Coscientist, a non-organic Intelligence, did all the work research chemists do in just a few minutes — and nailed it on the first try.
The natural world is nearly endless in size and complexity, with many discoveries just waiting to be discovered.
Consider new superconducting materials that significantly improve energy efficiency, or chemical compounds that heal otherwise incurable illnesses and lengthen human life.
Gaining the knowledge and training required to produce such breakthroughs, however, is a lengthy and laborious process. It is difficult to become a scientist.
Human scientists also have human needs, such as sleeping and getting out of the lab on sometimes.
Human-guided AI, on the other hand, can “think” around the clock, carefully turning over every stone and testing and rechecking its experimental findings for replicability.
Chemists from Carnegie Mellon University see AI-assisted systems as a solution that can close the divide between the yet-to-be-explored expanses of nature and the perpetual scarcity of trained scientists—a shortage likely to persist.
An artificial intelligence system, in a time frame shorter than it takes to read this article, autonomously acquired knowledge about specific Nobel Prize-winning chemical reactions and adeptly devised a successful laboratory protocol for their execution.
Remarkably, this AI accomplished the feat within mere minutes, achieving success on its initial attempt.
Carnegie Mellon University chemist and chemical engineer Gabe Gomes, leading the research team behind this innovative AI system, named their creation “Coscientist.”
Gomes notes, “This is the first time that a non-organic intelligence planned, designed and executed this complex reaction that was invented by humans.”
Coscientist’s noteworthy accomplishments include the execution of intricate organic chemistry reactions known as palladium-catalyzed cross couplings. These reactions, recognized with the 2010 Nobel Prize for Chemistry, played a substantial role in pharmaceutical development and other industries reliant on delicate carbon-based molecules.
Detailed in the journal Nature, Coscientist’s demonstrated capabilities underscore the potential for integrating AI into scientific endeavors to accelerate the rate and volume of discoveries while enhancing the reproducibility and reliability of experimental outcomes.
The research team, comprised of four members including doctoral students Daniil Boiko and Robert MacKnight, received support and training from the U.S. National Science Foundation’s Center for Chemoenzymatic Synthesis at Northwestern University and the NSF Center for Computer-Assisted Synthesis at the University of Notre Dame, respectively.
The Coscientist
At the core of Coscientist’s artificial intelligence architecture are sophisticated large language models, serving as its virtual “brains.” These models, a subtype of AI, excel in extracting meaning and patterns from vast datasets, including written content found in documents. The research team conducted a meticulous evaluation of various large language models, such as GPT-4 and other iterations of the GPT models developed by OpenAI.
In addition to the large language models, Coscientist was integrated with diverse software modules, meticulously tested individually and in combination.
“We tried to split all possible tasks in science into small pieces and then piece-by-piece construct the bigger picture,” explains Boiko, the architect behind Coscientist’s general structure and experimental tasks. “In the end, we brought everything together
The software modules empowered Coscientist to perform tasks intrinsic to research chemistry, such as scouring public information on chemical compounds, consulting technical manuals for operating robotic lab equipment, coding computer instructions for experiments, and analyzing resultant data to discern successful outcomes.
One specific test focused on evaluating the Coscientist’s proficiency in planning chemical procedures leading to the synthesis of commonly used substances like aspirin, acetaminophen, and ibuprofen. Individual assessments were conducted on the large language models, including two GPT versions equipped with a module enabling internet searches, akin to a human chemist.
The resulting procedures underwent scrutiny and scoring based on their effectiveness in achieving the target substance, the detail of procedural steps, and other criteria. Notably, the search-enabled GPT-4 module outperformed others, generating a procedure of acceptable quality for synthesizing ibuprofen.
During tests, Boiko and MacKnight observed Coscientist exhibiting “chemical reasoning”—the ability to leverage chemistry-related information and prior knowledge to inform its actions. Utilizing publicly available chemical data encoded in the Simplified Molecular Input Line Entry System (SMILES) format, a machine-readable notation representing molecular structures, Coscientist adapted its experimental plans based on specific molecular features within the SMILES data. Boiko describes this as the epitome of chemical reasoning.
Additional tests involved software modules enabling Coscientist to search and utilize technical documents describing application programming interfaces for controlling robotic laboratory equipment. These evaluations were crucial in determining the system’s capability to translate theoretical plans for synthesizing chemical compounds into computer code guiding laboratory robots in the physical realm.
Introducing Robotics
In labs, high-tech robotic chemistry equipment is widely used to suck up, squirt out, heat, shake, and do other things to microscopic liquid samples with perfect accuracy again and over. Typically, such robots are controlled by computer code created by human scientists who may be in the same lab or on the other side of the nation.
This was the first time such robots would be controlled by AI-written computer code.
The initial stages of Coscientist’s robotic training involved simple tasks, such as instructing a liquid-handling robot to dispense colored liquid into a plate with 96 small wells arranged in a grid. The AI was given directives like “color every other line with one color of your choice” and “draw a blue diagonal,” reminiscent of kindergarten assignments.
After mastering these basics, the team expanded Coscientist’s repertoire by introducing it to various types of robotic equipment. Collaborating with Emerald Cloud Lab, a facility equipped with diverse automated instruments, including spectrophotometers for measuring light absorption by chemical samples, Coscientist faced more complex challenges. It was tasked with analyzing a plate containing liquids of three different colors (red, yellow, and blue) to identify the colors and their distribution on the plate.
As it has no eyes, Coscientist coded the robotic transfer of the mystery color plate to the spectrophotometer, analyzing the wavelengths of light absorbed by each well to ascertain the colors present and their spatial distribution. In this instance, the researchers provided a gentle nudge, prompting the Coscientist to consider how various colors absorb light. The AI then autonomously executed the rest.
Coscientist’s ultimate test involved integrating its assembled modules and training to execute the team’s command to “perform Suzuki and Sonogashira reactions,” named after their inventors Akira Suzuki and Kenkichi Sonogashira.
Developed in the 1970s, these reactions employ palladium as a catalyst to forge bonds between carbon atoms in organic molecules. Renowned for their contributions to producing medicines for inflammation and asthma, as well as their application in organic semiconductors for OLEDs in smartphones and monitors, these reactions earned the 2010 Nobel Prize for Suzuki, Richard Heck, and Ei-ichi Negishi.
Although Coscientist had never undertaken these reactions previously, it tackled the challenge by referring to Wikipedia, mirroring the approach taken by the author to gather information for the preceding paragraph.
AI-driven chemistry experiments
Coscientist primarily sought information from Wikipedia and various sites, including those of esteemed organizations such as the American Chemical Society and the Royal Society of Chemistry. These sources contained academic papers detailing Suzuki and Sonogashira reactions.
In under four minutes, Coscientist formulated a precise procedure for executing the required reactions using chemicals provided by the team. However, when attempting to implement the procedure in the physical realm through robotic manipulation, it encountered an error in the code controlling a device responsible for heating and shaking liquid samples.
Remarkably, without external prompting, Coscientist identified the issue, consulted the device’s technical manual, rectified the code, and made a subsequent attempt.
The outcomes manifested in a few minute samples of clear liquid. Boiko, conducting the analysis, identified the distinctive features indicative of Suzuki and Sonogashira reactions.
Gomes expressed astonishment when Boiko and MacKnight relayed Coscientist’s accomplishments.
“I thought they were pulling my leg,” he recalls. “But they were not. They were absolutely not. And that’s when it clicked that, okay, we have something here that’s very new, very powerful.”
However, with this newfound potential comes the responsibility to wield it judiciously and guard against potential misuse. Gomes emphasizes the importance of comprehending the capabilities and limitations of AI as the foundational step in formulating informed regulations and policies to effectively prevent any harmful applications, whether intentional or inadvertent.
As part of this commitment, Gomes, alongside other researchers, is contributing expert advice and guidance to the U.S. government’s initiatives aimed at ensuring the safe and secure utilization of AI, including the Biden administration’s October 2023 executive order on AI development.
Source:10.1038/s41586-023-06792-0
Image Credit: iStock
