I spent last week at Advanced Therapies Week in San Diego, a four-day conference that brings together about two thousand people working on cell and gene therapies. Researchers, manufacturers, investors, regulators, patient advocates. The conference fills the San Diego Convention Center with five parallel tracks running simultaneously, covering everything from vector design to reimbursement strategy.
I went with a specific question: where does AI fit in the process of getting these therapies to patients? Not AI as a research topic in itself, but AI as a practical tool applied to the real problems people in this field face every day. The conference doesn't have a dedicated AI track, which made the question more interesting. AI showed up in conversations across many sessions, and I learned a few new valuable routes, especially around patients.
What makes advanced therapies different
For readers who aren't steeped in biotech, a brief orientation. "Advanced therapies" mostly means gene therapies and cell therapies. Gene therapies deliver genetic material into a patient's cells to treat disease, often using engineered viruses as delivery vehicles. Cell therapies involve modifying a patient's own cells (or donor cells) outside the body and reintroducing them. CAR-T therapy, where a patient's immune cells are engineered to recognize and attack cancer, is probably the most well-known example.
What makes these therapies categorically different from conventional drugs is that nearly everything about them is harder. A small-molecule drug is a chemical you can manufacture in bulk, ship in bottles, and store on shelves. A cell therapy might start with a specific patient's blood, require weeks of specialized manufacturing, ship frozen on a tight timeline, and be administered once. The pipeline from discovery to patient involves not just scientific challenges but logistical, regulatory, communicative, and organizational ones. Every interface between steps is a potential bottleneck.
The expected places
I'll work through the therapy development pipeline roughly in order. The AI applications at each stage were more or less what I expected, given where I spend most of my professional attention. But hearing the current state of things from people doing this work was valuable.
Discovery and target identification. Before you can build a therapy, you need to understand the disease well enough to know where to intervene. This means identifying the right genetic targets, understanding disease mechanisms, and predicting which modifications will produce therapeutic effects. Domain-specific models, often fine-tuned on genomic and proteomic data, are being used to narrow the search space. The value here is speed: exploring candidate targets computationally before committing to expensive wet-lab validation.
Construct and vector design. Once you have a target, you need a way to reach it. For gene therapies, this often means engineering an adeno-associated virus (AAV) to deliver the genetic payload. The design space is enormous, since you're optimizing for specificity (reaching the right cells), efficiency (delivering the payload), and safety (avoiding immune responses). Several sessions at the conference discussed computational approaches to vector engineering, including a session on engineering safer AAVs for human gene therapy. AI is being used here to predict how modifications to the viral capsid will affect tropism and immunogenicity, which reduces the number of variants you need to test empirically.
Manufacturing. This was perhaps the most heavily represented area at the conference, with an entire theater dedicated to technology and automation. The core challenge is consistency at scale. These aren't pills you stamp out by the millions; they're biological products with inherent variability. Sessions covered the "smart factory" vision for cell and gene therapies, automation of cell therapy workflows, process analytics, and quality control. AI appears in process monitoring (detecting deviations in real time), predictive maintenance, and optimizing manufacturing parameters. One session on process control discussed embedding digital maturity across the entire manufacturing lifecycle. The aspiration is to move from reactive quality control, where you test the product after you've made it, to in-process analytics that catch problems as they develop.
Clinical development. Getting a therapy from the lab into patients requires trial design, patient selection, endpoint definition, and regulatory strategy. The conference had sessions on biomarkers and companion diagnostics, trial design challenges specific to advanced therapies, and choosing the right modalities. AI is being applied to patient stratification (identifying who is most likely to benefit), trial simulation (predicting enrollment and outcomes before committing resources), and regulatory document preparation. The challenges here are partly scientific and partly organizational, since trials for rare diseases often struggle with small patient populations and complex endpoints.
Commercialization and market access. Even after a therapy is approved, delivering it to patients who need it involves supply chain logistics, reimbursement negotiations, and navigating payer systems. Sessions covered pricing and reimbursement models, decentralized manufacturing, and supply chain resilience. AI is starting to be used in demand forecasting, logistics optimization, and modeling reimbursement scenarios. For therapies that cost hundreds of thousands or millions of dollars per patient, the commercial model itself is a problem that requires creative solutions.
What surprised me
The applications above map onto the technical pipeline in ways that feel natural. Scientists and engineers applying computational tools to scientific and engineering problems. This is where I expected to find AI in advanced therapies (and did).
What I didn't expect was how much of the conversation turned to problems that are more social than scientific. Problems of communication, bureaucracy, and the friction between people and patients in a system that is, by necessity, extraordinarily complex.
One story stayed with me. A patient receiving an advanced therapy lost access to their drug. The insurer switched them to an inferior but biocompatible replacement (without notification). When the change became obvious through poor outcomes, the patient and their family tried to get the original therapy reinstated, but the insurer wouldn't budge, and the care team seemed unable or unwilling to push back. The system had made a decision, and the system wasn't interested in revisiting it. The patient turned to AI to navigate the bureaucracy. They used it to research the specific regulations governing their situation, identify which laws the change potentially violated, and draft targeted communications to the people with authority to reverse the decision. The emails went out. No one responded directly. But a few days later, the care team switched the treatment back. It's hard to say whether the AI-drafted communications were the decisive factor, but the story illustrates something important about where bottlenecks actually occur in healthcare. This patient didn't need a better drug, just to get the right care from an inefficient system.
A second area that surprised me was how patients experience the therapies themselves. Advanced therapies involve science that sits outside most people's daily experience. Worse, they involve concepts that sound frightening when described casually. Viral vectors sound like infections. Gene editing sounds like tinkering with the essence of a person. CRISPR has cultural baggage that extends well beyond what it does in a clinical context. Several sessions on patient-centric development explored how patients and their families need ways to understand what these therapies actually do, what the real risks are, and how to evaluate them separately from hype and fear. One session, "Saving Sophie," traced a patient-driven path to advanced therapy cancer care, and the patient advocacy track explored how engagement with patients can shape not just communication but the design and prioritization of therapies themselves.
The gap between what a therapy developer understands about their product and what a patient needs to understand to make an informed decision is large. AI can serve as a translation layer here, not dumbing things down but meeting people where they are, explaining mechanisms in terms that connect to what a patient already knows, and adapting to their specific concerns rather than delivering a one-size-fits-all brochure.
Impedance matching
I've been looking for a phrase that captures the common thread across these applications, both the expected ones and the surprises. The best analogy I've found comes from electrical engineering: impedance matching.
In a circuit, impedance matching means adjusting the characteristics of connected components so that power transfers efficiently between them. When impedances are mismatched, energy is lost at the interface. It reflects back instead of passing through. The components might each work fine in isolation, but they lose something at the connection.
Advanced therapy development is a chain of interfaces between very different groups of people. Researchers talk to manufacturers. Manufacturers talk to regulators. Regulators talk to companies. Companies talk to insurers. Insurers talk to doctors. Doctors talk to patients. Each of these groups has its own language, incentives, constraints, and ways of understanding the world. The bottlenecks in getting a therapy from bench to bedside often aren't within any single group; they're at the interfaces between them.
AI, and large language models in particular, have a natural capacity for this kind of interface work. Translating between technical contexts is something they do well. A model that can read a regulatory document and help a small biotech understand what's required, or take a complex mechanism of action and explain it to a patient in terms that are accurate without being terrifying, is performing impedance matching. It's not replacing the people on either side, just reducing the energy lost at the connection points.
This framing also explains why the patient-facing applications surprised me. I'd been thinking about AI in advanced therapies as a tool for solving scientific and engineering problems, which it is. But the hardest bottlenecks in delivering these therapies to patients aren't always scientific. Sometimes they're a patient with a doctor who doesn't know about these options or an insurer who won't listen. Sometimes they're a family trying to understand whether a therapy involving a virus is safe for their child. Sometimes they're a small therapy developer trying to communicate with a regulatory system designed for large pharmaceutical companies.
These are all impedance mismatches. The people on both sides of the interface may be competent and well-intentioned, but they're operating with different information, different languages, and different constraints. AI doesn't solve the underlying organizational problems. Those involve people, and will continue to for the foreseeable future. But it can reduce the friction at the interfaces where communication breaks down, and that turns out to matter more than I expected.