NASA TOPS-T ScienceCore: AI/ML in Space Biology — Notebook Suite

Overview

This notebook suite was completed as part of the NASA Transform to Open Science (TOPS-T) – ScienceCore: AI/ML in Space Biology Training.
It walks from data fundamentals → visualization/EDA → tabular ML (classification & regression) → image data → clustering/unsupervised learning → explainable AI (XAI), with emphasis on open-science practices and NASA space-biology datasets (OSDR / GeneLab).

---

Notebook Index (what each file covers)

• Introduction to Data
  Intro to data types & metadata; context from space-biology use cases (e.g., SANS).

• Data Visualization
  Exploratory plots, distributions, pairplots, correlation heatmaps; communicating results.

• Tabular Data
  Loading/cleaning tabular data; pandas DataFrames; basic feature engineering.

• Classification
  Supervised learning for discrete labels (train/test split, metrics, model comparison).

• Regression
  Supervised learning for continuous targets (error metrics, baselines, regularization).

• Image Data
  Bioimaging pipelines: reading, preprocessing, and basic deep-learning scaffolding.

• Clustering
  Unsupervised learning; dimensionality reduction (e.g., PCA), clustering & evaluation.

• Explainable AI
  XAI for space-bio signals: SHAP/LIME, permutation importance, partial dependence.

• Bioinformatic Tools
  Space-bio tooling overview; dimensionality reduction; ties to downstream ML/XAI.

---

Space Biology Topics Covered

• Spaceflight vs. Ground Control comparisons (rodent models)
  Using OSDR/GeneLab datasets to contrast flight and ground samples, especially in the eye/retina context (GLDS/OSD-255 RNA-seq; metadata labels “Space Flight” vs “Ground Control”).

• Spaceflight-Associated Neuro-ocular Syndrome (SANS)
  Intro and background on SANS, its relevance to astronaut vision, and why ocular/retinal endpoints matter in spaceflight studies.

• Rodent Research-9 (RR9) mission focus
  RR9 mission context and data usage; analyses built around RR9 ocular outcomes and retinal tissue.

• Retina/ocular phenotyping

  • Tonometry / Intraocular Pressure (IOP) measures (OSD-583).
  • Immunostaining microscopy (PECAM/CD31) as an endothelial/vascular marker (OSD-568).
  • Tomography/ophthalmic imaging references (OSD-557).
    These phenotypes are linked to flight status and used as targets/labels in ML.

• Transcriptomics under spaceflight (RNA-seq)

  • Normalized gene-expression matrices from GLDS-255 / OSD-255 (mouse retina).
  • Differential expression filtering (DE/“DGEA”) to enrich flight vs. ground signals.
  • Integrating RNA-seq with phenotypes (e.g., predicting PECAM or thresholds from transcriptomic features).

• Biological pathways & signatures
  Pathway-level interpretation (e.g., MSigDB inputs) suggesting ties to vascular/angiogenesis and muscle/structure phenotypes relevant to microgravity adaptation.

• Unsupervised signatures of spaceflight

  • PCA visualizations separating flight/ground samples.
  • k-means clustering on RNA-seq (with outlier checks), highlighting whether flight status forms distinct molecular clusters.

• Explainable models connecting genes → ocular phenotypes

  • Linear/Ridge regression and Random Forest models that map RNA-seq to PECAM microscopy readouts.
  • SHAP and permutation importance to identify gene drivers most associated with ocular/vascular endpoints under spaceflight.

• Bioimaging data handling
  Basic pipelines for microscopy/image data (loading, preprocessing), with ties to molecular data where possible.

---

Datasets Used

NASA Open Science Data Repository (OSDR) / GeneLab references:

• OSD-255 (NASA OSDR)
• OSD-557 / OSDR-557 (NASA OSDR)
• OSD-568 (NASA OSDR)
• OSD-583 (NASA OSDR)

---

Methods, Algorithms & Techniques

Data handling & EDA

• pandas, numpy for tabular data
• matplotlib, seaborn for visualization (histograms, box/violin plots, pairplots, correlation heatmaps)

Preprocessing

• Feature scaling/normalization: StandardScaler, MinMaxScaler, RobustScaler
• train_test_split; Pipeline / ColumnTransformer patterns (where applicable)

Dimensionality reduction

• PCA (principal component analysis)
  (t-SNE/UMAP may be added if used in future iterations.)

Supervised learning — Classification

• LogisticRegression, SVC, KNeighborsClassifier
• DecisionTreeClassifier, RandomForestClassifier
• Linear models like Perceptron / SGDClassifier
• Model selection: GridSearchCV / RandomizedSearchCV

Supervised learning — Regression

• LinearRegression, Ridge, Lasso (regularization)
• Tree-based/ensemble regressors when included
• SVR / SGDRegressor

Unsupervised learning

• KMeans
• silhouette_score (cluster quality)
• PCA projections for visualization of clusters

Deep learning

• Framework scaffolding with PyTorch and Keras/TensorFlow
• Basic image preprocessing (PIL / scikit-image)

Explainable AI (XAI)

• SHAP
• LIME
• Permutation importance
• Partial dependence

---

Reproducibility & Open-Science Practices

• Usage of open datasets (OSDR / GeneLab IDs cited in code).
• Clear cell ordering and markdown annotations to promote transparent workflows.
• Explicit model metrics for comparability across runs.
• Programmatic access patterns (e.g., OSDR API) for repeatable data retrieval.

---

How to run locally

1. Create a Python 3.10+ environment and install the dependencies above (pip install -U pandas numpy matplotlib seaborn scikit-learn shap lime torch torchvision torchaudio tensorflow keras scikit-image s3fs as needed).
2. Launch Jupyter: jupyter lab or jupyter notebook.
3. Open a notebook and run top-to-bottom.
   • For OSDR/GeneLab data, ensure network access and any required API permissions/paths.
   • If a dataset is missing, consult the data-loading cell for links/instructions.

---

Acknowledgments

Built during the NASA TOPS-T ScienceCore AI/ML in Space Biology training. Thanks to the open-science ecosystem (OSDR/GeneLab) enabling reproducible education & research.