Flow-Guided Biochemical Structure Generation With Action Values

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional systems for generating complex structures, such as molecular compounds, suffer from inaccuracies, inefficiencies, and operational inflexibilities due to an imbalance between reward-seeking and space exploration, leading to issues like mode collapse and computational instability.

Innovation Solution

The QGFN generation system combines a generative stochastic model with an action-value function model to balance reward-seeking and space exploration by using flow measures and action-values, adjusting their combination based on hyperparameters to select constructive object options, thereby improving the accuracy and efficiency of generating biochemical structures.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If conventional systems use generative methods to explore complex feature spaces, then diversity in sampling structures is improved, but accuracy and efficiency deteriorate due to mode collapse and computational instability

Engineering Contradiction:

Improvediversity in sampling structuresVSAvoidaccuracy in generating structures

Core Design Contradiction:

Adaptability or versatilityVSManufacturing precision

Solution Approach 1:

The patent introduces flow measures as an intermediary mechanism between the generative model and the reward function. The flow measure acts as a mediator that guides the sampling process through complex feature spaces while maintaining both diversity and accuracy. It provides a stable computational pathway that prevents mode collapse by distributing probability flow across multiple modes rather than concentrating on单一 high-reward regions

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The system dynamically adjusts sampling parameters based on flow measures and action-value estimates. By changing the balance between exploration and exploitation parameters during generation, the system can adapt to different regions of the feature space, maintaining diversity when needed and accuracy when high-reward structures are identified

Inventive Principle:

Principle #35Parameter changes

2Manufacturing precision

If conventional systems prioritize reward-seeking in generation, then accuracy in achieving objectives is improved, but operational flexibility deteriorates due to mode collapse

Engineering Contradiction:

Improveaccuracy in achieving objectivesVSAvoidoperational flexibility in exploration

Core Design Contradiction:

Manufacturing precisionVSAdaptability or versatility

Solution Approach 1:

The system implements dynamic balancing between reward-seeking and exploration through the mixing hyperparameter. The flow measure and action-value model work together to dynamically adjust the generation process, allowing the system to shift between exploiting high-reward regions and exploring new areas based on the current state and learned patterns

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The action-value model provides feedback about the expected reward of different actions, which is combined with flow measures to guide the generation process. This feedback mechanism allows the system to learn from past experiences and adjust its exploration-exploitation balance, preventing mode collapse while maintaining accuracy in achieving objectives

Inventive Principle:

Principle #23Feedback

3Productivity

If conventional systems increase computational resources for training generative methods, then productivity in generating structures is improved, but efficiency deteriorates due to computational instability

Engineering Contradiction:

Improveproductivity in generating structuresVSAvoidcomputational efficiency

Core Design Contradiction:

ProductivityVSLoss of energy

Solution Approach 1:

The patent replaces traditional reinforcement learning mechanisms with a flow-based generative approach combined with action-value estimation. This substitution avoids the computational instability inherent in conventional RL methods while maintaining the ability to learn from rewards. The flow measure provides a more stable computational framework that converges more reliably

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The system performs preliminary computation of flow measures and action-values during training to establish a stable foundation for generation. By pre-computing these guiding signals, the system reduces the computational burden and instability during the actual generation process, improving both productivity and efficiency

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS20250322917A1Utilizing flow measures of a generative stochastic model and action values of an action-value model to generate structural representations

Publication Date: 2025.10.16 RECURSION PHARMACEUTICALS INC

AI summary

The present disclosure relates to systems, non-transitory computer-readable media, and methods that utilize a generative stochastic model and an action-value function model to build a biochemical structure. Indeed, in one or more implementations, the disclosed systems generate a flow measure for a constructive object option in building a biochemical structure and further generate an action-value for the constructive object option. For instance, the disclosed systems combine the flow measure and the action-value to select the constructive object option from a plurality of constructive object options. Moreover, in some instances, the disclosed systems generate the biochemical structure using the selected constructive object option.