The Discovery Process: AI as Investigative Tool

What is being discovered is changing. How biomedical discovery is itself instantiated is changing faster, and the second of the two is the more consequential.

Traditional hypothesis-driven, sequential, and relatively low-throughput experimental cycles are giving way to AI-augmented, highly iterative, multi-scale, and increasingly automated approaches that compress timelines, expand the space of testable ideas, and integrate data across levels of biological organization.

• Protein structure and interaction prediction at scale, which has turned structural biology from a bottleneck into a rapid computational step guiding experimental design and drug discovery.
• Foundation models for cellular state and gene regulation, which learn the grammar of cell behavior across millions of cells and permit in silico perturbation before any bench work begins.
• Deep research agents and mechanistic co-investigation, positioning AI as scientific co-investigator, exploring literature and knowledge graphs, proposing testable hypotheses, and compressing months of reasoning into minutes.
• Closed-loop AI and automated robotic laboratories, in which hypothesis generation, execution, and interpretation cycle without human intervention, with explicit guardrails against confirmation bias.
• Biological digital twins and virtual cells, allowing interventions to be rehearsed computationally against a model of the specific system before being attempted in it.
• Generative design of biological systems, in which proteins, circuits, and organisms are specified by desired function.
• High-content patient-derived systems integrated with AI analytics: organoids, xenografts, and organ-on-chip platforms read out at scale, turning the individual patient into an experimental system.
• Mechanism-of-action triangulation, which approaches a disease backwards from the several unrelated exposures that produce it.

Taken together, the eight change the physician's relationship to discovery and not merely its pace. What they absorb is the part that used to require an institution: the apparatus, the cohort, the core facility, the years. What they cannot supply is the question, and the barrier to entry shifts accordingly, from access to instruments toward access to data and the capacity to frame a question worth putting to it. Both of those are things a practicing physician holds in unusual measure. This is what makes discovery democratizable in the sense Chapter VII argues, and why the reconfiguration proposed there is a matter of permission and infrastructure. Talent is not the constraint.

With respect to the last of these, uveitis is the clearest example: a possible sequela of long COVID, and triggered as well by Ebola, chikungunya, dengue, and reactive arthritis. The initiating mechanisms are drastically different, which is what makes the convergence informative, since whatever the shared endpoint requires must lie downstream of all of them. Data and systems now exist to track such loops at scale, and triangulating across them promises a deeper mechanistic account than any single-exposure study reaches, together with considerably better identification of drug targets.

The example reflects a shift of orientation. Discovery has organized itself around the patient and the disease entity. It is now reorganizing around mechanisms and systems, so that the informative unit becomes a pattern running across diseases, pathways, and interventions instead of any single condition. The several exposures therefore matter more as a set than as cases, and the shared endpoint serves as an instrument.

The clearest near-term demonstration of purposing is the rediscovery of an old molecule rather than the design of a new one. Biovista, founded by Andreas Persidis and led with his brother Aris Persidis, works from the observation that the pharmacopeia already contains compounds whose mechanisms would serve diseases they were never developed for. Its Project Prodigy platform reasons across clinical, preclinical, pharmacological, and toxicological evidence from the published literature to generate mechanistic hypotheses connecting an existing compound to a new indication, with the chain of evidence attached.

Aris Persidis has made a methodological argument: much machine learning is built to extract patterns already present in training data, whereas biomedical discovery requires models built to surface relationships that are not yet known to anyone. The company has applied the approach across neurodegenerative disease, epilepsy, oncology, and orphan conditions, and used it to nominate candidates during the COVID-19 pandemic. A purposed compound arrives with its safety profile already established, which collapses the translational gap the next section describes.

Not all computational discovery is AI, and the current enthusiasm merges two things that need separating. The emerging discipline of computational medicine is broader than machine learning, and some of its most consequential results have come from pattern discernment applied to the genome, with no model trained on labeled examples involved.

The work of Isidore Rigoutsos, who co-founded IBM's computational biology program at the Thomas J. Watson Research Center and now directs the Computational Medicine Center at Thomas Jefferson University, is the compelling proof point. Beginning from the architecture of the nuclear and mitochondrial genomes and the repetitive elements that riddle them, his laboratory identified entirely new classes of short non-coding RNAs where the field had assumed there was nothing to find: the isoforms of microRNAs known as isomiRs; fragments derived from transfer RNAs; and fragments derived from ribosomal RNAs.

These were computational predictions before they were biological entities. They have since been shown to exist, to be constitutively expressed, and to vary with tissue, cell type, sex, ancestry, and disease, and they have proven informative enough that the presence or absence of small RNA isoforms alone can discriminate among more than thirty cancer types. Tools and databases built by the group are now used across the field. The frontier belongs to those who can look at a genome and notice that its architecture is telling them something, as much as to those who train models.

An explicit statement of where this leads comes from Hiroaki Kitano, president of the Systems Biology Institute in Tokyo and director of the Laboratory for Disease Systems Modeling at the RIKEN Center for Integrative Medical Sciences. His Nobel Turing Challenge proposes the development of AI systems capable of conducting top-level science with high autonomy, making discoveries of a quality indistinguishable from the best human work and, in principle, worthy of the highest recognition.

His reasoning is specific to biomedicine. Two decades of systems biology produced an overflow of data and knowledge beyond what human cognition can hold. He concluded that the next major advance would require AI-driven discovery, not AI-assisted analysis, introduced first as assistance and resolving over time into genuine autonomy. Whether or not one accepts the terminus, the diagnosis is difficult to dispute, and it reframes the physician's question from whether machines will participate in discovery to what remains distinctively human within it.

A system that generates and tests hypotheses answers questions; it does not determine which questions are worth answering, and that determination is not a technical act. It requires knowing what would actually change a decision at a bedside, which anomalies are worth pursuing and which are noise, and what the field has stopped asking because it stopped noticing.

Those judgments come from proximity to patients, not proximity to data. That is why the physician's place in the discovery system is generative and not residual: the machine becomes the more capable investigator, and the clinic remains the more productive source of investigable questions. Chapter VII takes up question-asking as a competency to be trained, and Chapter IX argues that curiosity, deliberately exercised, becomes the scarcest act in medicine.

A caution comes from the institutional side. Otmar Wiestler, who led the German Cancer Research Center and then the Helmholtz Association before joining the Champalimaud Foundation as Director of Champalimaud Science, has spent a career on the organizational rather than the technical half of this problem: how large multidisciplinary research environments are built and coordinated so that discovery actually crosses into clinical use.

His move to the Champalimaud Centre for the Unknown, an institution founded on the deliberate integration of laboratory and clinic, was framed as strengthening the connections among scientific programs, clinical activity, and education. The instruments described here will not by themselves produce translation. Translation requires organizations designed so that the people using the instruments sit close to the people treating the patients.