Event Details
- This event has passed.
HMS Town Hall- AI for Protein Structure Prediction
March 5 @ 10:00 am - 1:00 pm
The Center for Computational Biomedicine (CCB) is hosting a town hall to discuss recent
developments in artificial intelligence for protein structure prediction and its applications
in biology and medicine.
If you were able to attend, please share your feedback by submitting a survey.
Recording now available!
Topics
Missense variant effect prediction with AlphaMissense
Protein complex prediction with AlphaFold-Multimer
Large language models for predicting post-translational modifications
AI foundation models for protein design and drug discovery
Schedule (all times ET):
Talk: Accurate proteome-wide missense variant effect prediction with AlphaMissense
Talk: Large Scale Protein-Protein Interaction Screening with Alphafold Multimer
Talk: Foundational AI Models for Biological Systems
Talk: BioPlex– Predicting structures for protein-protein interactions across the human interactome
Talk: Multi-Modal and Generative Models for Drug Design
Talk: Equifold– Fast Antibody Structure Prediction for Drug Discovery and More
After Town Hall, 1-2 pm
Josh Pan Silva and Edward Huttlin will discuss user stories and use cases of AlphaMissense, AlphaFold, and Bioplex. You are welcome to stay in the same room using the same Zoom link and join the discussion!
Speakers
Talk Summary– AlphaMissense is a deep learning tool that utilizes protein structure predictions and biological sequence data to predict whether single-amino acid changes in proteins will have deleterious effects.
Talk Summary– In this talk, I will share my recent work along the direction of foundational AI models for biological systems and introduce the 100-billion parameter xTrimo family of pretrained models leveraging billions of data points from protein sequences, structures, protein-protein interactions, and single-cell transcriptomics. The pretrained models can be used as the foundation to address many predictive tasks arising from protein design and cellular engineering and achieve SOTA performances.
Talk Summary– EquiFold is a protein structure prediction model that relies on a novel frame representation and geometric deep learning models. Its scalability makes it suitable for high-throughput applications including analysis of large NGS repertoires and MD-simulation like ensemble generation.
Talk Summary– In this talk, I describe our AI research to advance molecular drug discovery. Large language models and generative AI are transforming drug design. Instead of training separate models for every task across molecular modalities, we can now adapt pretrained large foundation models to many tasks through fine-tuning and few-shot prompting. Central to our approach is the integration of molecular structures, biological knowledge, and genomic data into AI models. We are advancing self-supervised learning to leverage vast unlabeled datasets and geometric deep learning to model the geometry of biochemical data. I will describe our multimodal sequence-structure generative models that help design molecules to maximize binding affinity with biological targets, serve as optimal binders, and have specific biochemical properties. For drugs to be effective, they must act on biological targets in relevant biological contexts. I describe PINNACLE AI models for precise and cell-type specific protein representation learning. PINNACLE models can be fine-tuned to enhance 3D structures of protein-protein interactions in immune oncology, predict drug effects across cell types and cell states, and nominate therapeutic targets in a cell-type-specific manner.
Talk Summary– Combining our BioPlex network of human protein interactions with AlphaFold structural models of >100,000 interacting protein pairs reveals functional insights into myriad complexes, proteins, and post-translational modifications.
Ernst completed his undergraduate studies at the University of California, Los Angeles majoring in molecular biology. After initially focusing on wet lab biology, he transitioned to computational research upon joining the Walter lab as a PhD student in 2021. His primary focus has been on developing in-silico screening platforms to find novel protein protein interactions in biological pathways.
Talk Summary– Protein-protein interactions (PPIs) underly nearly all biological processes including genome maintenance. Although existing experimental methods have attempted to comprehensively survey these interactions, many remain elusive and/ or poorly characterized. The recent development of in-silico protein structure prediction algorithms such as AlphaFold multimer and others offer a potential solution by unlocking new ways to reveal novel interactions at unprecedented speed and scale. In order to take advantage of these new deep learning based methods, we have built an internal software pipeline for quickly generating and evaluating hundreds of thousands of AlphaFold multimer structures. This platform has enabled us to perform several large interaction screens including a pairwise all by all screen of 286 human genome maintenance proteins that we have made public via web platform on predictomes.org.
Flyer
