But medicine still has its own “valley of death”: the gap between bench and bedside. Even after accounting for failure rates and the cost of capital — the risk-adjusted figure that captures what it actually takes to get one drug approved — clinical development is the largest line item. In oncology it accounts for 55.9% of expected capitalized cost, against 41.1% for nonclinical work. The clinic is where therapeutic hypotheses are translated from controlled lab environments into heterogeneous, real-world biology, and it is where most programs die.

That makes the clinic the right place to ask what AI can actually change. As Teslo has argued in her "clinic-in-the-loop" framing, going forward the clinic should perhaps be seen as a site of biological learning. The history of CAR-T illustrates the point: early clinical failures were not dead ends, but readouts that revealed why the therapy failed and how it should be redesigned. In this view, faster and more information-rich trials create a feedback loop between patients, models, and therapeutic design (*).

But learning from patients only matters if the rest of the system can respond. The clinic-in-the-loop idea therefore points to a full-stack therapeutic system. Such a therapeutic system would have at least four layers: read, design, build, and learn. It must read the patient's disease state, design an intervention against it, build the therapeutic, and learn from the clinical response efficiently enough to close the loop.

This should not be read as a replacement for standard medicine. The first applications of a full-stack therapeutic system are diseases where the framework reaches its limits: rare genetic diseases, end-stage cancers, refractory infections, and molecularly defined conditions that evolve faster than conventional development cycles.

Read: disease as molecular time series

The first layer of the stack is measurement. Genetics revolutionized oncology by classifying tumors through molecular drivers, such as HER2 or EGFR, enabling targeted therapies. But a single biopsy is a local snapshot of a moving target. Tumors exhibit pronounced intratumoral heterogeneity, adapting over time to therapeutic pressures and varying geographically within the same lesion.

Two complementary measurements push past the snapshot. Circulating tumor DNA (ctDNA) gives the temporal axis: malignant cells shed fragmented DNA into the bloodstream, and because that DNA has a short half-life, serial liquid biopsies act as a near-real-time readout of tumor state. This makes it possible to track clonal evolution and detect minimal residual disease. What ctDNA cannot do is tell you where in the body something is happening. Functional imaging techniques, such as targeted PET and multi-parametric MRI fill in the spatial axis, quantifying functional and metabolic variation across lesions in vivo.

The point is that the disease is no longer a category assigned at diagnosis. It is a dynamic trajectory in space-time that can be sampled.

Design: patient-tailored molecules

Once disease can be read as a dynamic molecular state, the next layer is specifying an intervention. Structure prediction helps translate dynamic molecular states into actionable information. With high-confidence structural models of targets and their conformational states, the design problem has been inverted. Rather than screening libraries against a target protein, we can specify the binding mode, geometry, and selectivity profile and ask whether a molecule can be built to satisfy that specification. Combined with advances in generative models, this allows for rapid iteration.

Recent demonstrations are striking in how aggressively they collapse screening into specification. RFpeptides, an extension of RFdiffusion to macrocyclic peptides, designs cyclic binders to a chosen epitope at a specified affinity. RNA is on the same trajectory: improved 3D structure prediction has begun to enable structure-guided generative design of aptamers. The relevant change is not only speed but interpretability: each design carries an explicit hypothesis about why it should bind, and what it should do. This allows every clinical observation to then be projected back onto the model that generated the molecule.

What this enables, ultimately, is a dynamic treatment plan rather than a singular treatment. The Read layer specifies a target, such as a particular driver mutation in a clonal population. The Design layer produces a candidate that fits that specification. As patient state evolves under selection pressure from therapy, the plan can evolve with it.

Build: closing the loop physically

None of this matters if the therapeutic cannot be made on the timescale of the disease. For sequence-defined medicines, the relevant primitive is fast, parallel synthesis of heterogeneous molecules. Such synthesis is not yet clinically possible in the full therapeutic sense. A deployable system would also need formulation, sterility, chain-of-identity, and regulatory oversight.

However, the technical substrate for such a build system is already emerging. Two advances in bioengineering are notable here, one in synthesis and another in QC. Twist Bioscience’s platform miniaturizes DNA synthesis chemistry onto a single CMOS chip, reducing reaction volumes by a factor of 1 million. This works by gating standard solid-phase phosphoramidite synthesis through a local electrochemical reaction.

At the same time, high-throughput assays are turning synthesis into a measurable process. Molecular barcodes separate writing errors from readout errors; shifted oligo libraries separate sequence-dependent failures from position-dependent synthesis bias. In a full-stack system, that error model should feed back into design, so the therapeutic sequence is optimized not only for activity, but for writability.

The build stack culminates in the transition from centralized mass production to decentralized, Point-of-Care (PoC) manufacturing, turning hospital pharmacies into software-defined foundries. By deploying validated, synthesis devices directly at the bedside, this model collapses the distribution time-lag and preserves biological fidelity. This shift is already codified by the UK MHRA and the Australian TGA’s framework for ATMP (Advanced Therapy Medicinal Product), which provide the legal blueprints for diffuse manufacturing.

Learn: navigating the clinic

The Read layer produced a patient state; the Design layer produced a candidate intervention based on that state; and the Build layer turned that intervention into a manufacturable therapeutic. The Learn layer has to take each clinical observation back as evidence about both “Read” and “Design”, that is the patient’s trajectory and the model that generated the molecule. One schema for doing so is navigational AI: systems that translate patient state and history into a recommended action for this patient. Although navigational AI is at a nascent stage, two methodological advances promise to make it tractable.

The first is causal inference, supported by methods such as double machine learning. These methods allow us to move beyond average treatment effects toward patient-specific estimates of benefit and harm while ensuring careful consideration of confounding factors. The second is reinforcement learning for sequential decision-making: a clinical trajectory is a series of state–action–outcome triples, which is exactly the structure RL is built for. Komorowski et al. trained an RL agent on ICU data to recommend vasopressor and IV fluid doses for sepsis, and patients whose actual treatment matched the learned policy had the lowest mortality.

From Molecules to Adaptive Systems

The transition from a linear bench-to-bedside pipeline to a closed-loop system represents an ontological shift in how we define medicine. For decades, drug development has been governed by Eroom’s Law, the observation that discovery becomes exponentially more expensive.

The four-layer stack of Read, Design, Learn, and Build offers a structural exit from this trend. In this regime, the therapeutic is no longer a static chemical entity. Rather, it becomes a dynamic software-hardware interface. Ultimately, this full-stack view hopes to bridge medicine’s “valley of death” with more data.

Further Reading

Asimov Press

Clinic-in-the-Loop

For the last several years, I have been trying to understand why biomedical progress, especially in therapeutics, has become less productive despite staggering advances in basic science…

Read more

4 months ago · 102 likes · 9 comments · Ruxandra Teslo and Asimov Press