Isomorphic Labs — Sir Demis Hassabis's company, which grew out of Google DeepMind in 2021 — published an article on September 29, 2026, titled "Building a new path to make medicines with AI," describing the Isomorphic Labs Drug Design Engine (IsoDDE): a platform in which physics-biological world models, frontier reasoning models, and generative design agents autonomously design molecules according to a given specification, taking 2–4 days for one run. In the presented case, a molecule found by the agents in a chemical space of approximately 10^60 compounds successfully modulated a biological target in real wet-lab tests and outperformed a literature reference that took researchers 3–5 years to create. For now, this is confirmed only by a corporate post without methodology or independent verification, so the key thing to watch next is the first transition of Isomorphic Labs' preclinical candidates into clinical trials.
What happened
Isomorphic Labs published a corporate post titled "Building a new path to make medicines with AI," announcing the transition of AI drug creation from hypothesis to a proven engineering discipline. The central platform is the Isomorphic Labs Drug Design Engine (IsoDDE): it combines physics-biological world models, frontier reasoning models, and generative design agents that autonomously and iteratively design molecules according to a specification — which includes the target, type of modulation, selectivity, solubility, and cell permeability. One run takes 2–4 days, resulting in a narrow set of candidates for further laboratory testing. In the described case, a molecule designed by the agents in days successfully modulated a biological target and outperformed a reference molecule from the literature that took researchers 3–5 years to create. The company also reported that preclinical data is preparing it for a transition to clinical development.
Context
Isomorphic Labs emerged in 2021 as a spin-off from Google DeepMind, and the team's scientific reputation is based on AlphaFold 2 — a protein structure prediction system that earned them the 2024 Nobel Prize. IsoDDE is the next claim at the intersection of AI and biology, but the task here is different: if AlphaFold 2 predicts already known structures, the Drug Design Engine must generate new molecules with specified properties, and the result can only be verified in the laboratory. Classical drug discovery relies on screening fixed libraries of 10^5–10^9 compounds, and each step of molecule optimization usually takes from one to three months. Isomorphic Labs instead describes a targeted agentic search in a space of approximately 10^60 possible molecules, where agents balance between exploration and exploitation and move the multi-objective Pareto frontier, for example, by the balance of potency and bioavailability.
Why this matters for the industry
For the industry, the main signal is the declared change in the cost structure in the early phase of drug discovery: if the molecule design step is indeed compressed in time from months to days, the bottleneck becomes not candidate generation, but synthesis, biological validation, and clinical trials, and vertically integrated AI-pharma transforms from a pitch slide into a working economic model. The phrase "days instead of years" is already becoming a benchmark in AI-pharma startup pitches and investor expectations, and companies without an answer to this narrative will face growing pressure. At the same time, there is currently no practical surface for assembly: IsoDDE is a closed internal platform without an external API, prices, or SLA, and there are no reproducible benchmarks either. Engineers can benefit from the pattern of "world model plus reasoning agents plus a verifier in the real world," as well as the tasks of designing the UX of multi-day autonomous runs — progress, checkpoints, intermediate agent results, and interfaces for comparing candidates across multiple goals — all of which are applicable in their own products today.
Why this matters for users
For the reader, this is the next series in the AlphaFold story: the same Demis Hassabis, but now AI is looking for working molecules where laboratory teams spent years. If the declared pace survives to the clinical stage, the path from a molecule idea to clinical trials could be significantly shortened, and new drugs could potentially reach patients faster. However, as of today, nothing changes for patients: there are no ready-made drugs created by IsoDDE, and the first candidate is only preparing to transition to the clinic. Practically, it is worth watching one marker — the first transition of Isomorphic Labs' preclinical candidates into clinical trials: it is this that will show whether AI drug design works beyond simulations and wet-lab validations.
What is still unknown / limitations
The evidence base is limited to one corporate case: the post has no methodology, datasets, or peer review, and the success rate, number of failed runs, target class, and comparison conditions with the reference remain unresolved, so selective selection of the best example cannot be ruled out. Wet-lab validation of target modulation is not equal to clinical efficacy. The 2024 Nobel Prize for AlphaFold 2 confirms the team's competencies, but does not validate IsoDDE: structure prediction and generative design of molecules with specified modulation are different tasks with different verification complexity. There is no independent discussion around the publication yet: on Hacker News, the material has 1 point and 0 comments, and the community cannot reproduce or refute the result. It is unclear whether technical details or a preprint on IsoDDE will be released, what the target of the candidate is, and exactly when it may enter clinical trials.
Sources
- Isomorphic Labs — article "Building a new path to make medicines with AI"
- Hacker News — discussion of "Building a new path to make medicines with AI" (1 point, 0 comments)
Author
Look at AI, editorial team