We are developing an AI robot that, after conceiving experiments based on research hypotheses, estimates specific procedures in cyberspace and executes them in physical space.
Specifically, we will develop (1) “AI robots for organic synthesis” that conduct automatic experiments based on generated hypotheses, (2) “AI for predicting and extending organic synthesis reactions” that plans the entire experiment while predicting synthesis reactions,(3) “Reaction condition prediction AI using a general-purpose organic synthesis robot,” which autonomously estimates the conditions for each synthesis experiment, (4) “Flow synthesis control AI,” which provides a wide range of synthesis experiment options, and (5) “Mechano-chemical synthesis control AI.”

As an AI for automated synthesis experiments, our fiscal 2023 feasibility study demonstrated an AI that can plan experiments from papers and estimate specific experimental parameters as researchers do. Since literature often describes only general experimental settings, we are implementing extensions for hypothesis generation in fiscal 2024 and beyond. We aim to develop an AI robot that can plan and execute experiments by having literature-trained AI update synthesis routes and conditions from another AI's experimental plans, autonomously performing syntheses including flask, flow, and mechanochemical methods.

(1) Exploration of AI robots for organic synthesis

We constructed a network-type database (Molecular Reaction Graph) with synthesis pathways as edges and molecules as nodes and developed a simplified automatic synthesis device called "Chemputor" (Figure 1). In fiscal 2023, we automated experimental procedure input using ChatGPT to convert texts into Mermaid notation, approaching automatic generation program realization. We successfully performed esterification, acetylation, and amidation on 0.3 mol compounds. In fiscal 2024, we are connecting hypothesis generation to automated experiments by interacting with hypothesis-generating AI to create non-opioid analgesic structures and conduct bioassays.

(2) AI for predicting and expanding organic synthesis reactions

We constructed an AI model for considering the composition of organic materials synthesized based on generated hypotheses and studied methods for expanding organic materials to discover new hypotheses. We curated the large-scale chemical reaction database Pistachio to construct a dataset and developed a reverse synthesis pathway prediction model using a language model. We confirmed its practicality in a case study targeting the creation of analgesics.

(3) Reaction condition prediction AI using a general-purpose organic synthesis robot

We developed reaction condition prediction AI while constructing and utilizing a general-purpose organic synthesis robot. We quantified the reactivity of organic molecules using a reactivity prediction AI and discovered a new reagent, “Antipyrine.” Additionally, we constructed a high-precision reaction condition prediction model and successfully achieved the automatic synthesis of molecules promising as analgesics.

(4) Flow synthesis control AI

We aimed to build an integrated system from hypothesis generation to automatic flow synthesis, evaluation, and feedback using AI. We achieved a yield error of less than 5% in the flow synthesis and analysis system and demonstrated gram-scale synthesis using a plunger pump. We were responsible for the synthesis and activity evaluation of ADRA2B inhibitors as analgesic candidates and established a more efficient synthesis method than the batch method using flow synthesis of asymmetric urea.

(5) AI for predicting and extending organic synthesis reactions

Aiming to demonstrate organic synthesis AI using mechanochemistry, we constructed an autonomous experiment platform that combines a precise mechanical control system with reaction mechanism elucidation. We achieved high reproducibility in standard mechanochemical reactions such as the Knaefner-Nager condensation and demonstrated an AI prototype that elucidates reaction mechanisms without human intervention by fully automating powder X-ray diffraction and spectroscopic measurements.

Future developments focus on improving throughput and exploring broader chemical spaces. For throughput enhancement, we must evolve our automated synthesis AI to provide autonomous feedback without human intervention. To explore wider chemical spaces, we need to advance synthesis devices that offer diverse synthetic methods in both cyber and physical spaces.