Authors:

Aakarsh Misra

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Understanding a chemical's intrinsic properties and characteristics in various environments is a critical step in the chemical discovery pipelines. Obtaining such properties can be challenging due their high costs and reliance on expertise for analysis. In recent years, Deep Learning (DL) has become frequently employed in obtaining fast, accurate predictions of various chemical properties. However, current paradigms lack predictive consistency, generalizability, and are unwieldy in data scarce scenarios. Large Language Models (LLMs) trained on chemical graphs and fine-tuned on bespoke property datasets have shown some promise. But there is no consensus that the training framework behind LLMs would correspond to accurate property prediction. Since many chemical properties are derived from fundamental properties, it is possible that finetuning bespoke models from models trained on fundamental datasets would lead to accurate predictions in both low and high data contexts. Therefore, we train a Large Property Model (LPM) based on our novel Edge-Featured Graph Attention Network (EGAT) architecture and 11.7 million chemicals pulled from PubChem - each with 23 fundamental properties collected from DL-driven quantum chemistry (QC) calculations. Initial benchmarking with finetuned LLM's on predicting the solvation energy and lipophilicity of molecules has shown relatively good performance compared to other techniques. Early training of the LPM model shows reasonable performance on withheld testing sets. If our hypothesis holds true, our findings can predict properties in large-data and data- scarce contexts for applications in drug discovery, material discovery, and environmental safety. Keywords: Chemical Property Prediction; Deep Learning; Large Property Models; Edge-Featured Graph Attention Neural Network; Large Language Models

Source:

Purdue University / 2024

Topics:

No topics listed

Co-authors:

Aakarsh Misra