Inferring experimental procedures from text-based representations of chemical reactions

The experimental execution of chemical reactions is a complex and time-consuming process, heavily reliant on the experience of chemists. While AI-driven retrosynthetic models help design synthetic routes, converting these routes into detailed experimental procedures remains a significant challenge. This bottleneck hinders the widespread adoption of automated synthesis platforms in chemistry, which require virtual assistants capable of generating precise execution programs for individual reactions. Existing AI approaches primarily focus on predicting specific reaction conditions (solvents, temperature) rather than the entire procedural sequence. This study introduces Smiles2Actions, an AI model aiming to bridge this gap by directly generating detailed experimental procedures from text-based chemical equations (SMILES notation) for batch organic synthesis.

Recent years have seen successful applications of AI in chemistry, including generative models for molecule design and retrosynthetic models for suggesting synthetic routes. However, translating these routes into executable experimental procedures remains a challenge. Previous works have explored AI models for predicting specific reaction conditions like solvents or temperatures for limited reaction classes. Some efforts have coupled retrosynthetic tools with nearest-neighbor searches in databases of existing procedures, but these often require manual revision. The lack of sufficiently curated data and the complexity of the domain have limited progress in creating AI models capable of predicting complete experimental procedures with minimal human intervention.

The Smiles2Actions model was trained on a dataset of 693,517 chemical equations and associated action sequences. This dataset was generated from the Pistachio database, which contains reaction records from patents. The experimental procedure text from these records was processed using a state-of-the-art natural language processing model (Paragraph2Actions) to extract action sequences. Subsequent steps involved data cleaning, standardization, and tokenization of key parameters (temperature, duration). Compound names were replaced with tokens indicating their position in the reaction input, and numerical values for temperature and duration were replaced with tokens representing predefined ranges. Three models were trained: a nearest-neighbor model based on reaction fingerprints, and two deep-learning sequence-to-sequence models based on the Transformer and BART architectures. Model performance was evaluated using metrics such as BLEU score, Levenshtein similarity, and accuracy at different similarity thresholds. Finally, a blind assessment of 500 predicted action sequences was conducted by an expert chemist to evaluate the practical usability of the model's predictions.

The three models outperformed random baselines, demonstrating their ability to learn patterns in chemical reaction procedures. The Transformer model achieved the best performance, with a 50% Levenshtein similarity for 68.7% of reactions, a 75% match for 24.7%, and a 100% match for 3.6%. The expert chemist assessment revealed that the predicted action sequences were deemed adequate for execution without human intervention in over 50% of the cases. Analysis of discrepancies between predicted and ground truth sequences showed that many differences involved variations in action ordering or minor property adjustments, rather than fundamental procedural errors. The model successfully predicted key aspects like solubility, precipitate formation, and heating/cooling requirements without explicit training on these concepts. The length of the predicted action sequences had a similar distribution to ground truth.

The results demonstrate the feasibility of using deep-learning models to predict complete experimental procedures for chemical reactions from SMILES notation. The high success rate of the Transformer model in the expert chemist assessment, surpassing the ground truth in adequacy in some aspects, shows its potential for real-world application. The relatively low percentage of exact matches (100% accuracy) is likely attributable to inherent noise in the dataset (multiple valid ways to perform a reaction), variations in linguistic description, and potential errors in the original patent data. The analysis suggests that further improvements in data quality would significantly enhance model performance. The study highlights the potential of AI to automate parts of the chemical synthesis process, although human verification remains crucial for safety reasons.

This work presents Smiles2Actions, a successful deep-learning model for predicting comprehensive experimental procedures for organic synthesis from chemical equations. The model's performance, validated by expert assessment, demonstrates its potential for accelerating chemical synthesis and facilitating the wider adoption of automated synthesis platforms. Future work could focus on improving data quality, incorporating additional procedural details (e.g., compound concentration, atmosphere), and integrating the model with automated synthesis systems.

The accuracy of the model is limited by the quality of the training data extracted from patent literature. Inaccuracies or inconsistencies in the original experimental procedures can propagate to the model's predictions. The model currently does not explicitly handle all aspects of experimental procedures (e.g., specific atmospheric conditions) due to variations in reporting practices within the patent data. Furthermore, the expert assessment, while valuable, was conducted on a limited subset of reactions and might not generalize perfectly across all chemical reaction classes.