Natural language processing automation has made significant advancements in recent years, impacting various industries. One of the key technologies behind this is Language Language Models (LLMs), which have been applied to different NLP applications with impressive results. However, LLMs have their limitations, particularly when it comes to tasks like arithmetic and chemical calculations. These limitations arise from the structure of the models, which focus on predicting upcoming words. To overcome these restrictions, additional third-party software can be integrated with extensive language models.
In the field of chemistry, expert-designed artificial intelligence (AI) systems have made a significant impact. These systems have been used in areas such as reaction prediction, retrosynthesis planning, molecular property prediction, materials design, and Bayesian Optimization. Code-generating LLMs have a degree of understanding of chemistry due to their training. However, the limitations of computational tools and the artisanal nature of chemistry require the integration of closed settings like RXN for Chemistry and AIZynthFinder. These integrations are made possible by corporate mandates prioritizing internal use.
Introducing ChemCrow: LLM-powered Chemistry Engine
The Laboratory of Artificial Chemical Intelligence (LIAC), National Centre of Competence in Research (NCCR) Catalysis, and the University of Rochester have developed ChemCrow. It is an LLM-powered chemistry engine that simplifies the reasoning process for typical chemical tasks in areas like drug and materials design and synthesis. ChemCrow supplements LLMs, such as GPT-4, with task- and format-specific prompts and leverages chemistry-specific expert-designed tools. The LLM is given a list of tools, their purpose, and information regarding data input and output.
ChemCrow incorporates the Thought, Action, Action Input, and Observation pattern to guide the model’s decision-making process. It considers the present state of the task, its relation to the end objective, and plans the next steps accordingly. The model asks for an action and its input based on its previous thought process. After a short break, the text generator resumes its search for an appropriate function to apply to the given data. The result is sent back to the LLM with an “Observation” tag, and the process repeats.
The researchers deployed thirteen different tools to aid in research and discovery. While the given toolset is not comprehensive, it can easily be extended by adding new tools and describing their purpose in natural language. ChemCrow provides a user-friendly interface to reliable chemical information, benefiting both professional chemists and those without specialized training in the field.
Evaluation of ChemCrow’s Features
The features of ChemCrow were evaluated across twelve different use scenarios, including synthesizing a target molecule, safety controls, and finding compounds with similar modes of action. An LLM-based evaluation found that GPT-4 and ChemCrow were nearly equally effective in completeness and quality of thought. However, human evaluations demonstrated that ChemCrow outperformed GPT-4 significantly in terms of successful task completion.
For more details, you can read the full paper here.
Don’t forget to join our ML SubReddit, Discord Channel, and subscribe to our Email Newsletter for the latest AI research news and projects. If you have any questions or if there’s anything we missed, feel free to email us at Asif@marktechpost.com.
Check out more AI tools in AI Tools Club here.