We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
GPT4 aided biomaterials research use case: stabilization of selenium nanoparticles with proteins.
- Authors
Stojanović, Zoran; Filipović, Nenad; Kuzmanović, Maja; Lukač, Sara; Stevanović, Magdalena
- Abstract
Recent advancements in LLMs based on various transformer architectures such as BERT and GPT family models, brought many new possibilities for application in scientific research. The specific architecture and broad knowledge of these models give them the ability to understand concepts, to plan and solve different kinds of problems, including various chemistry-related tasks. In this work, we are evaluating a case of GPT4 performance for recommending proteins suitable for the stabilization of selenium nanoparticles (SeNPs). SeNPs exhibit diverse beneficial bioactivities, including antioxidant, antibacterial, and anticancer properties, and stabilization of SeNPs with suitable proteins may be an effective approach to improve their bioactivities. Initially, we made a series of "zero-shot" tests to evaluate knowledge, problem-solving ability, and identify weaknesses of GPT4 on the research topic and subsequently, we optimized prompts if needed to get correct responses. Test questions in related domains of crystallography, colloidal chemistry, and biochemistry were selected from free sources online (e.g. sanfoundry.com) or created by us. Finally, we have evaluated model performances on the main task which is to suggest the best protein candidates for stabilization of SeNPs that we synthesized in laboratory experiments. The design of prompts was done according to proposed tactics for prompt engineering suggested by OpenAI and DAIR.AI. For interaction with GPT4 model OpenAI, API and Python programming language were employed. The model successfully completed all our benchmark tests with optimized prompts. It also demonstrated the ability to analyze procedures for the synthesis and stabilization of SeNPs. GPT4 exhibits the capability to recognize protein records by identifiers, correctly identify amino acid sequences, describe their properties and functions to some extent depending on data trained on, identify available binding sites for interaction with Se, and give, by our judgment, good proposals for SeNPs stabilizers. The study demonstrates the successful application of advanced transformer architecture models like GPT4 in addressing relatively complex tasks in materials research. Despite GPT4 capabilities being largely dependent on the quality and size of training data, utilization of strategically designed and optimized prompts significantly improves it's performance.
- Subjects
OPENAI Inc.; TRANSFORMER models; PYTHON programming language; AMINO acid sequence; BIOMATERIALS; SELENIUM; ELECTRIC transformers; PROTEINS
- Publication
Chemical Industry / Hemijska Industrija, 2024, Vol 78, p68
- ISSN
0367-598X
- Publication type
Article