[Paper Review] Machine Learning-guided Lipid Nanoparticle Design for mRNA Delivery

Daisy Yi Ding, Yuhui Zhang+2 more|arXiv (Cornell University)|Aug 2, 2023

RNA Interference and Gene Delivery12 citations

TL;DR

The paper develops ML models to predict LNP transfection efficiency from chemical structures, demonstrating that expert-crafted fingerprints with ML classifiers can achieve high accuracy on unseen designs, enabling in silico prioritization of LNP candidates for mRNA delivery.

ABSTRACT

While RNA technologies hold immense therapeutic potential in a range of applications from vaccination to gene editing, the broad implementation of these technologies is hindered by the challenge of delivering these agents effectively. Lipid nanoparticles have emerged as one of the most widely used delivery agents, but their design optimization relies on laborious and costly experimental methods. We propose to in silico optimize LNP design with machine learning models. On a curated dataset of 622 LNPs from published studies, we demonstrate the effectiveness of our model in predicting the transfection efficiency of unseen LNPs, with the multilayer perceptron achieving a classification accuracy of 98% on the test set. Our work represents a pioneering effort in combining ML and LNP design, offering significant potential for improving screening efficiency by computationally prioritizing LNP candidates for experimental validation and accelerating the development of effective mRNA delivery systems.

Motivation & Objective

• Motivate the use of ML to optimize lipid nanoparticle (LNP) designs for mRNA delivery.
• Create and curate a dataset of LNPs with transfection efficiency labels for ML modeling.
• Evaluate molecular representations and ML classifiers to predict transfection efficiency.
• Demonstrate feasibility of in silico screening to prioritize LNP designs for experimental validation.

Proposed method

• Formulate LNP design as a binary classification task predicting transfection efficiency from four molecular components and their ratios.
• Develop two molecular representations: expert fingerprints based on chemical domain knowledge, and neural fingerprints from Grover (GNN-based).
• Encode component ratios with one-hot embeddings and concatenate component representations.
• Train and compare classifiers (SVM, Random Forest, XGBoost, MLP) using AUROC as the evaluation metric.
• Curate a dataset of 622 LNPs from published studies and split into train/validation/test (approx. 8/1/1).
• Assess performance across fingerprint types and classifiers, reporting AUROC on validation and test sets.

Experimental results

Research questions

• RQ1 Can ML models accurately predict LNP transfection efficiency from chemical structure and composition data?
• RQ2 Do domain-knowledge-based expert fingerprints outperform neural fingerprints derived from graph neural networks for this task?
• RQ3 What is the best combination of fingerprint type and classifier for predicting LNP effectiveness on unseen designs?
• RQ4 Is there evidence that ML-guided design can meaningfully prioritize LNP candidates for experimental validation?

Key findings

• ML models can predict LNP transfection efficiency with high accuracy on unseen designs.
• MLP with expert fingerprints achieves the highest test AUROC of 0.9815 on the validation/test split.
• Expert fingerprints (domain-based) are sufficiently predictive, while Grover neural fingerprints do not consistently outperform them.
• Across experiments, several classifiers using expert fingerprints reach test AUROCs above 0.95.
• Dataset of 622 LNPs with transfection labels was curated from published studies and is publicly available for reproducibility.