AI for Chemical Synthesis: Graph Learning and Practical CASP

1.1 Why Graph AI Matters for Industrial Chemistry

• Limitations of traditional QSAR and fingerprints
• Structural learning vs engineered descriptors
• Where graph models outperform classical ML

1.2 Representing Molecules for Machine Learning

• Molecular graphs as structured data
• Practical feature engineering for industrial datasets
• Handling salts, mixtures, and real formulation data

1.3 How Graph Neural Networks Work

• Intuitive explanation of message passing
• Molecular embedding generation
• Interpreting learned representations

1.4 Hands-On: Building a Molecular Property Predictor

• Implementing a graph model in PyTorch
• Training on a small property dataset
• Evaluating robustness and model confidence

1.5 Deployment Considerations

• Data requirements
• Computational cost
• When GNN is overkill

2.1 Reaction Modelling with Graph AI

• Forward reaction prediction
• Reaction yield prediction
• Reaction classification

2.2 Retrosynthesis: How Modern Systems Work

• Template-based vs template-free methods
• Role of graph models in route generation
• Scoring and ranking synthetic pathways

2.3 Hands-On: Reaction Feasibility Modelling

• Training a reaction prediction model
• Interpreting outputs
• Understanding failure cases

2.4 Data Quality and Industrial Challenges

• Incomplete reaction conditions
• Batch vs flow differences
• Bias in public datasets
• Handling proprietary data

2.5 Integrating AI into Existing R&D Pipelines

• Where graph models fit in medicinal chemistry
• Supporting process development
• Human-in-the-loop AI systems

3.1 Architecture of Industrial CASP Systems

• Reaction prediction engine
• Feasibility scoring model
• Search algorithms
• Route ranking

3.2 Practical Route Evaluation

• Step count vs operational complexity
• Protecting group management
• Functional group compatibility
• Scale-up risk

3.3 Hands-On: Mini CASP Workflow

• Generating candidate routes
• Scoring synthetic feasibility
• Comparing routes using industrial criteria

3.4 Limitations of Current CASP Platforms

• Ignoring solvent systems
• Isolation and purification challenges
• Temperature and safety constraints
• Process-scale realities

3.5 Best Practices for Industrial Adoption

• Validation strategies
• Cross-team integration
• Regulatory and traceability considerations

Option A – 3D Graph Models and Materials Applications

Ideal for materials, catalysis, and advanced R&D teams.

• Incorporating 3D structural information
• Conformer-aware modelling
• Predicting energy-related properties
• Quantum-informed machine learning

Option B – AI-Driven Reaction Optimisation and Flow Integration

Ideal for process and manufacturing teams.

• Yield prediction models
• Condition recommendation
• Bayesian optimisation overview
• Connecting graph models to automated flow systems
• Towards semi-autonomous synthesis platforms

This course is designed for industrial chemists, process scientists, medicinal chemists, and computational R&D professionals who already have a working knowledge of machine learning and neural networks, and who are ready to take a practical, hands-on step into graph-based AI and computer-aided synthesis planning.

Attendees should be comfortable with core ML concepts (such as training/validation, model evaluation, and feature engineering) and have prior experience using Python in a notebook environment (e.g. Deepnote, Jupyter, or Google Colab). This is typically gained through attendance at our Empowering Chemists with AI & ML course, or equivalent experience.

Those working in route design, synthesis optimisation, flow chemistry, or materials R&D will find the content directly applicable to their day-to-day workflows.

Participants will:

• Understand when and why to use graph neural networks in chemistry
• Build and evaluate molecular and reaction prediction models
• Understand how modern retrosynthesis systems operate
• Critically assess AI-generated synthetic routes
• Identify realistic integration pathways into industrial R&D workflows
• Recognise limitations and risk factors before deployment

All hands-on sessions will be run in Python using a cloud-based notebook environment (Deepnote or equivalent, no local installation required). The following libraries and frameworks will be used:

PyTorch and PyTorch Geometric – for building and training graph neural networks RDKit – for cheminformatics and molecular graph generation DeepChem (optional) – for benchmarking and molecular property datasets Pandas / NumPy / Matplotlib – for data handling and visualisation

Participants should be comfortable running Python code in a notebook environment prior to attending. Starter notebooks and datasets will be provided in advance.