ASKCOS: an open source software suite for synthesis planning

TL;DR

ASKCOS presents an open-source, modular software suite for computer-aided synthesis planning that supports both interactive and automatic retrosynthesis through four one-step strategies. The framework integrates template-based and template-free approaches, condition and outcome predictions, and comprehensive pathway evaluation utilities, all backed by a microservice-oriented refactor for easy extension and deployment. The work demonstrates broad industrial and academic adoption, emphasizes transparency and traceability to literature precedents, and provides data, models, and code under permissive licenses to foster community-driven development. Its open access design and diverse predictive modules enable chemists to ideate, evaluate, and optimize synthetic routes with quantified metrics and configurable planning horizons. The combination of interactive exploration, automated search, and auxiliary predictive tools positions ASKCOS as a practical, scalable platform for modern CASP research and application.

Abstract

The advancement of machine learning and the availability of large-scale reaction datasets have accelerated the development of data-driven models for computer-aided synthesis planning (CASP) in the past decade. Here, we detail the newest version of ASKCOS, an open source software suite for synthesis planning that makes available several research advances in a freely available, practical tool. Four one-step retrosynthesis models form the basis of both interactive planning and automatic planning modes. Retrosynthetic planning is complemented by other modules for feasibility assessment and pathway evaluation, including reaction condition recommendation, reaction outcome prediction, and auxiliary capabilities such as solubility prediction and quantum mechanical descriptor prediction. ASKCOS has assisted hundreds of medicinal, synthetic, and process chemists in their day-to-day tasks, complementing expert decision making. It is our belief that CASP tools like ASKCOS are an important part of modern chemistry research, and that they offer ever-increasing utility and accessibility.

Figures (5)

• Figure 1: ASKCOS overview. a) The typical target-oriented synthesis workflow for which ASKCOS is designed. A target molecule (blue circle) is recursively expanded retrosynthetically until buyable starting materials (green circles) are reached. Agents/conditions can be predicted for each proposed reaction, which can then undergo further evaluation, e.g., to anticipate reaction products. b) High-level flowchart for ASKCOS usage. A user can either interact via the graphical user interface or programmatic endpoints, which call various prediction modules. c) Summary of modules available in ASKCOS, which fall under themes including one-step retrosynthesis (green), multi-step search (red), condition recommendation (purple), reaction outcome prediction (blue), as well as utilities and supplementary capabilities (black). The modularity of the software design enables the straightforward extension of functionality (e.g., to new data-driven models) as they are developed in a research setting and made production-ready.
• Figure 2: Annotated screenshot of the interactive planning canvas in ASKCOS. The input target molecule shown is cenobamate, a drug for partial-onset seizures, defined by entering the SMILES string NC(=O)OC(Cn1ncnn1)c1ccccc1Cl in the search bar or using an external name resolver if permitted by security policies. An interactive retrosynthetic search can be initiated with the ONE STEP button. An automated multi-step retrosynthetic search can be initiated with the BUILD TREE button. Search settings for both modes can be set with the STRATEGY SETTINGS button. Here, results are shown for an interactive search performed using the template-based single step model trained on the Reaxys dataset. The top 4 suggestions by the model are displayed on the canvas. A legend for the frame colors is displayed at the bottom left of the page. The displayed two-tier tree was generated by following the initial single step expansion of the target molecule with the expansion of one of the suggested precursors. Additional retrosynthetic expansions can be performed by double clicking on a chemical node. The chemical node info window provides users with additional information and actions (e.g., adding a comment for that chemical or banning it from future suggestions). If a one-step model has already been used to suggest precursors for that chemical, the node info window displays (right inset, red dashed frame) suggested precursors and allows users to sort or filter these suggestions using a variety of metrics (e.g., score, synthetic complexity, number of rings). A reaction node info window (left inset, red dashed frame) provides users with information about the suggested reaction, including predicted scores and database precedents if they exist. The window also allows users to perform actions on the node such as removing or hiding it. Results can be saved, exported, or reorganized into different graphical views using additional tools in the bottom-right of the canvas.
• Figure 3: Annotated screenshot of tree search results in ASKCOS. The Tree Builder job can be initiated with the BUILD TREE button on the interactive path planning page. In My Results page, the status for the Tree Builder job is shown as Started. Once the job is complete, the status changes to Completed, and the trees found can be visualized/analyzed in the Tree Explorer (VIEW TREES) or the IPP canvas (VIEW IN IPP). The settings used for the Tree Builder job can be viewed with the VIEW SETTINGS button, and the results can be shared with others with the SHARE button. Here, the Tree Builder job for the target molecule NC(=O)OC(Cn1ncnn1)c1ccccc1Cl is performed using the MCTS algorithm with the template-based single step model trained on the Reaxys dataset, using a maximum depth of 5, a maximum branching factor of 25, and an expansion time of 30s (right bottom inset, red dashed frame). The results can be visualized in the IPP canvas (left bottom inset, red dashed frame).
• Figure 4: Annotated screenshot of the condition recommender (top) and the forward predictor (bottom) in ASKCOS. In the condition recommender, the SMILES strings of reactants and products are input in the Reactants and Products panels. Condition recommendation can be initiated with the GET RESULTS button. Prediction models and their training set can be set with the MODEL and SETTINGS buttons, respectively. Each proposed condition has shortcuts to run product prediction. In the forward predictor, the SMILES strings of reactants and agents are pre-populated using the shortcut from the condition recommendation page. Product prediction can be initiated with the GET RESULTS button. Prediction models and their training sets are chosen via SETTINGS button. Here, the results are shown for three predicted products with their probabilities, which should be interpreted only qualitatively. Each prediction has shortcuts to run impurity and regioselectivity prediction as additional evaluations.
• Figure 5: Annotated screenshot of the Tree Explorer in ASKCOS. One of 200 routes returned for the target molecule cenobamate is displayed in the main panel (right) of the window. Pathway metrics for the route are shown in the top left corner of the main panel. The left panel provides users with additional actions to perform on the returned synthetic routes. Users can transfer information from the Tree Builder to the Interactive Path Planner to continue exploring retrosynthetic suggestions beyond what was returned (top); users can initiate longer-running pathway-level scoring calculations (middle); finally, users can sort or filter the discovered routes based on various calculated metrics or the presence or absence of specific starting materials and intermediates (bottom).