TURMIND

MareNostrum 5 Resource Pilot Access (TÜBİTAK ULAKBİM) Project Call 2024

The aim of this project is to develop a universal foundation model for molecular representation learning, capable of handling all molecules related to life, including biomolecules such as DNA, RNA, proteins, and small-molecule ligands and metabolites. The proposed foundation model will utilise cutting-edge deep learning techniques to encode atomic and molecular-level information, serving as the backbone for downstream tasks. By leveraging multi-modal data and state-of-the-art architectures, this project aims to advance machine understanding of biology and accelerate drug discovery and biotechnology innovation. The foundation model will adopt a multimodal learning approach to integrate diverse data types, including sequences (e.g., nucleotide and amino acid sequences), structures (e.g., atomic 3D coordinates), interactions (e.g., protein-protein or protein-ligand networks), textual annotations, and imaging data (e.g., immunohistochemistry). We will explore advanced transformer-based architectures as the primary modelling approach, including encoder-decoder models and dense attention mechanisms. Alternatives, such as graph neural networks (GNNs) and diffusion models, will be considered to capture the relational and spatial information inherent in molecular systems. This will enable the model to operate at the atomic level, using atoms and bonds as the fundamental building blocks to provide a unified representation for all molecules. Following pretraining, the model will be fine-tuned for generative tasks such as designing novel proteins with specified functions and drug candidates targeting biomolecules of interest. Generative tasks will utilise autoregressive or diffusion-based generative models. Discriminative tasks, such as protein function prediction and molecular property prediction, will leverage fine-tuned classifiers and regressors. Potential applications include enzyme design, metabolite synthesis, materials science, and drug discove