A conditional protein diffusion model generates artificial programmable endonuclease sequences with enhanced activity

Abstract

Computation or deep learning-based functional protein generation methods address the urgent demand for novel biocatalysts, allowing for precise tailoring of functionalities to meet specific requirements. This emergence leads to the creation of highly efficient and specialized proteins with wide-ranging applications in scientific, technological, and biomedical domains. This study establishes a conditional protein diffusion model, namely CPDiffusion, to deliver diverse protein sequences with desired functions. While the model is free from extensive training data and the sampling process involves little guidance on the type of generated amino acids, CPDiffusion effectively secures essential highly conserved residues that are crucial for protein functionalities. We employed CPDiffusion and generated 27 artificially designed Argonaute proteins, programmable endonucleases applied for easy-to-implement and high-throughput screenings in gene editing and molecular diagnostics, that mutated approximately 200-400 amino acids with 40% sequence identities to those from nature. Experimental tests demonstrate the solubility of all 27 artificially-designed proteins (AP), with 24 of them displaying DNA cleavage activity. Remarkably, 74% of active APs exhibited superior activity compared to the template protein, and the most effective one showcased a remarkable nearly nine-fold enhancement of enzymatic activity. Moreover, 37% of APs exhibited enhanced thermostability. These findings emphasize CPDiffusion's remarkable capability to generate long-sequence proteins in a single step while retaining or enhancing intricate functionality. This approach facilitates the design of intricate enzymes featuring multi-domain molecular structures through in silico generation and throughput, all accomplished without the need for supervision from labeled data.

Protein Engineering with Lightweight Graph Denoising Neural Networks

“…As an alternative, researchers often opt to pretrain models to encode protein representations from their sequences and/or structures. The learned embeddings can be utilized for de novo protein design, − variant effects prediction, − higher-level structure prediction, , and so on. Analogous to coaching a novice in a new field, a model’s learning efficiency is maximized by initially exposing it to a substantial data set of WT, allowing it to grasp the broader context before delving into specific proteins and their functionalities.…”

Protein Engineering with Lightweight Graph Denoising Neural Networks

“…As an alternative, researchers often opt to pre-train models to encode protein representations from their sequences and/or structures. The learned embeddings can be utilized for de novo protein design [11][12][13][14][15], variant effects prediction [16][17][18][19], higher-level structure prediction [20,21], etc. Analogous to coaching a novice in a new field, a model's learning efficiency is maximized by initially exposing it to a substantial dataset of WT, allowing it to grasp the broader context before delving into specific proteins and their functionalities.…”

Semantical and Geometrical Protein Encoding Toward Enhanced Bioactivity and Thermostability

“…Deep learning approaches have been instrumental in advancing scientific insights into proteins, predominatly categorieed into two primary domains: sequence-based [30], [34], [60] and structure-based [33], [80], [81] methods. Analogous to the analysis of semantics and syntax in natural language, protein language models (PLM) interpret protein sequences as raw text and employ autoregressive inference techniques [34], [43] with self-attention mechanisms [3], [41], [51].…”