Claude assists in resolving the prokaryotic Vault structure: AI is further integrated into the biological experimental workflow.
On November 5, Anthropic announced a new achievement of its biology research team: with the help of Claude, researchers have for the first time resolved the complete 3D structure of Vault (vault ribonucleoprotein particles) related to prokaryotes. The research covers multiple links including protein database retrieval, experimental protocol design, protein purification and cryo-electron microscopy image analysis. Among them, Claude helped the team shorten the purification time of Vault particles from about 3 days required by the traditional process to 90 minutes, and the first structural model was completed within 34 hours.
On October 27, Anthropic disclosed that it has cooperated with six drug R&D institutions to explore the application of Claude in different links of drug development. This Vault research further demonstrates its application path in basic biological experiments.
This study obtained a previously incompletely resolved structure of protein complex, and also demonstrated the specific way AI participates in biological research: The model can connect data retrieval, experimental design and result analysis, and participate in the real experimental process under the guidance of researchers. This capability is expected to be further extended to experimental iteration links in target research, protein engineering and drug R&D.
Finding Vault from massive protein sequences
Vault is a large vesicle-like protein complex that exists in almost all human cells. Since it was discovered in 1986, its specific biological functions have not been fully explained. Researchers have proposed that it may be involved in processes such as intracellular substance transport and cell protection, but there is no final conclusion at present.
Previously, researchers have found protein sequences evolutionarily related to Vault proteins in prokaryotes and their related viruses, but there is still a lack of direct evidence on whether these proteins can assemble into a complete Vault structure.
In this study, Anthropic chose to start with bacteriophages. Bacteriophages are a type of viruses that infect bacteria, and their proteins need to be expressed with the help of bacterial hosts. The research team hopes to find candidate proteins that can form the Vault structure, and take advantage of the characteristics of prokaryotes with faster growth rate and relatively shorter experimental cycle to establish an experimental system that is easier to carry out research.
In this process, Claude participated in large-scale protein sequence retrieval and candidate screening. The research team searched for targets from about 15 billion predicted protein sequences, then selected 44 candidate proteins for experimental verification, and finally confirmed that the L05 protein from bacteriophage can assemble into Vault particles.
Researchers expressed the L05 protein in Escherichia coli and resolved its structure by cryo-electron microscopy (cryo-EM). The results show that this Vault is composed of 96 identical protein copies, with a total molecular weight of about 9.5 million Daltons, making it one of the larger protein shells in prokaryotes reported so far.
As a control, human Vault consists of 78 major Vault protein copies. The two differ in size, internal space and repeated protein structures.
Claude participates in the experimental process to shorten the structure resolution cycle
The more noteworthy part of this study is that Claude participated in multiple links from computational analysis to experimental operation scheme design.
In the early stage of the study, Claude assisted in searching protein databases, designing DNA constructs, and helping researchers determine the experimental verification path for candidate proteins. After entering the experimental stage, Claude participated in optimizing the protein separation and purification methods.
One key improvement comes from the purification process. The traditional method requires relatively cumbersome sample processing and usually takes about 3 days. Claude proposed that the mixture after cell disruption can be directly loaded onto the chromatography column, omitting some pre-purification steps. The researchers evaluated and experimentally verified this scheme, and finally compressed the process from cell sample to cryo-EM grid preparation to about 90 minutes.
In the stage of structure resolution, Claude also participated in processing cryo-EM data. The research team used relevant software to analyze a large number of particle images, and gradually reconstructed the three-dimensional structure of the protein complex. Anthropic disclosed that the first complete structural model was generated in 34 hours, while according to the team's estimate, the traditional process usually takes several weeks to complete this stage of work.
After obtaining the complete structure, the research team further verified the assembly mechanism of Vault through protein variant experiments. The team designed 15 L05 protein variants, most of which were modified for the contact regions of adjacent protein molecules. Experiments found that a single point change at the top of some particles is enough to prevent the shell from closing normally, turning the originally closed Vault into an open garland-like particle.
The researchers also changed the size of Vault by adding repeated fragments inside the protein. The L05 protein originally contains 7 repeating units, the team copied two of the fragments, increased the repeating units to 9, and successfully assembled larger Vault particles, whose internal volume increased by about 20%. This experiment shows that the structure of Vault has a certain space for engineering modification, and also provides clues for understanding the possible size changes of Vault in the process of evolution.
The boundary of AI participating in scientific research is expanding, but accurate professional judgment is required
This study not only resolved the three-dimensional structure of Vault, but also Claude helped researchers identify a previously unnoticed small molecule pocket inside the protein. Cryo-EM maps show that there is a small molecule in it that is similar in shape to coenzyme A but whose identity has not yet been determined. This provides clues for further research on the biological functions of Vault.
Anthropic believes that the hollow structure of Vault is expected to be developed as nanocapsules in the future for delivering biomolecules such as enzymes. However, the formation mechanism and specific biological functions of Vault in natural hosts are not yet clear, and researchers have not proved that it undertakes the role of material transport or delivery in the natural environment. Whether it can be transformed into a stable and controllable nanocarrier and realize effective loading and delivery of biomolecules such as enzymes in the future still requires further engineering research and experimental verification.
Anthropic said that this study demonstrates a new way of scientific research collaboration. Its team pointed out that the members previously lacked research experience in large protein complexes and had limited knowledge reserve in structural biology. Claude helped the team advance the research tasks they were not familiar with by providing professional knowledge retrieval, experimental protocol design and data analysis support, and quickly completed the analysis work that would otherwise take a longer time. The entire project took less than four months from candidate screening, sample preparation to the completion of the first prokaryotic Vault structure resolution.
However, when AI participates in experimental science, Claude cannot give correct answers every time. During the research process, Claude once proposed a medium scheme containing components that interfere with DNA entering cells, causing a two-week delay in the project; in the early cryo-EM data processing, it also skipped standard steps and generated problematic structural maps. Researchers need to identify and correct these errors in time.
The Anthropic team fed back experimental records to model trainers to help them understand Claude's errors in scientific research tasks and improve model training data. With the successive launch of new versions of models, Claude's performance in related tasks has improved, the errors in DNA construct design have decreased, and its ability to identify potential experimental risks has also improved.
Therefore, this scientific research collaboration method still relies on the professional judgment of researchers. The Anthropic team concluded that in familiar research fields, researchers can quickly identify Claude's errors; while in links with relatively insufficient knowledge reserve, these errors may bring additional time costs.
In other words, Claude did not help the research team eliminate the risks brought by insufficient professional judgment.
In the past, discussions on AI drug R&D mostly focused on computational tasks such as protein structure prediction, molecular generation, target identification and candidate compound screening. In this study, AI focused on a specific scientific problem, continuously participated in literature and database retrieval, experimental protocol design, sample preparation method optimization and experimental data processing, connecting different links together.
For the AI pharmaceutical industry, the indicators worthy of observation in the future will also include whether the model can continuously improve experimental efficiency, improve the quality of R&D iteration, and whether it can help the research team obtain experimental results that were previously difficult to achieve.
This article is from the WeChat official account "Bright Company" (ID: suchbright), the author: Chief Editor 24h Online, published with authorization from 36Kr.