned, will require fit-for-purpose data, transparent governance, appropriate privacy protections, and rigorous benchmarks that explicitly test performance across the populations AI seeks to serve.
He stressed the importance of multidisciplinary collaboration, involving "AI scientists with a deep understanding of the social and socioeconomic determinants of trial participation and success, working closely with clinicians, trialists, epidemiologists, behavioral and social scientists and, most importantly, patient communities."
The overarching vision, Reis-Filho concluded, is an "AI-enabled clinical development ecosystem with patients at its very center" — one that advances AstraZeneca's stated ambition in oncology: "to eliminate cancer (搜索) as a cause of death."