Discrete Diffusion Protein Models for Generation and Representation Learning

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current language models struggle with protein sequence generation and prediction due to limitations in capturing complex global interactions and bidirectional context, leading to inadequate performance in both generation and prediction tasks.

Innovation Solution

A discrete diffusion probability model-based language model (DPLM) is proposed, combining discrete diffusion probability models with language models, pre-trained on real protein sequences to capture global interactions and enable both generation and prediction capabilities.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional language models are used for protein sequence generation, then the model structure is simple, but the performance in capturing global interactions and bidirectional context is insufficient

Engineering Contradiction:
Improveperformance in capturing global interactionsVSAvoidmodel structure complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent merges the discrete diffusion probability model with the language model structure to create a unified model that simultaneously captures global interactions through diffusion mechanisms and bidirectional context through transformer architecture. This combination resolves the contradiction by integrating multiple functional components into a single cohesive system that achieves both improved reliability and controlled complexity.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The proposed model serves multiple functions: it performs both sequence generation and sequence understanding tasks, captures global interactions through diffusion mechanisms while maintaining bidirectional context through self-attention. This multi-functionality allows the model to address various protein analysis requirements with a single unified architecture, improving reliability without proportionally increasing complexity.

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Reliability

If traditional language models are used for protein sequence generation, then the model is easy to train, but the performance in both generation and prediction tasks is inadequate

Engineering Contradiction:
Improveperformance in generation and prediction tasksVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The model employs pre-training on large-scale protein sequence data followed by fine-tuning on specific tasks. This preliminary action allows the model to learn general protein sequence patterns and representations beforehand, which then transfer to improve performance on specific generation and prediction tasks, reducing the need for extensive task-specific training time.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The discrete diffusion process enables continuous refinement of sequence generation through iterative denoising steps. This continuity allows the model to progressively improve generation quality over multiple steps, achieving high performance in generation tasks while the pre-trained representations provide a strong foundation that reduces overall training time requirements.

Inventive Principle:
Principle #20Continuity of useful action

Data Source

PatentEP4629248A1Protein sequence generation and representation learning
Publication Date: 2025.10.08 BEIJING YOUZHUJU NETWORK TECH CO LTD
  • EP4629248A1 patent drawingFigure 1
  • EP4629248A1 patent drawingFigure 2
  • EP4629248A1 patent drawingFigure 3

AI summary

Embodiments of the present disclosure relate to protein sequence generation and representation learning. The method includes: obtaining a target model which is based on a discrete diffusion probability model and a language model structure and is pre-trained using a real protein sequence set; and performing, using the target model, at least one of the following: generating a first target protein sequence using the target model based at least on an input sequence in a form of protein sequence, or extracting a sequence feature representation corresponding to a second target protein sequence using the target model based on the second target protein sequence.