Discovering New Molecules Using Graph Neural Networks by Rocío Mercado

Summary

Rosio Mercado from AstraZeneca's Molecular AI group discusses her work on molecular design using deep learning, focusing on a recently developed platform called Graph Invent. The platform leverages graph-based neural networks for iterative molecular generation, aiming to accelerate the drug discovery process. Mercado explains how AI can optimize the design of drug-like molecules, describing the methodology and architecture of Graph Invent, including its use of graph neural networks. She also shares performance benchmarks, challenges, and future improvements in scalability and computational costs.

Takeaways

• 🤖 The speaker, Rosio Mercado, from AstraZeneca's Molecular AI group, focuses on deep learning and graph-based molecular design.
• 🔬 Her talk introduces 'Graph Invent,' a platform using graph neural networks (GNNs) for molecular design, aiding the drug discovery process.
• 💡 Graph Invent employs deep molecular generative models to explore promising areas of chemical space, optimizing drug-like molecules.
• 🧪 The model tackles challenges in drug discovery, such as optimizing molecules' binding properties and navigating vast molecular spaces (estimated between 10^20 to 10^60 possible molecules).
• 🧠 There are two main classes of deep molecular generative models: string-based and graph-based approaches, with Graph Invent using a graph-based, iterative approach.
• ⚛️ Graph Invent generates molecules one atom or bond at a time by building molecular graphs iteratively, ensuring high validity and diversity of results.
• 📊 The model operates with a Message Passing Neural Network (MPN) followed by a feed-forward network, predicting how molecules should grow based on input data.
• 📈 The platform performs on par with state-of-the-art models in terms of novelty, diversity, and learning from limited data, but aims to improve its computational efficiency.
• 🧬 Graph Invent’s action prediction is based on probabilities for adding nodes, connecting nodes, or terminating molecule generation, optimizing the molecular structure.
• 🛠️ Although Graph Invent shows high success, future work includes scaling up its capacity to handle larger molecules and exploring other graph architectures.

Q & A

What is the primary focus of Rosio Mercado's work at AstraZeneca?

Rosio Mercado's work focuses on using deep learning methods for graph-based molecular design, specifically utilizing graph neural networks (GNNs) for molecular generation.

How does AI accelerate the drug development process, according to Rosio Mercado?

AI can accelerate the drug development process by generating molecules that meet desired properties, reducing the time spent in the early discovery phase and increasing the likelihood of success in pre-clinical and clinical trials.

What is Graph Invent, and how does it contribute to molecular design?

Graph Invent is a deep molecular generative model platform developed by Rosio Mercado's team. It uses graph-based, iterative approaches to generate new molecules one atom or bond at a time, offering high validity and diversity in generated structures.

What challenges exist in optimizing molecules during drug discovery?

The challenges in optimizing molecules include managing a large number of interdependent properties like target selectivity, novelty, and physical-chemical properties, while navigating an enormous space of possible drug-like molecules.

What are the two main classes of molecular generative models mentioned in the talk?

The two main classes of molecular generative models are string-based approaches and graph-based approaches. Each has its own advantages, with string-based models drawing from natural language processing methods and graph-based models using molecular graphs.

How does the iterative generation process work in Graph Invent?

In Graph Invent, molecules are generated iteratively by sampling actions to grow an input graph node by node. The model outputs a probability distribution of possible actions, and the process continues until the model decides to terminate the generation.

What advantages does Graph Invent offer in terms of molecular generation?

Graph Invent generates structures with high validity and diversity, can start with empty or non-empty graphs, and doesn't rely on encoded chemical rules, allowing it to learn molecular properties directly from the training data.

What are some areas where the Graph Invent model can be improved?

Graph Invent could be improved in terms of increasing the percentage of valid molecules generated (currently 95%) and reducing the computational cost, which is generally higher than string-based methods.

How does the Graph Neural Network (GNN) architecture work in Graph Invent?

The GNN architecture in Graph Invent has two main components: a message-passing neural network (MPN) that generates node and graph embeddings, and a global readout phase that predicts the next action (adding a node, connecting nodes, or terminating).

Why is the KL divergence used as the loss function in Graph Invent?

KL divergence is used in Graph Invent because it worked best in minimizing the difference between the predicted action probability distribution and the target action probability distribution, leading to better molecular generation performance.