College
College of Engineering
Mentor Information
Rituparna Samanta
Description
An important group of proteins to design are membrane proteins, and a specific type of membrane protein is the transmembrane β-barrel (TMB), commonly used as nanopores. Nanopores are used for analysis and sequencing of DNA molecules. To analyze specific targets, the chemical properties of the TMB need to be compatible with the target, otherwise the nanopore will be unable to detect change in current. De novo design would allow for making these versatile and functional nanopores. With current methods, it is difficult to design these nanopores de novo due to the lack of PDB representations. Our goal is to computationally design sequences for membrane embedded nanopores by combining the structure-based protein design model ProteinMPNN with the membrane protein language model MemDLM. We expect that using a combination of embeddings from sequence-based language models and weights of structure-based design models will improve membrane protein design. Our hypothesis is that integrating explicit physical 3D-geometric principles with evolutionary language models produces proteins that are both highly stable and biologically functional, overcoming limitations of relying on just one approach. After training MemDLM, an adapter was added to it. This adapter is a layer that can accept structural information from ProteinMPNN and feed it to the pre-trained MemDLM with frozen parameters. The system was then tested on a blind dataset of both designed and naturally occurring beta barrels, and its ability to create effective TMBs will be further tested through sequence identity, hydropathy of the barrel core, and propensity for beta sheet formation.
Language Model Adapter for de novo Trans-Membrane Beta-Barrel Protein Design
An important group of proteins to design are membrane proteins, and a specific type of membrane protein is the transmembrane β-barrel (TMB), commonly used as nanopores. Nanopores are used for analysis and sequencing of DNA molecules. To analyze specific targets, the chemical properties of the TMB need to be compatible with the target, otherwise the nanopore will be unable to detect change in current. De novo design would allow for making these versatile and functional nanopores. With current methods, it is difficult to design these nanopores de novo due to the lack of PDB representations. Our goal is to computationally design sequences for membrane embedded nanopores by combining the structure-based protein design model ProteinMPNN with the membrane protein language model MemDLM. We expect that using a combination of embeddings from sequence-based language models and weights of structure-based design models will improve membrane protein design. Our hypothesis is that integrating explicit physical 3D-geometric principles with evolutionary language models produces proteins that are both highly stable and biologically functional, overcoming limitations of relying on just one approach. After training MemDLM, an adapter was added to it. This adapter is a layer that can accept structural information from ProteinMPNN and feed it to the pre-trained MemDLM with frozen parameters. The system was then tested on a blind dataset of both designed and naturally occurring beta barrels, and its ability to create effective TMBs will be further tested through sequence identity, hydropathy of the barrel core, and propensity for beta sheet formation.
