Modeling the complex interplay between cellular networks and resulting physiological conditions requires integration of multi-modal data types. In Li et al npj Microgravity 2023, we leverage graph-based explainable AI through AbzuAI QLattice to provide transparent characterization of the relationship between biological features. We integrate multi-modal biological data (gene expression, protein expression, genomic methylation, and calcium uptake levels) from spaceflown mouse muscle to identify the relationships between molecular features to predict the changes in calcium uptake in muscle. In future work, we plan to develop advanced AI-based methods for integrating multi-modal complex biological datasets.

Understanding causal relationships is essential for biomedical applications. We leverage an ensemble of causal inference machine learning methods (CRISP) to identify genes whose expression is potentially causal of increased fat content in liver from spaceflown mice. Currently, we are working with a Crowdsourcing Contenders award to develop a version of CRISP that can identify causal features in both tabular and bio-imagery data.

We provide open-source, no-cost, zero-prerequisite, virtual and asynchronous training in AI/ML for space biology. Enroll here: https://canvas.instructure.com/enroll/8JYKD7

Collaboration with Robots Go Mental – Guided Transfer Learning (GTL) improves on traditional transfer learning by teaching the neural network “how” to learn through inductive biases that inform how the network learns in the future. We use GTL to achieve high classification performance in the high-dimensionality, low-sample-size settings intrinsic to space biology. Current and future work will focus on expanding GTL to few- and zero-shot learning.

Collaboration with Frontier Development Lab and Intel – The 2021 Frontier Development Lab Astronaut Health team, co-funded by NASA and Intel, developed an approach for federated learning to train predictive models with data on physically separated nodes. This approach enables model training in settings with low compute such as spaceflight, or settings with privacy concerns such as astronaut health. We leverage this approach for training and inference with the first federated machine learning model using data in space on the International Space Station and data held on Earth.

Collaboration with Baranzini Lab at UCSF – Leveraging multi-omics data from NASA GeneLab and integrating it with the Scalable Precision Medicine Oriented Knowledge Engine (SPOKE), the study below marks a step forward in identifying critical physiological changes and symptoms linked to spaceflight. SPOKE is being integrated into OSDR, creating a GeneLab-SPOKE fabric. Building on this approach, we aim to generalize the use of knowledge graphs (KGs) for monitoring and mitigating potential health risks for astronauts. By combining pre-, in-flight, and post-flight biomarker data (cholesterol, PTH, magnesium, phosphorus, etc.) from Earth and space, one should be able to refine actionable countermeasures to ensure astronaut health and safety in future missions.

Leveraging the Generative Adversarial Network (GAN), we develop realistic synthetic RNA sequencing gene expression data to amplify the signals in tiny space biological datasets. Current and future work focuses on evaluating additional architectures such as the Conditional Variational Autoencoder (CVAE).

Supported by the NASA Science Mission Directorate, we publish a standardized, AI-ready RNA sequencing dataset for AI/ML benchmarking, training, and data challenges.

Supported by the NASA Science Mission Directorate, we publish a standardized, AI-ready fluorescence microscopy dataset for AI/ML benchmarking, training, and data challenges.