Berkeley AI Researchers Develop PLAID for Controlled Protein Generation
Researchers from the Berkeley Artificial Intelligence Research (BAIR) lab have introduced PLAID, a novel multimodal generative model designed to create new proteins by leveraging the latent space of existing protein folding models. Unlike previous methods that primarily predict structure, PLAID simultaneously generates both discrete 1D amino acid sequences and continuous 3D all-atom structural coordinates. This approach addresses critical limitations in real-world drug design, such as the need for organism-specific humanization and precise control over functional constraints like solubility. A key innovation is PLAID's ability to train exclusively on sequence databases, which are significantly larger and more accessible than experimental structure datasets. The model accepts compositional prompts regarding function and organism type, enabling controlled generation of useful proteins. By mirroring textual control interfaces used in image generation, PLAID aims to streamline the complex process of developing biological therapeutics, marking a significant step forward from structure prediction to practical, controllable protein engineering.
Wire timeline
Berkeley AI Researchers Develop PLAID for Controlled Protein Generation
Researchers from the Berkeley Artificial Intelligence Research (BAIR) lab have introduced PLAID, a novel multimodal generative model designed to create new proteins by leveraging the latent space of existing protein folding models. Unlike previous methods that primarily predict structure, PLAID simultaneously generates both discrete 1D amino acid sequences and continuous 3D all-atom structural coordinates. This approach addresses critical limitations in real-world drug design, such as the need for organism-specific humanization and precise control over functional constraints like solubility. A key innovation is PLAID's ability to train exclusively on sequence databases, which are significantly larger and more accessible than experimental structure datasets. The model accepts compositional prompts regarding function and organism type, enabling controlled generation of useful proteins. By mirroring textual control interfaces used in image generation, PLAID aims to streamline the complex process of developing biological therapeutics, marking a significant step forward from structure prediction to practical, controllable protein engineering.
The Berkeley Artificial Intelligence Research Blog