AI drug discovery depends on what happens after a model makes a prediction. A proposed molecule or biological hypothesis still needs experimental evidence.
Reuters reported in September 2026 that Anthropic had established a wet lab in the San Francisco Bay Area, extending its life sciences work into physical experimentation. The report brings renewed attention to a practical research question: how can computational design and laboratory testing work together more effectively? Reuters reporting
That question is at the heart of closed-loop AI drug discovery: connecting predictions, experiments, and results so that each round informs the next.
Why Does an AI Company Need a Wet Lab?
Computational tools can help researchers review literature, prioritize targets, and design candidate molecules. Experimental work establishes whether those ideas hold up under the conditions being tested.
The attraction of a connected workflow is straightforward. If experimental results can be returned in a consistent, usable format, researchers can use them to refine the next hypothesis or candidate set. XtalPi, for example, describes its platform as linking scientific models with robotic experiments in an iterative dry–wet workflow. XtalPi’s platform
Our interpretation: closer coordination between computational and experimental teams could reduce delays between proposing an experiment and learning from its results. The benefit depends on the workflow, data quality, and integration. Building a wet lab alone does not establish a particular speed or cost advantage.
From Research Software to Physical Experiments
Anthropic’s scientific tools provide context for the reported laboratory expansion. Its October 2025 Claude for Life Sciences announcement described support for literature research, hypothesis development, bioinformatics, and protocol drafting, with connections to resources such as PubMed and Benchling. Anthropic announcement
Claude Science, now described by Anthropic as a beta research workbench, adds scientific database connections, Python and R workflows, and records linking outputs to the code and environment that produced them. Claude Science
On August 27, 2026, Anthropic also introduced a research preview of the Model Hardware Standard, or MHS, for connecting AI agents with programmable physical equipment. One published example is a Genentech proof of concept for a BCA protein assay, coordinating a liquid handler, robotic arm, and plate reader. MHS research preview
That example makes the opportunity tangible: software coordinates a sequence of instrument tasks that produces an experimental measurement. It also defines the limits of the evidence—a demonstrated workflow is not proof that every biological experiment can run autonomously.
Enterprise adoption is a separate development. In May 2026, Bristol Myers Squibb announced plans to deploy Claude across more than 30,000 employees in functions spanning research, development, manufacturing, and commercial operations. That figure refers to employees across the business, not 30,000 research scientists already using the system. BMS announcement
What Is Closed-Loop AI Drug Discovery?
The Chinese expression “干湿闭环” describes a feedback loop between computational work—the “dry lab”—and physical experiments—the “wet lab.” In discovery chemistry, it can be understood through a design-make-test-analyze cycle:
- Design: Use existing evidence and computational tools to propose candidates or experiments.
- Make: Prepare the compounds, samples, or biological materials needed for testing.
- Test: Collect measurements through suitable biochemical, cellular, or other experimental assays.
- Analyze: Evaluate the results and use them to guide the next round.
Automation can connect selected steps, but the essential feature is feedback: experimental results change what the team does next. A closed loop does not require every result to retrain a foundation model, nor does it require every experiment to be performed in-house.
The distinction matters when evaluating an “AI lab.” Generating more candidate designs, integrating instruments, and improving a drug’s clinical outcome are different accomplishments.
XtalPi and Insilico Show Different Paths
XtalPi describes a platform combining AI agents, scientific models, and robotic laboratory execution. Its emphasis illustrates one route: provide infrastructure that connects computational decisions with physical testing. XtalPi
Insilico Medicine offers a therapeutic-program example. Its investigational TNIK inhibitor rentosertib, developed for idiopathic pulmonary fibrosis, combines an AI-prioritized target with a molecule designed using generative chemistry. A randomized Phase IIa study was published in Nature Medicine in 2025. On July 7, 2026, the company announced initiation of Phase III testing. Research publication record · Insilico’s Phase III announcement
These examples represent different business models: experimental infrastructure and drug development. Neither establishes the eventual scale, operating model, or clinical impact of Anthropic’s laboratory.
Faster Discovery Still Needs Clinical Proof
Progress into clinical testing is meaningful, but it is not the same as an approved treatment. Insilico’s Phase III announcement explicitly describes rentosertib as investigational and not approved by any regulatory authority. Insilico announcement
The FDA explains that later-stage studies evaluate treatment benefit and collect broader safety evidence, including adverse effects that smaller or shorter studies may miss. Faster computational work does not remove those questions. FDA: Clinical Research
For researchers following this field, it helps to ask which milestone a claim describes: a better prediction, a faster experimental cycle, a nominated candidate, or a demonstrated clinical benefit. Each requires different evidence.
What This Means for Everyday Lab Work
D2L’s practical interpretation: as research connects more tightly with computation, the quality and traceability of the underlying experiments remain central. The model can only learn from the measurements and context the workflow captures.
Match Labware to the Experiment
For a cell-based experiment, start with the biological model and assay requirements. Consider the culture surface, well format, working volume, and readout. A standard plate footprint alone does not establish compatibility with a particular robot, reader, or assay.
D2L’s sterile, TC-treated cell culture plates are one option to review for routine cell culture and cell-based work. Select the relevant format and qualify it within your own method. This is a workflow-based product connection; it does not imply use by Anthropic, Genentech, XtalPi, or Insilico.
Preserve the Experimental Record
Record sample identities, plate maps, relevant consumable lots, protocol versions, controls, and deviations alongside results. This is a practical recommendation: those records help teams distinguish a biological result from a change in how the experiment was performed. The emphasis is consistent with Claude Science’s stated focus on tracing analytical outputs to their underlying work. Claude Science
For lab teams, the immediate opportunity is to make each experimental cycle easier to execute, interpret, and repeat. Explore D2L’s research and lab-workflow articles for more practical perspectives from the bench.
Related Lab Essentials
Explore labware for automated liquid handling, assay setup, and cell-based research.
- Hamilton-Compatible Conductive Pipette Tips, Sterile — For compatible automated liquid handling and assay setup.
- Reagent Reservoir Plates — For reagent dispensing and multichannel or robotic pipetting.
- Assay Plates — Explore formats for biochemical assays and plate-based readouts.
- Cell Culture Plates, Sterile, TC-Treated — For routine adherent cell culture and cell-based assays.
