RetroLens: A Human-AI Collaborative System for Multi-step Retrosynthetic Route Planning

Title of the Paper

RetroLens: A Human-AI Collaborative System for Multi-step Retrosynthetic Route Planning

Paper Information

• Research Domain: Human-AI Collaboration, Chemical Synthesis, Artificial Intelligence
• Keywords: Human-AI Collaboration, Multi-step Problem Solving, Multi-criteria Decision Making, Computational Chemistry, Chemical Synthesis Route Planning

Research Background and Problem Statement

• Identified Problems or Challenges:

  • Multi-step retrosynthetic route planning (MRRP) is a core task in chemical synthesis, but due to the vast search space and high planning complexity, chemists often spend significant amounts of time on this process.
  • Automated AI models, while fast, can only handle simple target molecules and struggle with complex ones.
  • For complex molecule synthesis, relying solely on manual planning by chemists is time-consuming and prone to errors.

• Importance and Research Motivation:

  • Chemical synthesis is widely applied in fields such as drug discovery, but existing challenges lead to long synthesis times and low efficiency.
  • AI's limited capability in handling complex molecules makes human-AI collaboration a potential solution to these issues.
  • Exploring effective human-AI collaboration methods can help improve chemists' planning efficiency and reduce cognitive load.

• Related Work:

  • Existing AI models use template-based or deep learning methods to address localized chemical retrosynthesis problems, but they face limitations in long-term route planning.
  • Human-AI collaboration has primarily focused on single-step problems, while dynamic multi-step problem-solving remains an exploratory area.

Solution

• Methods and Proposed Solution:

  • A human-AI collaborative system, RetroLens, is proposed to support chemical retrosynthetic route planning for complex molecules through two collaboration modes: joint action and AI-assisted decision-making.
  • By task allocation, combining chemists' expertise with AI capabilities, the system addresses the limitations of AI in handling complex molecules.

• Innovations:

  • Integration of two human-AI collaboration modes: joint action and AI-assisted decision-making, for dynamic multi-step problem-solving.
  • A candidate step recommendation and ranking mechanism based on user-specified multi-criteria, aiding chemists in quickly resolving failed routes.

• Implementation Steps and Key Techniques:

  1. Use heuristic rules to allocate retrosynthesis tasks: AI handles simpler molecules, while chemists focus on more complex ones.
  2. If route planning fails, the AI module analyzes and identifies the issue, recommending possible corrective steps.
  3. Employ multi-criteria decision-making algorithms (e.g., Simple Additive Weighting, SAW) to rank corrective steps and provide optimization suggestions based on multiple user-specified criteria.

Research Outcomes

• Specific Results:

  • RetroLens significantly improved chemists' planning speed for complex molecule synthesis.
  • The system expanded the design space exploration for chemists, enabling users to discover more synthesis routes.
  • Provided a multi-dimensional evaluation method for route modification, making corrections more efficient.

• Advantages:

  • Compared to manual planning and single AI-based route planning solutions, RetroLens reduced chemists' cognitive load while enhancing the comprehensiveness and practicality of route planning.
  • Participants reported higher confidence in planning results and reduced stress during the planning process after using the system.

• Experimental and Evaluation Results:

  • User studies showed that participants completed tasks significantly faster and explored a broader design space using RetroLens compared to manual conditions.
  • The system improved chemists' user experience in the MRRP process, including decision-making efficiency and satisfaction.

• Limitations and Future Directions:

  • RetroLens currently lacks the ability to edit AI-generated routes, limiting user flexibility in planning.
  • The candidate step ranking algorithm lacks certain refined criteria, necessitating the expansion of standards or the introduction of more sophisticated ranking methods in the future.
  • Future experiments need to further validate the practical experimental effectiveness of these synthesis routes and explore the impact of users' chemical expertise on collaboration modes.