Summary of Seminar: Multimodal deep learning for protein engineering (Kevin Yang)

This is an AI generated summary. There may be inaccuracies.
Summarize another video · Purchase summarize.tech Premium

00:00:00 - 01:00:00

In this seminar, Kevin Yang discusses multimodal deep learning for protein engineering. He explains how features computed by Rosetta can be used to predict protein structures more accurately. He also discusses the benefits of pre-training and the importance of domain generalization.

00:00:00 Dr. Kevin Yang is a senior researcher at Microsoft Research in Cambridge, Massachusetts, who works on problems at the intersection of machine learning and biology. He did his PhD at Caltech with Francis Arnold on applying machine learning to protein engineering. He taught math and physics for three years at a high school in Inglewood, California, through Teach for America. With that, we'd like to extend a warm welcome to Dr. Yang. In this seminar, he will be discussing multimodal deep learning for protein engineering. Dr. Yang will discuss the history and application of directed evolution, and how it can be used to design proteins with specific functions.
00:05:00 In this seminar, Kevin Yang explains how multimodal deep learning can be used to improve the efficiency of protein engineering. First, the different types of data that can be used for deep learning are discussed. Next, a machine learning-guided deep learning approach for sequence optimization is described. Finally, the benefits and limitations of this approach are highlighted.
00:10:00 This seminar discusses how deep learning can be used to improve protein engineering. The methods discussed include a deep learning model for mass language modeling, Transformer architectures for representational learning, and zero shot prediction.
00:15:00 The author discusses multimodal deep learning for protein engineering and its applicability to influenza. They discuss the Transformers deep learning architecture and how it scales linearly with the length of the input sequence, while the convolutional neural network (CNN) runs quadratically with sequence length.
00:20:00 This 1-paragraph summary explains how multimodal deep learning can be used to improve the performance of protein engineering tasks. The video describes how a pre-training task can be deconvoluted to separate the effects of the task and the architecture being trained on, and how cnns are more efficient for long sequences. Finally, the video shows how cnns can be used to predict the structure of proteins.
00:25:00 This seminar discusses the benefits of multimodal deep learning for protein engineering. In particular, it highlights how pre-training can help improve performance. It also discusses the importance of domain generalization and the need for a large training set.
00:30:00 The presenter discusses multimodal deep learning for protein engineering, highlighting the importance of pre-training and the benefits of using a CNN for corruption instead of a simple sequence-to-sequence model. They also mention that large models can sometimes outperform pre-trained weights, but that's not always the case.
00:35:00 The video discusses the benefits of multimodal deep learning for protein engineering, with particular emphasis on the benefits of using sequence transfer to improve performance. The video also provides code for a model that uses sequence transfer to improve performance on human data.
00:40:00 In this seminar, Dr. Kevin Yang discusses how multimodal deep learning can be used to improve protein engineering. Dr. Yang explains how features computed by Rosetta can be used to predict protein structures more accurately.
00:45:00 This seminar discusses the benefits of multimodal deep learning for protein engineering, including how to reduce pre-trained downstream task mix match. Pre-training is not always necessary, and using energetics features can give a big boost on out-of-the-main-generalization performance.
00:50:00 In this seminar, Kevin Yang discusses how multimodal deep learning can be used to pre-train protein fitness predictions. He also mentions that there is a trade-off between accuracy and the cost of energetics to compute.
00:55:00 In this seminar, Kevin Yang discusses multimodal deep learning for protein engineering. He notes that high quality MD simulations are necessary to begin, but that pairwise physical chemical properties tables are not very useful for guidance in building machine learning models. He discusses a project in which he used contrastive learning objectives to try to improve the accuracy of predictions made by a machine learning model. He notes that while this approach has been successful in some cases, it has not been effective in others. He recommends that, if possible, multimodal deep learning should be framed in a way that focuses on positive pair relationships.

01:00:00 - 01:00:00

In the seminar, Kevin Yang discusses multimodal deep learning for protein engineering. He emphasizes that there are many features that are important for downstream fitness prediction, but the most important feature will depend on the Downstream task. He also shares that Rosetta has tested other hyperparameters, but has not yet tested numbers in between.