Recent advancements in deep learning have enabled improvements in chemical data extraction, in silico drug discovery, reaction prediction, and data annotation. Our team's efforts address challenging issues like complex installation procedures, poor documentation or lack of maintenance by developing accessible, user-friendly, open-source software that bridges the gap between developers and end-users in cheminformatics, making advanced tools and resources more approachable for the broader scientific community.
Our development process emphasises collaboration through GitHub for version control and Zenodo for dataset hosting, ensuring transparency and reproducibility. To cater to users without programming expertise, we have created intuitive graphical user interfaces (GUIs) and web applications for tools like DECIMER (Optical Chemical Structure Recognition)¹ and STOUT (SMILES-IUPAC name translation), allowing for easy file upload, processing, and download. Recent developments include the Cheminformatics Microservice, STOUT V2.0², and COCONUT V2.0³, which integrate API endpoints for remote task execution, reducing the need for complex software installations. With TensorFlow Lite, deep learning models have been optimised for CPU performance, improving accessibility for researchers without high-performance computing resources. Moreover, all our work is fully containerised using Docker, allowing for seamless local setup and execution.
The current ongoing work looks into integrating large language models (LLMs) for text extraction from journal articles, aiming to develop end-to-end workflows for chemical data processing for Natural Products. This includes extracting structured, machine-readable data from PDFs using a single pipeline with minimal installation requirements. These advancements contribute to the open science movement in chemical data extraction, fostering collaboration and innovation in cheminformatics and related fields.

Achim Zielesny

Appearances