To view previous webinars use the links below
2020 Webinars | 2021 Webinars | 2022 Webinars | 2023 Webinars | 2024 Webinars | 2025 Webinars
ISCBacademy is an online webinar series including the ISCB COSI, COVID webinars, Indigenous Voices and practical tutorials. We aim to inspire, connect, and communicate the science while providing a hands-on experience accessing and using newly developed bioinformatics tools while ensuring best practices for rigour and reproducibility.
May 2, 2025 at 11:00 AM EDT
A continuous phylogenetic tree space is a geometric space whose points correspond to phylogenetic trees with branch lengths. Such a space provides a framework for analyzing both the tree topology and branch lengths together, which is important for processes such as the multi-species coalescent which depend on tree branch lengths. This talk will introduce the most well-known such space, the Billera-Holmes-Vogtmann (BHV) tree space. BHV tree space has several nice properties, including unique shortest paths (geodesics) between pairs of trees, and a unique tree (the Fréchet mean) minimizing the sum of the squared distances to a set of trees. I will discuss some recent work on algorithms and applications of BHV tree space, including using the geodesics to explore and visualize phylogenetic landscapes and to find relevant high likelihood trees.
Join the Webinar by logging in to ISCB Nucleus
Hosted by:
May 27, 2025 at 11:00 AM EDT
Understanding the language of proteins has been a major focus in computational biology, with recent advances in protein language models (pLMs) leading to increasingly powerful sequence representations. These developments hold great promise for critical tasks such as protein-protein interaction (PPI) prediction, which plays a fundamental role in biological processes. However, despite progress in sequence representations, significant challenges remain.
In our previous work, we demonstrated that reported performances of sequence-based PPI prediction models were largely inflated due to data leakage. When evaluated on a strongly leakage-reduced dataset, models performed randomly, highlighting the field's open challenges. This motivated further method development, which leveraged the now widely used ESM-2 protein sequence embeddings. In our recent publication, we evaluated the contribution of ESM-2 embeddings compared to model architecture. While the embeddings led to a substantial performance boost, accuracy plateaued at 0.65—regardless of model architecture. This suggests that the improvements stem from better sequence representations rather than increased model complexity.
We argue that sequence-based embeddings alone are insufficient to drive PPI prediction forward. Since protein interactions occur in three-dimensional space, incorporating structural information is crucial for generalizing to unseen proteins.
In this talk, I will introduce the field, discuss the pitfalls we encountered, and present our findings on the current limitations of sequence-based approaches.
Join the Webinar by logging in to ISCB Nucleus
Hosted by:
