Tango*: Constrained synthesis planning using chemically informed value functions (2412.03424v1)

Abstract: Computer-aided synthesis planning (CASP) has made significant strides in generating retrosynthetic pathways for simple molecules in a non-constrained fashion. Recent work introduces a specialised bidirectional search algorithm with forward and retro expansion to address the starting material-constrained synthesis problem, allowing CASP systems to provide synthesis pathways from specified starting materials, such as waste products or renewable feed-stocks. In this work, we introduce a simple guided search which allows solving the starting material-constrained synthesis planning problem using an existing, uni-directional search algorithm, Retro*. We show that by optimising a single hyperparameter, Tango* outperforms existing methods in terms of efficiency and solve rate. We find the Tango* cost function catalyses strong improvements for the bidirectional DESP methods. Our method also achieves lower wall clock times while proposing synthetic routes of similar length, a common metric for route quality. Finally, we highlight potential reasons for the strong performance of Tango over neural guided search methods

• The paper introduces Tango*, a novel method for constrained synthesis planning that generates pathways from specific starting materials using a chemically informed cost function.
• Tango* demonstrated superior solve rates and efficiency on various datasets, outperforming existing methods while requiring fewer computational resources like node expansions and wall clock time.
• The success of Tango* suggests computed, non-neural cost functions can reliably guide synthesis planning, offering potential applications in medicinal chemistry and waste valorization.

Understanding Tango*: A Novel Approach for Constrained Synthesis Planning

The paper presents a significant development in the field of Computer-Assisted Synthesis Planning (CASP) through the introduction of Tango*, a method aimed at solving the challenges associated with constrained synthesis planning. Unlike traditional CASP systems, which typically generate retrosynthetic pathways to any available precursors, Tango* innovatively focuses on generating pathways from specified starting materials. This feature is particularly useful for transforming specific waste products or utilizing renewable feedstocks.

Key Contributions and Methodology

1. Node Cost Function - TANGO: A core component of Tango* is the TANGO node cost function, which guides retrosynthetic searches by leveraging a computed molecular similarity metric, combining Tanimoto Similarity and Fuzzy Matching Substructure (FMS). This approach allows the synthesis planning process to prioritize pathways that utilize predefined starting materials effectively.
2. Integration with Retro: The Tango method is integrated with Retro*—an existing uni-directional search algorithm known for its efficiency in retrosynthesis. Tango* modifies this algorithm by incorporating the TANGO node cost function, thereby adapting it to handle starting material constraints with potentially better results than specialized methods.
3. Performance Evaluation: The authors conducted experiments using datasets such as USPTO-190, Pistachio Reachable, and Pistachio Hard, demonstrating Tango*’s superior performance in terms of solve rate and efficiency. Tango* consistently outperformed or matched existing approaches while showing reduced computational overhead, evidenced by lower average numbers of node expansions and wall clock times.

Experimentation and Results

Tango* was benchmarked against existing methods like DESP (Double-Ended Synthesis Planning) and Retro* enhanced with neural cost functions. Notably, Tango* achieved the highest solve rates across various constraints and datasets, proving especially efficient in pathway discovery on the challenging UPSO-190 dataset. The scalability and generalizability of the TANGO cost function were further supported by its effectiveness across both simple and complex synthetic routes.

Theoretical and Practical Implications

The introduction of a computed, non-neural node cost function like TANGO challenges the existing reliance on neural networks for guidance in synthesis planning. The consistent performance improvements noted in the paper suggest that such computed functions may offer more reliable search guidance due to their structural invariance and granularity. This work opens new research directions that can focus on the development of alternative, perhaps simpler, cheminformatics-based node cost functions that can drive substantial efficiency in synthesis planning.

Future Directions

Tango* sets a foundation for future advancements in CASP, particularly for applications demanding constrained and steerable synthesis paths. The implications are vast, offering potential utility in fields like medicinal chemistry, where semi-synthesis and waste valorization are vitally important. Moreover, integrating these approaches into more sophisticated multi-objective frameworks could optimize synthesis routes for additional criteria such as cost, yield, and environmental impact, further enhancing the practical utility of CASP systems.

In conclusion, Tango* marks a significant progression towards more efficient and flexible synthesis planning, leveraging the inherent strengths of cheminformatics in guiding complex retrosynthetic processes.