For most of the last decade, "AI drug discovery" lived in a strange middle space — too credible to dismiss, too unproven to bank on. When the BIO International Convention wrapped in San Diego this week, after four days that gathered more than 20,000 executives, investors, and researchers, that ambiguity had quietly resolved. The questions in the hallways had changed. For three years the dominant one was a hopeful hypothetical: could machines actually design medicines? This year it was operational and a little impatient: what is working, what is not, and how fast do we have to move?
What Happened
BIO is the biopharmaceutical industry's annual barometer. It does not set strategy so much as reveal it, because the people who allocate capital, license compounds, and argue with regulators are all in one building for four days, and the conversations there tend to predict the moves that follow over the next year and a half. The signal from BIO 2026 was unusually consistent: generative AI in drug discovery has crossed from an investment thesis into something closer to a clinical fact.
That shift had a concrete cause rather than a marketing one. Platforms built around generative protein design — companies such as Xaira Therapeutics, Generate Biomedicines, and Isomorphic Labs — have moved past the stage of simply filling their pipelines with computational candidates. They are now producing molecules that enter and clear early human testing. The distinction between a promising pipeline and a clinical result is the whole game in this field, and for the first time the conversation had real data to stand on.
Why It Matters
To understand why this is a genuine inflection rather than another hype cycle, it helps to separate two things that often get blurred together. The older, more familiar approach is library screening: a model evaluates millions of already-existing compounds against a biological target and ranks the most promising ones. It is fast and useful, but it is fundamentally an accelerated search through chemistry that already exists. Generative design is a different act altogether. A diffusion model begins with what is essentially random atomic noise and learns, step by step, to remove that noise until a coherent protein structure emerges — one shaped to bind a chosen target, fold into a stable form, or fit a pocket that was previously considered undruggable.
.png){kind=link}
The mechanism is not metaphorical. The breakthrough method, RFdiffusion, came out of the University of Washington's Institute for Protein Design and was published in Nature in July 2023. It borrows the exact denoising idea that powers AI image generation and applies it to protein backbone coordinates. During training, known structures are gradually corrupted with noise across roughly two hundred steps, and the network learns to run that process in reverse. At inference, it starts from random residue frames and converges on a real, foldable backbone. The practical payoff is dramatic: where earlier design methods might demand tens of thousands of candidate molecules before one performed as intended, this approach has reduced that to as few as one per design challenge in lab experiments.
That capability kept widening. A December 2025 update, RFdiffusion3, pushed the method down to atom-level design across every class of biomolecule — including DNA and small-molecule interactions — and was released as open-source software, which means the tools driving this moment are no longer locked inside a handful of well-funded labs. When a technique this powerful becomes freely available, the question stops being who can do it and becomes who can do it well.
The Field Reacts
The mood among investors at BIO 2026 reflected that maturation. Capital had already voted with conviction: more than $11 billion flowed into AI and machine-learning drug discovery and licensing in 2025 across hundreds of separate funding rounds. But money chasing a compelling architecture is a different thing from money rewarding clinical evidence, and this year the second kind started to dominate. The expectation hardening across the room was that future capital will concentrate in platforms with clinical-stage data, not just elegant models.
It is worth noting how much of this rests on open science rather than proprietary secrecy. David Baker, who shared the 2024 Nobel Prize in Chemistry for computational protein design, co-founded Xaira Therapeutics, which launched in April 2024 with more than a billion dollars in committed funding and builds directly on the diffusion architecture his lab pioneered and then gave away. The ecosystem around these methods has grown collaborative in revealing ways — a recent generative model from NVIDIA, validated against more than a hundred therapeutic targets, was developed alongside partners including Novo Nordisk, the European Bioinformatics Institute, and Seoul National University. The center of gravity is shifting from lone breakthroughs to a shared toolkit.
What Comes Next
The clinical milestone that gave BIO 2026 its confidence did not actually happen at the convention — it happened a year earlier, and the industry spent this week absorbing what it meant. Insilico Medicine's Rentosertib, a TNIK kinase inhibitor designed end to end by the company's generative platform, completed a Phase IIa trial in idiopathic pulmonary fibrosis and published peer-reviewed results in Nature Medicine on June 3, 2025. In the 71-patient, double-blind, placebo-controlled study, patients on the 60 mg daily dose showed a mean improvement in lung function of 98.4 mL, against a decline of 20.3 mL in the placebo group. What made it a landmark was not the effect size alone but the provenance: both the biological target and the compound itself were identified by AI — the platform proposed the target by reading multi-omics data and then designed the molecule by navigating chemical space, rather than a chemist using AI as an assistant.
None of this means the problem is solved. A single Phase IIa result in one indication is encouraging, not a validated technology, and no AI-designed drug has yet earned FDA approval. Full approval still requires Phase III completion, which routinely takes years. The regulatory picture adds its own friction: the FDA issued draft guidance on AI in drug development in January 2026, but it addressed how AI is used in regulatory decision-making rather than the discovery phase itself, leaving companies to integrate these methods ahead of clear standards for the data they produce. The honest summary is that the pipeline is no longer only filling — it can deliver — but the path from delivery to approved medicine remains long.
Closing Thoughts
There is a quieter story underneath the capital flows and the trial statistics, and it is the one most worth sitting with. For a century, drug discovery was defined by scarcity — of time, of viable starting points, of the sheer luck required to stumble onto a molecule that fit. Generative design does not abolish that difficulty, but it changes its shape. When a model can propose a protein built to order, the bottleneck moves from imagination to validation, from "can we even think of a candidate?" to "which of these many candidates is real?"
That is a more hopeful constraint to live under, and it is also a more human one than the breathless framing usually allows. In every one of these stories, people still ran the experiments, judged the results, and carried the responsibility for what reached a patient. The machine widened the space of what could be tried; it did not remove the need for someone to decide what should be. BIO 2026 will be remembered less for any single announcement than for the moment a field stopped asking whether the idea could work and started asking what it owes the people it might one day treat. That is not the end of the story. It is, finally, the beginning of a more serious one.
한글 요약
이번 주 샌디에이고에서 막을 내린 BIO 국제 컨벤션은 생성형 AI 신약 설계가 '투자 가설'에서 '임상적 사실'로 넘어가는 전환점을 보여주었습니다. 2만 명 넘는 업계 관계자들의 화두는 더 이상 "AI가 정말 약을 설계할 수 있는가"가 아니라 "무엇이 되고 무엇이 안 되며 얼마나 빨리 움직여야 하는가"로 바뀌었습니다. 기존의 라이브러리 스크리닝이 이미 존재하는 화합물을 탐색하는 방식이라면, 확산(diffusion) 모델 기반 생성 설계는 무작위 원자 노이즈에서 출발해 원하는 표적에 결합하는 단백질 구조를 처음부터 만들어 낸다는 점에서 근본적으로 다릅니다.
이 흐름의 토대는 데이비드 베이커 연구진이 2023년 Nature에 발표한 RFdiffusion이며, 2025년 12월 공개된 RFdiffusion3는 원자 단위 설계까지 확장돼 오픈소스로 풀렸습니다. 2025년 한 해에만 AI 신약 분야에 110억 달러 이상이 유입됐고, 자본의 무게중심은 '멋진 모델'에서 '임상 데이터를 가진 플랫폼'으로 옮겨가고 있습니다. 베이커가 공동 창업한 Xaira를 비롯해 Generate Biomedicines, Isomorphic Labs 등이 이 흐름의 중심에 있으며, NVIDIA의 최신 생성 모델은 노보 노디스크·유럽 생물정보학연구소·서울대 등과 함께 100개 이상 표적에 대해 검증됐습니다.
결정적 이정표는 인실리코 메디슨의 렌토서팁이었습니다. 표적과 화합물 모두 AI가 도출한 이 TNIK 억제제는 특발성 폐섬유증 2a상에서 폐 기능을 유의미하게 개선하며 2025년 6월 Nature Medicine에 결과를 실었습니다. 다만 아직 FDA 승인을 받은 AI 설계 신약은 없으며, 3상까지는 수년이 더 필요합니다. 그럼에도 이번 컨벤션의 의미는 분명합니다. 한 분야가 "이 아이디어가 작동할까"를 묻는 단계를 지나, 언젠가 치료하게 될 환자들에게 무엇을 빚지고 있는지를 진지하게 묻기 시작했다는 점입니다.
참고: Tech Times, Insilico Medicine, Baker Lab, BIO International Convention.