Lightning talks

About this talk

Lightning talks

La iniciativa Scientific Python
Oriol Abril Pla

Fa uns anys es va engegar la iniciativa Scientific Python, amb l'objectiu d'enfortir la col·laboració entre llibreries de Python en el camp de la computació científica. En els 5 minuts intentaré explicar què és i algunes de les seves iniciatives principals, que van des de gestió de dependències a consells d'accessibilitat passant per traduccions, 'nightly builds' o governança.

DrugScope: a ML-based platform for drug discovery
Verónica Larroy

DrugScope is a machine learning-powered tool designed to predict the binding affinity of small molecules to specific proteins, aiming to accelerate drug discovery and reduce costs. Using Graph Neural Networks (GNNs), the project models molecular structures as graphs, with atoms as nodes and bonds as edges. A public dataset from Leash Biosciences, containing around 300 million molecules and their interactions with three key proteins (HSA, BRD4, and SEH), was used for model training.

The project involved developing three separate GNN models, one for each protein, to capture the unique molecular interactions relevant to each target. Features for atoms and bonds were extracted using RDKit, and the data was prepared for training by converting these features into graph structures. The models were deployed using FastAPI, allowing researchers to input molecule data and receive predictions on binding affinity, potentially helping speed up the early stages of drug development.

Can LLMs Power Product Recommendations?
Babaniyi Olaniyi

I will discuss the use of large language models (LLMs) for product recommendation systems, leveraging their natural language understanding capabilities to predict items a user might purchase given their historical product reviews and purchases.

Pixel2Mechanics: Automated Biomechanical Simulations of High-Resolution Intervertebral Discs from Anisotropic MRIs
Sai Natarajan

This study presents Pixel2Mechanics, an automated pipeline for generating patient-specific biomechanical finite element (FE) models of lumbar intervertebral discs (IVD) from clinical MRI data. It utilizes a deep learning framework to create high-resolution meshes of the Annulus Fibrosus (AF) and Nucleus Pulposus (NP) across L1-L2 to L4-L5. The pipeline features a novel optimization method based on differentiable rendering and a Bayesian Coherent Point Drift++ morphing algorithm for volumetric FE meshes that maintain tissue topology. Load simulations validate the models against manual segmentation, demonstrating potential for clinical integration in diagnosing and treating IVD degeneration.