Anthropic's Claude Speeds Up Drug Discovery Work: CAR-T Sequence Cut 70%, Batch Record Review Time Down 89%

Developing a new drug generally takes more than a decade and well over a billion dollars. Along the way come countless tasks: finding drug targets, designing molecules, running in vitro binding assays, conducting animal studies, auditing manufacturing records, and carrying out long clinical trials. When people talk about AI in drug discovery, attention tends to go to predicting protein structures or generating new molecules. But molecular design is not the only thing slowing down drug development. Delays also arise throughout the process, in experimental data handling, quality control, and coordination among research organizations.

A series of research results Anthropic published on October 27, 2026 compiles cases in which its AI model Claude was applied across this wide range of drug discovery work. Six organizations took part in the collaborations: BigHat Biosciences, Bristol Myers Squibb (BMS), Divergence Labs, Enveda, Generate Biomedicines, and The Column Group (TCG). The results also include a muscular dystrophy drug development program at Mithril Medicine, which Anthropic acquired. Claude was used not only for molecular design but also for reanalyzing experimental data, supporting the drafting of regulatory documents, and coordinating work with outside research institutions.

The announced results include figures such as a roughly 70% reduction in the sequence length of a CAR-T cell signaling domain, a binding affinity improvement of 10x or more for an antibody, and an 89% reduction in time spent reviewing manufacturing records. However, these come from different research subjects and are measured by different metrics, so they cannot be compared on a single scale. The published materials are also mainly technical reports and preprints compiled by each company, not peer-reviewed academic papers. Most of the research is at the preclinical stage, consisting of in vitro evaluations and early animal studies. Which steps of drug discovery can AI make more efficient, and what challenges remain? Below, we look in detail at the experimental results of each program and the caveats needed to interpret them.

{
  "type": "bar",
  "title": "創薬支援プログラムにおける主要な定量的改善",
  "unit": "倍率または短縮率（%）",
  "categories": [
    "CAR受容体の細胞表面発現量（倍）",
    "抗体バリアントの結合親和性（倍）",
    "CARシグナル伝達ドメインの短縮率（%）",
    "製造記録レビューの所要時間短縮率（%）"
  ],
  "series": [
    {
      "name": "達成された改善水準",
      "values": [10, 10, 70, 89]
    }
  ],
  "caption": "各社技術報告書に基づく実験室および業務改善の実測値",
  "source": "Anthropicおよび提携各社の公開技術報告書（2026年10月）"
}

As the chart shows, the metrics for which improvements were reported span a wide range, from receptor expression on the cell surface and molecular binding affinity to the time needed to review manufacturing records. Because multiples and percentage reductions are mixed together, the length of the bars cannot be read directly as the relative size of each result. None of these results came from the AI model alone; they were obtained by combining existing laboratory equipment, analysis tools, and the judgment of human researchers.

Shrinking CAR-T receptors: signaling domain sequence cut by 70%

Chimeric antigen receptor T-cell (CAR-T) therapy is an immunotherapy in which T cells collected from a patient are genetically modified to attack cancer cells and then returned to the body. It has achieved high efficacy in blood cancers such as leukemia and malignant lymphoma, but challenges remain in applying it to solid tumors and in reducing manufacturing costs. For solid tumors in particular, there is a need to add functions such as recognizing multiple cancer antigens and protecting T cells from the immunosuppressive environment inside tumors. That requires delivering additional genetic information into the cells, but the vehicles that carry genetic material have capacity limits.

Generate Biomedicines focused on making the intracellular signaling domain of the CAR smaller. This domain relays information into the T cell when the receptor on the cell surface recognizes a cancer antigen, triggering an attack response. Viral vectors and other tools are used to introduce genetic material into CAR-T cells, and lipid nanoparticles (LNPs) encapsulating mRNA are also being studied as a delivery technology, but all of them limit the size of the nucleic acid they can carry. If an existing signaling domain can be made smaller, that frees up room to add other functions.

The company's research team explored ways to shorten an amino acid sequence while preserving the protein's function. The idea resembles "refactoring" in software, tidying code without changing what it does, but with proteins it is hard to judge in advance which parts can be removed without losing function. Intracellular signaling domains in particular contain intrinsically disordered regions that lack a fixed three-dimensional structure, and their roles are not fully understood. So six scientists built an AI agent system using tools such as Claude Code and had it consider which regions to keep and which could be deleted, based on prior studies, experimental data, and the researchers' hypotheses.

The team selected 186 candidate sequences from the designs proposed by the AI, introduced them into human T cells, and evaluated them in vitro. The best candidate was reported to be about 70% shorter in amino acid sequence length than the standard signaling domain widely used in approved CAR-T products. However, the specific number of amino acid residues in the reference sequence and the shortened sequence is not given in the publicly available materials we were able to check. In addition, T cells carrying the shortened domain were reported to show roughly 10 times more CAR receptor on the cell surface than the standard version. Here too, absolute receptor counts and measured mean fluorescence intensity have not been published, and the comparison is presented as a multiple.

In vitro, T cells with the shortened domain killed tumor cells more efficiently and released more of the signaling factors that promote immune responses. "Tonic signaling," in which the receptor activates autonomously in the absence of the target antigen, was also suppressed, and a tendency toward reduced T-cell exhaustion was observed. The research team suggests that the increased receptor expression on the cell surface may have contributed to the functional improvement, but this is an interpretation based on in vitro observations. Solid tumors have physical barriers that block immune cells from entering and a tumor microenvironment that suppresses immune responses, so better performance in cultured cells does not necessarily translate into therapeutic effect. For practical use, the next step is animal studies, such as in disease-model mice, to evaluate antitumor effect, safety, and durability of effect.

Pulling IgE off its receptor: a search for new food allergy therapies

An estimated 33 million people in the United States have food allergies, and severe cases carry the risk of life-threatening anaphylaxis. In 2024, the antibody drug omalizumab (brand name Xolair) was approved for multiple food allergies. It binds immunoglobulin E (IgE), which is involved in allergic reactions, and mainly blocks free IgE circulating in the blood from attaching to receptors on immune cells. However, at normal dosing concentrations it is difficult to pull off IgE that is already bound to the high-affinity IgE receptor (FcεRI) on the surface of cells such as mast cells.

Allergic reactions are triggered when an allergen binds IgE on the cell surface and cross-links the IgE-receptor complexes. For this reason, even if an existing drug suppresses IgE activity, the risk of a serious reaction from accidental ingestion does not disappear entirely. For patients who react strongly to trace amounts of peanuts and similar allergens, avoiding allergens and carrying an epinephrine auto-injector (EpiPen) for emergencies remain important.

Divergence Labs aimed to develop a new antibody that not only captures free IgE but also directly pulls off IgE already bound to the receptor. It adopted an asymmetric antibody design that combines two arms with different functions. One "anchor arm" grips IgE strongly, while the other "disruptor arm" acts on the interface between IgE and its receptor to dissociate the two. A distinguishing feature is that it targets IgE bound to the cell surface, which existing drugs have difficulty removing.

This design, however, comes with a serious safety challenge. Ordinary antibodies have two antigen-binding sites, so depending on the design they may cross-link IgE molecules on the cell surface. In that case, a drug meant to suppress allergic reactions could instead activate immune cells. The antibody must capture IgE reliably while avoiding unintended receptor cross-linking, which requires precise control over the properties needed in each of the two arms.

To address this, Claude supported literature analysis, screening of candidate structures, and antibody sequence design. For the disruptor arm, it refined sequences obtained from prior studies, while the anchor arm was designed from scratch with new sequences. A total of 22 autonomous exploration sessions were run, using about 2,750 hours of NVIDIA H100 GPU compute and generating roughly 35,000 antibody sequences. Candidates were then narrowed down using structure prediction models, and dozens for each arm were synthesized and measured in the lab.

In evaluations using biolayer interferometry (BLI), which optically measures molecular binding and dissociation, 20 of 91 disruptor arm candidates significantly dissociated receptor-bound IgE within 1 to 2 minutes. The top three of these showed greater dissociation activity than omalizumab under the in vitro measurement conditions. Among 57 anchor arm candidates, three were found that bound IgE strongly, with dissociation constants (\(K_{\mathrm{D}}\)) comparable to existing antibody drugs.

According to Divergence Labs, the search, which would have taken up to nine months with conventional methods, was completed in about one month. However, the molecules obtained are only early-stage discovery candidates known as "hits." The anchor and disruptor arms have been evaluated separately, and it has not yet been verified whether a complete antibody combining the two will show the intended dissociation effect. Especially important is confirming that it does not mistakenly cross-link IgE on the cell surface and trigger allergic reactions. Even excellent results in molecular binding tests do not by themselves justify concluding that allergic reactions can be safely suppressed in the body.

It was also reported that the AI did not always proceed efficiently. While designing a particular loop structure, Claude proposed five consecutive calculations that changed conditions only slightly, even though the experimental results had already given a clear conclusion. Recognizing such dead ends and switching to a different strategy required judgment from human researchers. Speeding up candidate generation with AI and the ability to steer an entire research effort in the right direction are not necessarily the same thing.

Not just stronger binding: finding errors in experimental data

In drug discovery, after a molecule that binds a target is found, "lead optimization" is carried out to improve the properties needed in a pharmaceutical. Candidate molecules must have not only high affinity for strong binding to the target but also stability against temperature changes and resistance to aggregation and precipitation in solution. These properties influence one another, however, so improving one can degrade another. For example, changing an amino acid sequence to strengthen target binding can destabilize the protein's three-dimensional structure, making it vulnerable to heat or prone to aggregation.

To improve multiple properties simultaneously, the research team at BigHat Biosciences and Anthropic linked Claude with "Reccy," an automated experimental system developed by BigHat. Two targets were used: an anti-albumin VHH (single-domain antibody), used to extend residence time in the blood, and a single-chain antibody (scFv) targeting mesothelin, which is highly expressed on the surface of solid tumor cells. By using a cell-free protein synthesis system, the company shortened antibody expression, which usually takes several weeks, to a few days. It also built a setup that can automatically synthesize 176 antibody variants per week and measure the physicochemical properties of each.

The study ran for about 10 weeks, repeating experiments and optimization over seven rounds. During that time, 1,010 unique antibody variants were evaluated, with a total of 1,232 measured samples. In the final variants selected for improving multiple properties in balance, binding affinity improved 6- to 9-fold and thermal stability indicators improved by 8 to 18°C compared with the starting molecules. However, the absolute dissociation constant (\(K_{\mathrm{D}}\)) values for the reference and modified molecules are not given in the publicly available materials we were able to check. There were also cases in which some variants improved binding affinity by 10-fold or more and others improved thermal stability by 20°C or more, but these maximum values were not all achieved simultaneously in a single candidate molecule.

The main drug discovery programs in this announcement, along with what was evaluated, the results, and the validation still needed, are summarized below.

Note 1: For sequence length and receptor expression, the absolute values for the reference and comparison are not given in the publicly available materials we were able to check. Note 2: The absolute dissociation constant values used to compare binding affinity are also not given in the publicly available materials we were able to check.

What stands out most in the BigHat study is that Claude not only supported antibody sequence design but also helped uncover a problem in how experimental data were interpreted. During optimization, results from the instrument measuring thermal stability and aggregation varied widely, and the quality control system excluded many measurements as unreliable. At first, problems with measurement conditions or sample quality seemed possible, but when the team examined the instrument's raw data closely, they found the problem lay in the very metric that the conventional automated analysis software had been calculating.

At issue was the interpretation of the protein's denaturation temperature (\(T_{\mathrm{m}}\)). The team reanalyzed raw data recording changes in fluorescence signals and light scattering as temperature rose. They found that in some antibody variants, aggregation began and the protein precipitated out of solution before it fully denatured with heat. The value the conventional software treated as the denaturation temperature may in fact have reflected the temperature at which light scattering changed sharply because of precipitation. For VHHs in particular, aggregation began at a relatively low temperature of about 40°C, greatly distorting the measurements.

Confusing protein denaturation with aggregation changes the design strategy for improving a molecule. If the protein itself is prone to denaturing with heat, stabilizing features such as its internal hydrophobic core can be effective. If molecules tend to aggregate with one another, different measures are needed, such as adjusting amino acid residues on the surface. Claude compared the raw data against simulation models and proposed analyzing aggregation not as a simple temperature threshold but as a reaction-rate problem that progresses over time.

To test this hypothesis, the team ran additional experiments varying sample concentration and heating rate. A true denaturation temperature should remain relatively stable regardless of heating rate, whereas a time-dependent aggregation phenomenon should cause the apparent collapse temperature to shift with the time samples spend at each temperature and with concentration. In the actual measurements, the collapse temperature of the scFv varied by 2 to 5°C depending on conditions, supporting the interpretation that aggregation had been affecting the conventional metric. The team then revised its analysis method and resumed selecting candidate molecules using a more appropriate metric.

What makes this case significant is that AI was used not only to propose new molecules but also to revisit existing experimental data and improve how measurements were interpreted. However, what was confirmed here is improvement in antibody binding affinity and physicochemical stability. Whether the candidate molecules show suitable pharmacokinetics (PK) in the body, whether they avoid being recognized as foreign by the immune system (immunogenicity), and whether they can be stably mass-produced in commercial-scale manufacturing facilities all remain to be verified.

Auditing 1,061 pages of manufacturing records: review time cut by 89%

For drug discovery to become more efficient, the heavy document work tied to manufacturing and regulatory compliance matters as much as molecular exploration and experiments. In pharmaceutical manufacturing, raw materials, manufacturing conditions, quality test results, and more must be recorded in detail in accordance with Good Manufacturing Practice (GMP). Every piece of information about the process is subject to recording, including equipment temperature settings, raw material lot numbers, mixing times, and product purity tests. As a result, the records for a single batch, the unit of one manufacturing run, can reach anywhere from several hundred to 1,000 pages.

At Enveda, which searches for drug candidates from natural sources such as plants, manufacturing chemists reportedly spent up to 75% of their working hours reviewing paperwork. Just scrutinizing a 500-page batch record received from a contract development and manufacturing organization (CDMO) could take an experienced person about 10 hours. Such review is essential for ensuring product quality and safety, but it also slows R&D and manufacturing schedules.

To address this problem, Enveda developed "CMC Document Processor (CDP)," a document processing system built around Claude. For performance validation, it used records totaling 1,061 pages across four lots, created in an actual manufacturing campaign for Phase 2 clinical trials. These records were later submitted to the U.S. Food and Drug Administration (FDA). The system's inspection results were compared with manual reviews by two internal reviewer teams and two external consultants. The quality assurance department then evaluated the validity of the detected issues from the standpoint of whether they could be cited by a regulatory inspector.

Across multiple validation runs, CDP detected all 11 action items identified by human experts. The findings included a case in which a manufacturing step exceeded the specified time and a case in which a test result was recorded as failed but the measured value field was blank. In a separate validation trial with strictly timed processing, it also found seven additional minor issues that had been missed in human review. These are results based on specific documents and evaluation conditions, and they do not confirm that equivalent detection performance can be obtained for every manufacturing record.

A major reduction in working time was also reported. A task that took about 26 hours with conventional manual review was completed in under three hours, including both CDP's analysis and human verification of its output, an 89% reduction in time. In another case involving a regulatory submission document, the first draft, which had taken about three months by hand, was completed in about four hours. Using AI to extract and cross-check document contents likely helped narrow down the issues experts needed to examine, contributing to the time savings.

This result does not mean human review has become unnecessary. In fact, when humans reviewed a first draft of a regulatory document generated by AI, they found one unit conversion error. CDP is also designed to run in a protected, view-only environment and has no authority to modify official manufacturing records. The final decision on whether to release a manufactured batch continues to rest with the head of the quality assurance department. The 89% time reduction reported here is a validation result using a specific set of 1,061 pages, and the same effect is not guaranteed for all GMP-based audit work.

One founder drives drug development, reaching mouse studies in under 18 months

AI-driven efficiency in drug discovery goes beyond molecular design and document processing. Streamlining tasks such as contracting with outside research institutions, ordering experiments, shipping samples, and analyzing results is another way to shorten development timelines. Mithril Medicine, acquired by Anthropic, is a biotechnology company that has been pursuing this kind of approach. Founded single-handedly by Cody Tranbarger, the company is working on therapies for two rare forms of muscular dystrophy that affect children.

Normally, when a drug company uses multiple outside research institutions, it faces a large amount of coordination work, from signing confidentiality agreements and handing over samples to compound synthesis, cell assays, and tracking the progress of animal studies. Handling this smoothly requires staff for legal affairs, business development, and research management. Tranbarger used Claude to draft contracts, coordinate sample shipments and work across 11 outside laboratories, and analyze experimental results. As a result, within 18 months of launching the development concept, he reportedly obtained data showing improved muscle function in disease-model mice.

Based on this result, Anthropic says that if development goes smoothly, an early-stage human clinical trial could begin as early as the second half of 2028. However, the 18-month development period was not achieved solely through AI-driven efficiency gains in administrative work. It resulted from a combination of factors, including researchers' judgment, the experimental capabilities of outside institutions, and progress in biological exploration. Determining how many months were actually saved by adopting Claude would require a comparison case developed under equivalent conditions, but no such data has been presented.

Even if improved muscle function is confirmed in disease-model mice, the same effect will not necessarily be seen in human patients. Disease processes and responses to drugs differ between animal models and humans, and candidates that look promising in animal studies often fail to show sufficient efficacy in clinical trials. For this report too, detailed data need careful evaluation, such as how close the treated mice's muscle function came to that of healthy mice and whether survival was affected. The outlook of starting a trial in the second half of 2028 also assumes that upcoming preclinical safety studies, toxicity assessments, and manufacturing preparations proceed as planned.

Drug discovery is getting more efficient, but clinical trials remain a wall

The results Anthropic announced show that AI can support concrete tasks across a wide range of drug discovery steps. Results were reported not only for generating candidate molecules but also for analyzing experimental results, improving measurement methods, reviewing manufacturing records, and coordinating with outside research institutions, work on which humans have traditionally spent much of their time. However, this does not prove that the success rate of drug development has risen substantially or that the overall development timeline has shortened.

In the collaboration with pharmaceutical giant Bristol Myers Squibb (BMS), Claude was used to plan early-stage clinical trials and estimate their probability of success. Initial analysis that previously took several days was reportedly completed quickly through prompted instructions. However, no controlled comparison data have been presented on whether using the AI's analysis actually raised the success probability of clinical trials or by how much it shortened the overall drug development timeline. At Pleiades, a new company established by The Column Group (TCG), validation has also begun on an AI skill that evaluates combinations of drug targets and diseases, but whether the 45 combinations evaluated will lead to actual drug development outcomes is left to future research.

In evaluating these results, the relationship between Anthropic and each research organization also deserves attention. The company provided Generate Biomedicines and Divergence Labs with model usage credits and engineering support, and covered part of Divergence Labs' computing costs. It also paid for BigHat Biosciences' experiments and acquired Mithril Medicine outright. Such support and capital ties do not mean the reported experimental results are inaccurate. Still, in assessing research results, independent replication and verification by third parties matter, in addition to reports from the developers themselves.

The results include a roughly 10-fold increase in CAR receptor expression, a binding affinity improvement of 10x or more in certain antibody variants, and an 89% cut in manufacturing record audit time. But these are improvements in individual steps, measured under different experimental conditions and work environments. For antibodies and cell therapies in particular, even excellent in vitro performance leaves open problems such as unexpected toxicity in the body, effects on targets other than the intended one (off-target effects), and differences in drug response between patients. Clinical trials must also confirm long-term safety and efficacy, and even if molecular design and experimental preparation get faster, the time needed for this validation cannot simply be cut in the same way.

At the same time, this research has concretely shown that AI may be able to reduce waste and rework in the early stages of drug discovery. In BigHat's case especially, Claude took part in an important step of scientific research: not merely generating candidate molecules but also spotting a problem in how measurement data were interpreted and verifying it through additional experiments. Still, most of the results shown so far are at the preclinical stage, and the efficacy and safety of these candidates as medicines have not been confirmed.

The focus going forward is whether these candidate molecules can show efficacy and safety in animal studies and then deliver clear benefits to patients in well-designed clinical trials. Only when they lead to clinical outcomes such as longer survival and improved quality of life can the value of AI-assisted drug discovery be judged from a medical standpoint. The announcement shows that AI has begun to change the practice of drug discovery, but it also highlights that a long period of further validation is still needed before these results reach patients as actual treatments.