Protein Language Model Supervised Precise and Efficient Protein Backbone Design Method

Abstract

Proteins are essential macromolecules that play crucial roles in nearly every type of biological function. Most of the protein functions are determined by their position topologies, indicating that new functional proteins might be generated by designing their tertiary structures initially. Over the past two decades, numerous energy-based and machine learning algorithms have been proposed forde novoprotein design. However, optimizing these models to achieve a balanced performance among three critical aspects, namely sample fidelity, diversity, and generation speed, remains a challenging task on account of the constrained residue connections and hydrophobic interactions. Here we used the protein language model (pLM) as supervision to develop two complementary deep learning algorithms, Generative Protein Design by Language Model-Inpainting (GPDL-Inpainting) and GPDL-Hallucination, that can sufficiently represent and generate the protein structures and sequences. We compared the models with the state-of-the-art protein design models (RFdiffusion, RF Inpainting, and RF Halluciantion) and demonstrated that our methods can generate more designable structures with more diversity and higher computation efficiency. We also explored the possibility of the combination of hallucination and inpainting which might further improve the model efficiency and diversity. The source code of GPDL is available athttps://github.com/sirius777coder/GPDL.

Scaffold-Lab: Critical Evaluation and Ranking of Protein Backbone Generation Methods in A Unified Framework

“…FrameFlow 51 was developed upon FrameDiff's architecture, with the former used a particle filtering strategy to raise conditional sampling applied to motif-scaffolding task, and the latter replaced diffusion models with a flow-matching process to significantly elevate the efficiency and designability compared with previous works. Akin to the strategy in Wang et al 45 , GPDL 48 leveraged the single structure prediction method ESMFold 16 combined with the constrained hallucination and inpainting strategies to generate highly designable and diverse scaffolds around given motifs. Given the distinct application contexts and abilities to generate long sequences among various methods, it is crucial to categorize them appropriately prior to evaluation.…”

“…We selected the most widely-used motif-scaffolding task as a representative under this category, whose goal is to generate novel scaffolds around one or multiple motifs extracted from native protein structures, which usually plays a key role in specific biological functions. This strategy is particularly useful in applications like designing novel enzymes 41 , binders 34,42,43 or vaccines 44 and has been widely explored and evaluated within the last few years 33,34,[45][46][47][48] . Within this problem set, several computational metrics designed to evaluate these methods were recast into this evaluation framework followed by a ranking stage performed to assess the capability of different methods.…”

A Unified Framework of Scaffold-Lab for Critical Assessment of Protein Backbone Generation Methods

“…Both TDS 46 and FrameFlow 51 was developed upon FrameDiff's architecture, with the former used a particle filtering strategy to raise conditional sampling applied to motif-scaffolding task, and the latter replaced diffusion models with a flow-matching process to significantly elevate the efficiency and designability compared with previous works. Akin to the strategy in Wang et al 45 , GPDL 48 leveraged the single structure prediction method ESMFold 16 combined with the constrained hallucination and inpainting strategies to generate highly designable and diverse scaffolds around given motifs.…”

“…We selected the most widely-used motif-scaffolding task as a representative under this category, whose goal is to generate novel scaffolds around one or multiple motifs extracted from native protein structures, which usually plays a key role in specific biological functions. This strategy is particularly useful in applications like designing novel enzymes 41 , binders 34,42,43 or vaccines 44 and has been widely explored and evaluated within the last few years 33,34,[45][46][47][48] . Within this problem set, several computational metrics designed to evaluate these methods were recast into this evaluation framework followed by a ranking stage performed to assess the capability of different methods.…”