Shulou(Shulou.com)12/24 Report--

AI big model "chemist" is published on Nature! The kind that makes aspirin, acetaminophen and ibuprofen. Even complex palladium-catalyzed cross-coupling reactions can be completed!

You know, the winners of the 2010 Nobel Prize in chemistry won the prize for their research on the reaction, which can efficiently construct carbon-carbon bonds and produce a lot of substances that were difficult or even impossible to synthesize in the past.

The AI system, which is now called Coscientist and based on GPT-4 and other large models, can quickly and accurately complete a whole set of processes of retrieving information, planning and designing experiments, writing programs, remote control automation systems to do experiments, and analyze data.

A netizen who marked himself as a chemist on his home page said:

Suan Q, you have created more unemployed doctoral students.

So how on earth did Coscientist do it?

What does "chemist" Coscientist look like? Coscientist was developed by a research team at Carnegie Mellon University.

Not long ago, AI "chemist" made by Google DeepMind also appeared on Nature, claiming to be able to predict 2.2 million new materials in one breath. Now Coscientist can actually complete all the subsequent experimental processes on its own.

The key to accomplish such a complex experimental task lies in the system architecture of multi-module interaction.

Coscientist contains five modules: Planner, Web searcher, Code execution, Docs searcher and Automation.

Among them, Planner module is the intelligent center of the whole system, which is based on GPT-4 and is responsible for planning and promoting the whole experiment according to the input of users, calling and coordinating other modules.

Planner can issue four instructions: GOOGLE, PYTHON, DOCUMENTATION and EXPERIMENT.

The GOOGLE instruction is responsible for using the Web searcher module to retrieve information about the experiment on the Internet, and Web searcher itself is a large model.

The PYTHON instruction controls the Code execution module. Code execution is an isolated Docker container that provides an independent Python execution environment, which can complete the calculation work related to the experiment.

The DOCUMENTATION instruction controls the Docs searcher module, which is also used to provide information to the hub.

But unlike Web searcher, Docs searcher is used for text retrieval and document understanding. It can locate the technical documents of the experimental equipment, such as the manipulator programming manual, and provide the necessary experimental parameters and operation details to the Planner module through text mining.

Then, the Automation module is responsible for automatically connecting the API of the actual experimental equipment, converting the experimental scheme formulated by Planner into the equipment control code, issuing and executing, and completing the experimental operation. For example, experiments are carried out to remotely control the liquid transfer robot in the "cloud laboratory".

In this way, suppose that when Coscientist is asked to synthesize a substance, Coscientist will search the Internet for the synthesis route; then design the experimental scheme for the desired reaction; next write code to guide the liquid transfer robot; and finally run the code to make the robot perform its predetermined task.

It is worth mentioning that Coscientist can also carry out iterative optimization, learn from the response results, and propose to modify the scheme to improve the experiment.

Overall, Coscientist can accomplish six major tasks:

Planning the synthesis of known compounds based on open data

Effectively search and browse a large number of hardware documents

Use the information in the documentation to execute advanced commands in the cloud lab

Precise control of liquid processing instruments with low-level instructions

Dealing with complex scientific tasks that require simultaneous use of multiple hardware modules and integration of different data sources

The optimization problem is solved by analyzing the previously collected experimental data.

What is the Coscientist performance of palladium-catalyzed cross-coupling reaction successfully? The researchers tested several modules.

In order to test the ability of Coscientist to design chemical reaction process, the research team asked Coscientist to generate drug molecules such as aspirin, acetaminophen and ibuprofen through retrieval learning, and compared different models of GPT-3.5, GPT-4, Claude 1.3 and Falcon-40B-Instruct.

The GPT-4-based Web Searcher significantly improved the synthesis plan, achieving the highest scores in all trials of paracetamol, aspirin, nitroaniline and phenolphthalein (figure b above, the number "5" represents a very detailed and chemically accurate procedure).

The key point also depends on the integration ability of Coscientist. For this reason, the researchers also designed the catalytic cross-coupling experiment.

The researchers set up available experimental equipment, including an OpenTrons OT-2 liquid processing robot and several microboards, including a source microboard for placing reactants and a target microboard placed on a heating vibration module.

The source plate is prepared with reagents for the experiment, including hexyl iodine, bromobenzene, chlorobenzene, phenylacetylene, phenylboric acid and other raw materials, as well as catalysts, bases and solvents.

The goal of Coscientist is to successfully design and run two common palladium-catalyzed coupling reactions, Suzuki reaction and Sonogashira reaction, using these reagents.

Coscientist first uses Web searcher module search to determine the best reaction conditions of Suzuki reaction and Sonogashira reaction, such as temperature, equivalence ratio and other parameters.

Then different reagents were reasonably selected, for example, bromobenzene was preferred to chlorobenzene in Suzuki reaction. At the same time, Coscientist provides the chemical basis for selection, such as reaction activity.

Next, Coscientist calls the Code execution module to calculate the required volume according to the concentration and equivalent of each reactant. Generate the Python code that controls the robot to transfer liquid, and specify the transfer volume between the source orifice plate and the target orifice plate.

But there was an episode in the middle, and the original method used to heat the shock module was misnamed.

After that, Coscientist quickly consulted the Opentrons device documentation to correct the method name, regenerated the correct code, and successfully completed the Suzuki response and Sonogashira response.

Finally, the product was verified by GC-MS technology, and the characteristic mass spectrum signal of the target product was detected, which confirmed the formation of the target product.

Links to papers:

Https://www.nature.com/articles/s41586-023-06792-0

This article comes from the official account of Wechat: quantum bit (ID:QbitAI), author: Xifeng

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.