Speaker: Pro. Zhang Haicang (Institute of computing technology, Chinese Academy Sciences)

Time: 13:30-14:30 p.m., April 18, 2024, GMT+8

Venue: Room 411, Jinguang Life Science Building, PKU

Abstract: 

Accurate prediction of damaging missense variants is crucial in genetic studies and clinical diagnosis. While many methods have been developed, their performance has been limited. We propose two deep learning-based methods, MVP and gMVP, for predicting the functional effects of missense variants. MVP utilizes a deep residual network to leverage large training datasets and numerous correlated predictors, while gMVP incorporates graph attention networks to incorporate co-evolutionary and structural information. Both MVP and gMVP significantly improve performance in deep mutational scanning data and cross-gene de novo variants enrichment. Additionally, we introduce DiffAffinity for predicting mutational effects on protein-protein binding. DiffAffinity employs a diffusion-based model to learn the generative process of side-chain conformations at protein-protein interfaces, achieving state-of-the-art performance.
De novo protein structure and sequence design play a critical role in protein engineering and drug discovery. We present CarbonDesign and CarbonNovo for protein sequence design and structure and sequence Co-design, respectively. CarbonDesign utilizes a novel architecture, Inverseformer, to learn representations from backbone structures and an amortized Markov Random Fields model for sequence decoding. CarbonNovo is the first end-to-end generative method for structure and sequence co-design, based on a unified energy-based model. Specifically tailored for antibody design, AbX integrates evolutionary, physical, and geometric constraints into a diffusion-based generative model, addressing the scarcity of labeled antibody-antigen complex data. All these methods achieve state-of-the-art performance in protein sequence and structure design.

Source: School of Life Sciences, PKU