Automated Drug Discovery Advances with AI-Driven Synthesis of 905,990 Reactions

Automated planning of chemical synthesis remains a significant hurdle in accelerating drug discovery and enabling fully robotic laboratories, largely because translating computational routes into practical, executable procedures proves challenging. Guoqing Liu, Junren Li, and Zihan Zhao, working with colleagues at Microsoft Research and Shanghai Jiao Tong University, now present a new model, QFANG, which directly generates detailed, structured experimental procedures from simple reaction equations. This achievement stems from a novel Chemistry-Guided Reasoning framework and a large dataset of nearly one million chemical reactions, carefully extracted from patent literature, allowing the model to learn complex chemical reasoning.

The team further refined QFANG using a reinforcement learning approach, resulting in a system that surpasses existing methods in generating accurate and adaptable synthesis procedures, and represents a crucial advance towards fully automated chemical synthesis workflows. A key challenge, however, is bridging the gap between computational route design and practical laboratory execution, particularly the accurate prediction of viable experimental procedures for each synthesis step. In this work, the team presents QFANG, a scientific reasoning language model capable of generating precise, structured experimental procedures directly from reaction equations, with explicit chain-of-thought reasoning. To develop QFANG, the researchers curated a high-quality dataset comprising 905,990 chemical reactions paired with structured experimental protocols. This dataset allows the model to learn the complex relationships between molecular transformations and the corresponding laboratory actions, ultimately enabling more reliable and efficient synthesis planning. AI Reasoning Advances Chemical Procedure Generation QFANG represents a significant advancement in AI-driven chemical procedure generation, moving beyond simple template retrieval to exhibit genuine reasoning capabilities. The system adapts to constraints, such as green chemistry principles and scale-up requirements, and generates novel procedures, consistently outperforming existing models.

Results demonstrate that QFANG doesn’t simply memorize chemistry, but understands the underlying principles, allowing it to generalize effectively. Performance on test sets reveals that QFANG excels even when presented with reactions dissimilar to those in its training data, indicating a capacity for independent reasoning. The system successfully designed a 50kg scale Suzuki coupling procedure, explicitly avoiding column chromatography and incorporating a palladium scavenger for industrial compliance. This demonstrates QFANG’s ability to bridge the gap between lab-scale discovery and process chemistry, correctly identifying the need for different purification strategies at scale and justifying its choices based on operational considerations. Furthermore, QFANG successfully adapted a Wittig reaction procedure to adhere to green chemistry principles, avoiding hazardous solvents and prioritizing a bio-derived solvent. This showcases its ability to translate abstract directives into concrete experimental steps, demonstrating a sophisticated understanding of green chemistry principles. The report details optimization of training parameters to achieve high throughput and low latency, demonstrating a commitment to practical implementation and scalability. A strong correlation between throughput and reward suggests that optimizing these parameters can significantly improve performance. QFANG’s strengths include its reasoning ability, adaptability, generalization, operational awareness, and scalability. However, the evaluation relies heavily on automated metrics, and human evaluation is crucial for verifying chemical validity and practicality. Evaluating QFANG’s ability to handle a wider range of constraints and conducting detailed error analysis would further improve its performance. Despite these areas for improvement, QFANG represents a significant step towards fully automated laboratory synthesis and optimized chemical procedures. QFANG Automates Synthesis Procedure Generation From Equations The research team has developed QFANG, a novel system designed to generate precise experimental procedures directly from chemical reaction equations, representing a significant step towards fully automated synthesis workflows. This work centers on bridging the gap between computational reaction design and practical laboratory execution, a longstanding challenge in chemistry. Scientists curated a large dataset of chemical reactions, each paired with a structured sequence of actions, automatically extracted and processed from extensive patent literature using large language models. The core of QFANG lies in a Chemistry-Guided Reasoning framework, which produces detailed, step-by-step reasoning grounded in established chemical knowledge at scale. This framework undergoes supervised fine-tuning to enhance its ability to perform complex chemical reasoning, and is further refined through Reinforcement Learning from Verifiable Rewards, a technique that specifically improves the accuracy of the generated procedures. Experiments demonstrate that QFANG outperforms both advanced general-purpose reasoning models and nearest-neighbor retrieval baselines, as measured by automated similarity metrics and a chemically-aware evaluator utilizing a large language model.

The team demonstrated QFANG’s ability to adapt to variations in laboratory conditions and user-specific constraints, and to generalize to certain reaction classes outside of its initial training data. This breakthrough delivers a system capable of generating high-quality synthesis procedures, paving the way for fully automated laboratory synthesis and accelerating the pace of molecular innovation. QFANG Generates Chemical Procedures From Equations QFANG, a new scientific reasoning model, successfully generates high-fidelity chemical procedures directly from reaction equations, addressing a critical gap between computational synthesis planning and practical laboratory execution.

The team achieved this by constructing a large dataset of chemical reactions paired with structured action sequences, automatically annotated using advanced language models. This dataset then informed a chemistry-guided reasoning framework, enabling the model to elicit chemically-principled reasoning, further refined through reinforcement learning with verifiable rewards. Evaluations demonstrate that QFANG outperforms existing methods, including state-of-the-art language models, both in automated similarity metrics and in assessments of chemical validity. Importantly, the model exhibits an ability to generalize to reactions outside of its original training data, adapt to user-defined constraints, and even identify and correct flawed procedures present within its training corpus, indicating a capacity for critical assessment. These achievements represent a significant step towards fully automated laboratory synthesis and next-generation platforms for scientific discovery. Further research will likely focus on expanding the dataset to encompass a wider range of chemical reactions and experimental conditions, and on developing methods to improve the model’s robustness and reliability in complex scenarios. 👉 More information 🗞 A Scientific Reasoning Model for Organic Synthesis Procedure Generation 🧠 ArXiv: https://arxiv.org/abs/2512.13668 Tags: