Bum Chul Kwon | 권범철 | @[email protected]

Multi-view biomedical foundation models for molecule-target and property prediction @IBMResearch • The paper introduces MMELON, a multi-view molecular foundation model combining graph, image, and text views to enhance prediction of molecular properties. Unlike single-view models, MMELON leverages multiple representations for a richer, more versatile molecular embedding. • The model performs exceptionally well on 18 diverse tasks, including ligand-protein binding, molecular solubility, metabolism, and toxicity, balancing the strengths of each modality. This versatility is critical in drug discovery and computational chemistry. • MMELON integrates three views—graph, image, and text—to learn comprehensive molecular representations. The image view uses ImageMol (pre-trained on 10 million molecules), while the graph and text views are based on advanced transformer architectures, pre-trained on datasets of 200 million molecules. • A novel aspect is the “late fusion” of these different modalities, ensuring each modality contributes optimally depending on the downstream task. This approach yields interpretable results and allows for an analysis of how each view supports different predictions. • For validation, MMELON was applied to screen compounds against a large set of G Protein-Coupled Receptors (GPCRs). Of these, 33 GPCRs related to Alzheimer’s disease were identified, and strong binders were predicted, validated through in silico structure modeling. • The multi-view model shows strong correlations between predicted and experimental affinities, achieving a Pearson correlation of 0.78 for GPCR binding. This suggests the model’s robust application for identifying new therapeutics. • Compared to single-view models, MMELON delivers superior performance across classification and regression tasks, making it an essential tool for complex molecular property predictions in drug discovery. @jamorrone3 @jianying_hu @FeixiongCheng @jeriscience @BCKwon @timrumbell @dplatt_maths @YunguangQiu @diwakarmahajan 💻Code: github.com/BiomedSciAI/bi… 📜Paper: arxiv.org/abs/2410.19704 #biomedicalAI #drugdiscovery #foundationmodel #multiviewlearning #GPCR #Alzheimers #machinelearning #bioinformatics

What is visualization literacy? How can we measure it? How can we improve it for everyone? Submit your work to our CHI 2024 workshop by Feb 29 and join our discussions on defining, studying, and enhancing visualization literacy for all. visualization-literacy.github.io/CHI2024/

2 GT students have been named 2023 IBM Ph.D. Fellows, and one of them is our own @graceguo43! Out of hundreds of applications, IBM awarded only 10 Fellowships for 2023. Congrats Grace!

How can we uncover hidden biases in language models that impact fairness? Our #ACL2023 demo paper introduces Finspector, an interactive visualization widget available as a Python package for Jupyter. Paper, Video, Code: bckwon.com/publication/fi… @nandanamihindu #nlp #fairness

Chart captioning is hard, both for humans & AI. Today, we’re introducing VisText: a benchmark dataset of 12k+ visually-diverse charts w/ rich captions for automatic captioning (w/ @angie_boggust @arvindsatya1) 📄: vis.csail.mit.edu/pubs/vistext.p… 💻: github.com/mitvis/vistext #ACL2023NLP