Introducing ChemCrow, where researchers extend large language models with chemistry tools

AI and ML Jobs


Source: https://arxiv.org/pdf/2304.05376.pdf

The automation of natural language processing brought to us by Language Language Models (LLMs) over the past few years has had a far-reaching impact on many industries. It is currently being applied to various NLP applications with excellent few-shot and zero-shot results. Recent advances have been made based on the Transformer architecture originally developed for neural machine translation.

Still, it’s important to remember that LLM has its limitations, making it difficult to learn things like elementary arithmetic and chemical calculations. The basic structure of the model, which focuses on predicting future words, is responsible for these shortcomings. One way he overcomes these limitations is to supplement his extensive language model with additional third-party software.

Artificial intelligence (AI) systems designed by experts to tackle specific problems are impacting the field of chemistry, especially reaction prediction, retrosynthetic planning, molecular property prediction, materials design, and most recently Bayesian optimization. . LLMs that generate code have demonstrated some understanding of chemistry due to the nature of their training12. The highly experimental and sometimes artisanal nature of chemistry limits the scope and applicability of computational tools even within designated areas. Tools such as RXN for Chemistry and AIZynthFinder are examples of closed setups where integration is common, made possible by corporate mandates to prioritize integration and internal use.

πŸš€ Check out 100’s of AI Tools at the AI ​​Tools Club

Researchers at the Laboratory for Artificial Chemical Intelligence (LIAC), the National Center for Research Competence (NCCR) Catalyst, and the University of Rochester developed the LLM-powered chemical engine ChemCrow, inspired by similar successful applications in other fields. Announced. It aims to simplify the reasoning process for many typical chemical operations in fields such as pharmaceutical and materials design and synthesis. By providing his LLM (GPT-4 in trial version) with task-specific and format-specific prompts, ChemCrow leverages the capabilities of a wide range of chemistry-specific, expert-designed tools. LLMs are given a list of tools, a brief description of their purpose, and information about data input and output.

The model is instructed to use thoughts, actions, behavioral inputs, and observed patterns. So you need to think about the current state of the task and how it relates to the end goal, and plan how to proceed. Concurrently with this preprint, 46 details a similar strategy for equipping LLM with chemistry-specific functionality, but without which it is beyond its scope. LLM then prompts for an action and input for this action (keyword “inference-based action completed in the thought step. After a short break, the text generator resumes searching for the appropriate function to apply). The results are prepended with the phrase “Observed” and sent back to the LLM, which repeats the previous step of “Think”.

Thus, LLM evolves from a confident source of information, albeit sometimes erroneous, to a thought engine that observes, reflects, and takes appropriate action based on what is learned. Researchers have deployed 13 different tools to aid research and discovery. The team acknowledges that the specified toolset is not comprehensive. Simply provide a tool, describe its intended purpose in natural language, and easily extend it to new uses. ChemCrow empowers professional chemists and those without specialized training in the field by providing a user-friendly interface to authoritative chemical information.

This paper evaluates the capabilities of ChemCrow across 12 different usage scenarios, including target molecule synthesis, safety control, and discovery of compounds with similar mechanisms of action. LLM-based assessments found that GPT-4 and ChemCrow had roughly equivalent effects on completeness and quality of thought. In contrast, human evaluations found ChemCrow to outperform his GPT-4 by nearly 4.4/10 points and by a significant 2.75/10 points in task completion success rate.


Please check paper.don’t forget to join 19,000+ ML SubReddit, Discord channeland email newsletterShare the latest AI research news, cool AI projects, and more. If you have any questions regarding the article above or missed something, feel free to email us. Asif@marktechpost.com

πŸš€ Check out 100’s of AI Tools at the AI ​​Tools Club

Tanushree Shenwai is a consulting intern at MarktechPost. She is currently pursuing her bachelor’s degree at the Indian Institute of Technology (IIT), Bhubaneswar. She is a data her science enthusiast and has a keen interest in the range of applications of artificial intelligence in various fields. She is passionate about exploring new advances in technology and its practical applications.

πŸš€ Transform Selfies into AI Generated Headshots: Try the #1 AI Headshot Generator Today



Source link

Leave a Reply

Your email address will not be published. Required fields are marked *