Mesenchymal stem cell (MSC) therapies hold the promise of not just addressing patients’ symptoms, but also healing and regenerating damaged tissues. The challenge for clinicians, however, is that it’s nearly impossible to predict whether or not MSC therapies will benefit an individual patient.
“Every MSC therapy is unique and highly dependent on the specific condition of the cell, treatment, and the recipient thus make it difficult for reproducibility,” explained Fredrik Liu, a Postdoctoral Research Fellow at the Massachusetts Institute of Technology and A*STAR scholar.
Cartilage acts as a shock absorber for joints and once damaged by injury or degenerative conditions, cannot repair itself, causing pain and restricting movement. While preclinical data using MSCs to repair cartilage have been promising, clinical trials have yielded mixed results. Because of this inconsistency, there are no clear clinical guidelines for how best to deploy MSC therapies.
“We wanted to analyze information from previous studies and suggest a guideline for MSC therapy,” said Liu. “However, to do so, a quantitative assessment of each therapy property was needed.”
To better understand this complicated data, Liu, together with a team led by Steve Oh, Director of the Stem Cell Bioprocessing Group at A*STAR’s Bioprocessing Technology Institute (BTI), turned to machine learning.
They first trained machine learning models using published data from 36 diverse animal and clinical studies on MSC therapy for cartilage repair. Next, they established neural networks that could draw connections between factors such as the number of stem cells administered, the defect area and patients’ body weight, measuring how these variables influenced therapeutic outcomes.
The model, the first of its kind, calculates the uncertainty of these properties simultaneously, making sense of these complex relationships to predict patient outcomes. “Our machine learning method works across all available information to extract useful knowledge,” Liu said. “It can capture multi-property correlations and leverage all of this historical experimental data to find relationships that humans may not see.”
Additionally, while standard machine learning methods can’t perform analyzes with incomplete datasets, the new model has the unique capability of filling in the blanks. “Our model identifies the link between existing properties and uses information from other completed entries to guide the extrapolation of the model,” explained Liu.
According to the authors, this machine learning tool can empower clinicians to personalize treatments for patients with cartilage damage, allowing them to predict whether MSCs will be the best option given an individual patient’s condition. The method is generic and can be adapted to address other biomedical questions, Liu added. To further improve this methodology, Liu has created a decentralized learning consortium to facilitate collaboration with international partners without the need to share any private data.
The A*STAR-affiliated researchers contributing to this research are from the Bioprocessing Technology Institute (BTI).
