Turning complex multi-omic trial data into evidence Candel could act on

Traditional clinical trials often focus on single "omics" layers, such as genomics, to identify disease-related genes or predict drug efficacy. However, this approach often fails to capture the complexity of biological systems, where multiple molecular processes interact to influence disease progression and treatment response. Multi-omics addresses this limitation by integrating data from different "omics" levels, providing a more holistic view of the underlying biology.

What is Multi-Omics Factor Analysis (MOFA)?

Multi-Omics Factor Analysis (MOFA) [1–3] is an unsupervised machine learning method designed to integrate and analyze diverse biological data types (e.g. genomics, proteomics, and transcriptomics) collected from the same set of samples. By identifying shared patterns across these datasets, MOFA reduces complex data into a few key factors that represent the main sources of variation.

Step 1: Data preprocessing

Ensure that the multi-omics datasets, including flow cytometry and proteomics data, are structured appropriately for MOFA. This typically involves organizing the data into matrices where rows represent features (e.g., cell types, proteins) and columns represent samples (patients).

To assess dynamic changes over time, compute the fold change for each feature by dividing the follow-up measurement by its corresponding baseline value. This fold change metric offers a standardized view of expression shifts, which enables meaningful cross-patient comparisons.

Apply appropriate normalization techniques to correct for technical noise and disparities inherent to each dataset type. This ensures consistent input for MOFA's integration process.

Step 2: Model training

Once the datasets are ready, we can input the preprocessed data into the MOFA framework. Specify the number of latent factors to be inferred, which represent the underlying sources of variation across the datasets.

Next, execute the MOFA algorithm to learn the latent factors. This process runs in an Amazon SageMaker notebook instance on our Sonrai Discovery platform and takes less than a minute to complete for a typical clinical trial biomarker dataset (hundreds of patients, thousands of features).

Step 3: Model interpretation

Once MOFA learns the factors, we can perform a series of exploratory analyses to understand their biological significance, and examine whether any of them are associated with clinical outcomes, such as overall survival.

Step 4: Machine learning to predict long survivors

MOFA simplifies complex multi-omics data by reducing hundreds or thousands of features into a handful of factors that capture the most meaningful information. These factors can then be used as inputs for machine learning algorithms to address specific questions.

For example, let's predict which patients are likely to be "long survivors" (patients surviving more than 30 months after treatment) — a demonstrative question used purely to illustrate the method, not Candel's own trial question. Two approaches are compared: one using the original concatenated multi-omics dataset, and the other using the 15 MOFA-derived factors.

Using full omics dataset: With all raw features concatenated, we trained a classifier using the no-code XGBoost app on the Sonrai Discovery platform. After testing several hyperparameter combinations, the model achieved a test F1-score of 0.67.

Using MOFA factors: By leveraging the simplified MOFA factors, the same XGBoost workflow saw immediate improvements, achieving a test F1-score of 0.74 — a substantial gain over the omics dataset. This demonstrates the utility of MOFA in enhancing downstream machine learning workflows.