EPFL scientists have presented an Artificial Intelligence system that simplifies and automates scientific research processes

Apply the enormous potential ofArtificial intelligence Chemistry isn't easy. In this field of research, the processes are so complex that the most sophisticated calculation tools are often left aside as they can only be operated by extremely specialized figures. And i Large Language Models (LLM) like GPT 4.0 they are not as performing as in other contexts.

Now, however, EPFL researchers have developed Artificial Intelligence that simplifies the processes of chemical research, making it more efficient for both experts and beginners. Is called ChemCrow, and integrates 18 advanced tools for tasks ranging from organic synthesis to the discovery of new drugs.

From EPFL a new multimodal model for more flexible AI
Development of new drugs, libraries enhanced thanks to chemistry

Chemistry and the need to overcome the many limitations of GPT-4

In recent years, large language models (LLMs) have completely transformed some industries work and research: just think of the introduction of tools like GitHub Copilot or StarCoder, which promise to revolutionize developers' tasks and perspectives.

In chemistry all this hasn't happened yet. In reverse, LLMs often struggle to carry out seemingly simple tasks such as basic chemical operations. For example, we read in the EPFL study published in the journal “Nature Machine Intelligence”, "GPT-4 and GPT-3.5 are unable to consistently and accurately multiply 12.345 × 98.765 or convert IUPAC names to the corresponding molecular graph".

To overcome these severe limitations, scientists have begun to supplement LLMs with gods external plugins specifically designed for these tasks, such as OPSIN, which allows the language model to convert IUPAC names (those that indicate the number of atoms of each species) into structural formulas. Until now, the intrinsic shortcomings of LLMs have only been compensated thanks to tools of this type, which have improved their performance and applicability.

Furthermore, there is the complexity of these tools computational: although they are mostly open-source and accessible via API, their integration poses considerable challenges to experimental chemists, who should acquire increasingly sophisticated computational skills distant from the scope of their research.

Artificial Intelligence also for the development of new drugs
Artificial intelligence and holograms: the new frontier of healthcare

ChemCrow, the Artificial Intelligence that automates chemistry

The levels of automation achieved in chemistry are relatively low compared to what happens in other sectors. This is mainly due to highly experimental nature of the discipline, but also to the lack of data and the fact that computational tools have a limited scope.

These "impediments", however, could be overcome thanks to ChemCrow, the AI ​​system presented by Philippe Schwaller's group at the Federal Polytechnic of Lausanne, Switzerland, which integrates 18 advanced tools for activities such as the design and synthesis of drugs and materials.

ChemCrow was developed by Andres Bran and Oliver Schilter (EPFL, NCCR Catalysis) in collaboration with Sam Cox and Professor Andrew White (FutureHouse and University of Rochester).

It is based on an LLM such as GPT-4, which scientists have integrated with a suite of specialized software tools already used in chemistry, including WebSearch, LitSearch and various molecular and reaction instruments for chemical analysis.

With these tools at its disposal, ChemCrow has proven capable of autonomously planning and executing chemical syntheses such as creating an insect repellent and various organocatalysts, and even assist in discovery of new chromophores, fundamental substances for the dyes industry.

Furthermore, we read in the paper, “ChemCrow can autonomously solve reasoning tasks in chemistry, ranging from simple drug discovery cycles to planning the synthesis of substances with a wide range of molecular complexities”, which shows its potential as future chemical assistant à la ChatGPT.

GPT-3, the algorithm that writes like a human arrives in Italy
Water, grass and humanity: the cognitive limits of Artificial Intelligence

A simpler and more democratic chemistry now and for the future

What sets ChemCrow apart is its ability to adapt and apply a structured reasoning process to chemical tasks.

"The system is analogous to a human expert with access to a calculator and databases that not only improve the expert's efficiency, but also make him more concrete, in the case of ChemCrow, reducing hallucinations“, He explains Andres Camilo Marulanda Bran, first author of the study.

ChemCrow receives a request from the user, plans in advance how to solve the task, selects the relevant tools and refine iteratively its strategy based on the results of each phase.

This methodical approach ensures that the system is not only based on theory, but also on practical application for real-world interaction with laboratory environments.

Democratizing access to complex chemical knowledge and processes, ChemCrow lowers the barrier of entry for non-experts, while increasing the toolkit available to veteran chemists. This can speed up the search and development in the pharmaceutical field, in materials science and other fields, making the process more efficient and safe.

Artificial Intelligence and the climate crisis: opportunity or threat?
In Alto Adige today EDIH NOI is the new point of reference for AI