MeMDLM: De Novo Membrane Protein Design with Property-Guided Discrete Diffusion

Shrey Goel, Vishrut Thoutam, Edgar Mariano Marroquin, Aaron Gokaslan, Arash Firouzbakht, Sophia Vincoff, Volodymyr Kuleshov, Huong T. Kratochvil, Pranam Chatterjee

Track: Full Paper Track

Keywords: Membrane protein design, discrete diffusion, property guidance sampling

TL;DR: MeMDLM is a property-guided masked diffusion language model for de novo membrane protein generation, enabling the design of soluble membrane proteins and transmembrane domains with high biological relevance.

Abstract: Masked Diffusion Language Models (MDLMs) have recently emerged as a strong class of generative models, paralleling state-of-the-art (SOTA) autoregressive (AR) performance across natural language modeling domains. While there have been advances in AR as well as both latent and discrete diffusion-based approaches for protein sequence design, masked diffusion language modeling with protein language models (pLMs) is unexplored. In this work, we introduce MeMDLM, an MDLM tailored for membrane protein design, harnessing the SOTA pLM ESM-2 to de novo generate realistic membrane proteins for downstream experimental applications. Our evaluations demonstrate that MeMDLM-generated proteins exceed AR-based methods by generating sequences with greater transmembrane (TM) character. We further apply our design framework to scaffold soluble and TM motifs in sequences, demonstrating that MeMDLM-reconstructed sequences achieve greater biological similarity to their original counterparts compared to SOTA inpainting methods. Finally, we apply a generalized Bayesian optimization procedure that uniquely uses saliency maps to facilitate the generation of soluble membrane proteins, paving the way for experimental applications. In total, our pipeline motivates future exploration of MDLM-based pLMs for protein design.

Attendance: Shrey Goel

Submission Number: 25