Poster Preview
College
College of Engineering
Mentor Information
Rituparna Samanta
Description
Membrane proteins, particularly β-barrel membrane proteins, play essential roles in cellular processes and have growing applications in biotechnology, biosensing, and sequencing technologies. However, de novo design of membrane proteins remains underexplored because current inverse folding models, are primarily trained on soluble proteins and have limited exposure to membrane protein structures. The scarcity of experimentally determined membrane protein structures further limits the development of deep learning models for membrane protein design. In this work, we investigate strategies to retrain ProteinMPNN for membrane protein sequence design by retraining the model on a number of membrane protein datasets consisting of α-helical and β-barrel membrane protein complexes. Model performance is evaluated using blind-test set composing of natural and de-novo backbones, on sequence recovery, hydropathy, and β-sheet propensity, with particular emphasis on β-barrel proteins. To complement structure-based sequence generation with membrane-specific sequence information, we integrate ProteinMPNN with MemDLM, a protein language model trained on large-scale membrane protein sequence data. We introduce a distributed adapter architecture that injects ProteinMPNN-derived structural embeddings into every layer of MemDLM, enabling backbone information to shape sequence representations throughout the language model. By leveraging structural features and membrane-specific sequence representations, we hypothesize that the model will improve sequence recovery and better reproduce membrane-associated sequences, particularly for β-barrel proteins. More broadly, this framework provides a general strategy for combining inverse-folding models with domain-specific protein language models for de novo membrane protein design. At the end, we will test these new models on our blind test set to determine its accuracy.
Designing β-barrel Membrane Protein using Distributed Structure-to-Sequence Adapters.
Membrane proteins, particularly β-barrel membrane proteins, play essential roles in cellular processes and have growing applications in biotechnology, biosensing, and sequencing technologies. However, de novo design of membrane proteins remains underexplored because current inverse folding models, are primarily trained on soluble proteins and have limited exposure to membrane protein structures. The scarcity of experimentally determined membrane protein structures further limits the development of deep learning models for membrane protein design. In this work, we investigate strategies to retrain ProteinMPNN for membrane protein sequence design by retraining the model on a number of membrane protein datasets consisting of α-helical and β-barrel membrane protein complexes. Model performance is evaluated using blind-test set composing of natural and de-novo backbones, on sequence recovery, hydropathy, and β-sheet propensity, with particular emphasis on β-barrel proteins. To complement structure-based sequence generation with membrane-specific sequence information, we integrate ProteinMPNN with MemDLM, a protein language model trained on large-scale membrane protein sequence data. We introduce a distributed adapter architecture that injects ProteinMPNN-derived structural embeddings into every layer of MemDLM, enabling backbone information to shape sequence representations throughout the language model. By leveraging structural features and membrane-specific sequence representations, we hypothesize that the model will improve sequence recovery and better reproduce membrane-associated sequences, particularly for β-barrel proteins. More broadly, this framework provides a general strategy for combining inverse-folding models with domain-specific protein language models for de novo membrane protein design. At the end, we will test these new models on our blind test set to determine its accuracy.
