Multi-task Reinforcement Learning for Molecular Design

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current reinforcement learning methods for molecular design, such as MolDQN, are computationally expensive and require extensive training time, making it inefficient to identify molecules with desired physico-chemical properties, and are highly dependent on initial molecules, necessitating re-training for each new initial molecule.

Innovation Solution

A method utilizing multi-task reinforcement learning for pre-training a neural network on multiple training molecules, allowing for a pre-trained network to be fine-tuned for specific initial molecules, reducing the number of iterations required for optimization and enabling the network to generalize better across different initial molecules.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If reinforcement learning is used for molecular design optimization, then molecules with desired physico-chemical properties can be identified, but training time is excessively long (thousands of iterations required)

Engineering Contradiction:
Improveoptimization qualityVSAvoidtraining time
Core Design Contradiction:
Manufacturing precisionVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-training the neural network on a large dataset of molecules before fine-tuning it for specific optimization tasks. This pre-training phase establishes a robust foundation that accelerates subsequent training convergence, reducing the thousands of iterations traditionally required to merely hundreds or tens of iterations while maintaining optimization quality.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The training process is segmented into distinct phases: pre-training on diverse molecular data, then fine-tuning on task-specific objectives. This segmentation allows the model to learn general molecular patterns first, then specialize for specific optimization goals, significantly reducing the time required for each phase compared to traditional end-to-end training.

Inventive Principle:
Principle #1Segmentation

2Manufacturing precision

If a deep Q-network is trained for a specific initial molecule, then optimization performance is achieved, but the model cannot be applied to other initial molecules without re-training from scratch

Engineering Contradiction:
Improveoptimization performanceVSAvoidapplicability to different initial molecules
Core Design Contradiction:
Manufacturing precisionVSAdaptability or versatility

Solution Approach 1:

The patent implements universality by designing a neural network architecture and pre-training procedure that creates a model capable of adapting to multiple different initial molecules and optimization tasks. The pre-trained network serves as a universal foundation that can be fine-tuned for various molecules without requiring complete re-training, making the system multi-functional and broadly applicable across different molecular design scenarios.

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

Solution Approach 2:

By performing preliminary pre-training on diverse molecular structures and properties, the network acquires generalizable knowledge that enables it to adapt to new initial molecules. This preliminary action creates a versatile model that can be quickly fine-tuned for specific molecules, eliminating the need to start from scratch for each new optimization task.

Inventive Principle:
Principle #10Preliminary action

3Manufacturing precision

If thousands of candidate initial molecules are considered for molecular design, then better molecules can be identified, but training a deep Q-network for each candidate would take weeks or months

Engineering Contradiction:
Improvemolecule qualityVSAvoidtotal training time
Core Design Contradiction:
Manufacturing precisionVSLoss of time

Solution Approach 1:

The patent performs preliminary pre-training on a comprehensive dataset encompassing diverse molecular structures and properties. This pre-trained model can then be rapidly fine-tuned for each candidate initial molecule, reducing the training time from weeks or months per molecule to a fraction of that time. This approach enables efficient evaluation of thousands of candidates while maintaining the ability to identify high-quality molecules.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The pre-trained neural network serves as a reusable template or copy that can be instantiated and fine-tuned for each candidate molecule. Instead of training entirely new models for each candidate, the system copies the pre-trained architecture and adapts it to specific candidates through efficient fine-tuning, dramatically reducing computational time while maintaining optimization quality.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4152336A1Method and computing system for molecular design via multi-task reinforcement learning
Publication Date: 2023.03.22 TOTALENERGIES ONETECH
  • EP4152336A1 patent drawingFigure 1a~2
  • EP4152336A1 patent drawingFigure 3~4
  • EP4152336A1 patent drawingFigure 5a~5c

AI summary

The present disclosure relates to a computer implemented method (10) for molecular design via reinforcement learning, for designing molecules optimizing a predetermined criterion representative of at least one physicochemical property, wherein molecular design is performed by selecting a sequence of modifications to an initial molecule according to a selection policy, wherein the selection policy uses a Q-function which predicts a cumulative criterion score associated to subsequent molecules obtained by subsequent modifications, said Q-function being modeled by a trained neural network, said method (10) comprising: - a pre-training phase (11) of a neural network modeling the Q-function, yielding a pre-trained neural network, wherein the pre-training phase comprises N > 1 reinforcement learning pre-training tasks applied to N respective different training molecules, - a training phase (12) of the pre-trained neural network using reinforcement learning, yielding the trained neural network, by applying different sequences of modifications to the initial molecule.