Chimdi Walter Ndubuisi
University of Missouri, Columbia
Research Affiliate, Lawrence Berkeley National Laboratory
I develop methods at the intersection of topology, differential geometry, and deep learning for scientific imaging, primarily in plant phenotyping. Working with maize lesion mimic mutants as a model system, I build physics-informed segmentation pipelines, including a novel Navier-Stokes active-contour framework, to extract high-fidelity lesion masks at scale. These masks are mapped onto phenotypic manifolds that link visual morphology to underlying biochemical pathways. Closing the loop between what a camera sees and what a genome encodes.
The long-term aim is to resolve breeding bottlenecks that limit crop improvement and food sustainability. The geometric and topological toolkit transfers directly: I apply the same ideas to single-cell transcriptomics, oncology, omics feature selection, and causal discovery across biological domains.
Physics-Informed Segmentation
GPU-accelerated Navier-Stokes active contours with vorticity-based proxy scoring, wavelet-diffusion-SSL fusion pipelines, and semi-amortized inference for uncertainty-guided phenotyping, deployed on HPC clusters via containerized SLURM workflows.
Topology & Geometry for Vision
Product-manifold latent spaces (hyperbolic × Euclidean × spherical), HyperTopo-Adapters for frozen encoders, and persistent-homology descriptors that fuse morphological, colorimetric, and topological features for lesion characterization.
Omics & Computational Biology
Geometry-sensitive sparse persistent representatives (CC-SPR) for interpretable omics feature selection; Ricci-curvature classifiers for diffusion geometry; causal discovery on genome-scale Perturb-seq; single-cell modeling at the geometry-dynamics-interpretability tradeoff.
Multimodal Protein AI , LBNL
Building SE(3)-aware multimodal foundation models that integrate SAXS profiles, protein sequence, 3D structure, and conformational ensembles. Generating physically plausible conformers via elastic-network normal-mode displacement and validating against experimental scattering data.
- Ndubuisi, C.W. and Kazic, T. (2026). HyperTopo-Adapters: Geometry- and Topology-Aware Segmentation of Leaf Lesions on Frozen Encoders. arXiv:2601.06067. arXiv Code
- Ndubuisi, C.W., Kazic, T., Kharismawati, D.E. (2023). Imaging Maize Lesions. North American Plant Phenotyping Network (NAPPN). DOI
- Somoye, O.I., Ndubuisi, C.W., Donatus, S., Amakye, F. (2025). Scalable Machine Learning Algorithms for Processing High-Dimensional Data. Asian J. Adv. Res. & Reports. DOI
- Eleweke, I., Umakor, M.F., Ndubuisi, C.W., et al. (2025). AI-Driven Threat Detection and Prevention in Cloud Computing Environments. AJISE.
- Ndubuisi, C.W. (2026). Geometric Decoding of Pathology Foundation Models Reveals Morphology-Encoded Gene Programmes in Breast Cancer. Code
- Ndubuisi, C.W. (2026). Task-Sufficient Manifold Distillation of Learned Representations.
- Ndubuisi, C.W. (2026). When Does Nonlinearity Matter? Structural Accessibility and the Geometry-Dynamics-Interpretability Tradeoff in Single-Cell Modeling.
- Ndubuisi, C.W. (2026). Ontology-Guided Pathway Attention Reveals a Cell-Intrinsic Inflammatory Axis of Kinase Inhibitor Sensitivity. Code
- Ndubuisi, C.W. (2026). Geometric-Relational Diagnosis Spaces for Multi-LLM Clinical Reasoning. Code
- Ndubuisi, C.W. (2026). A Reproducible Benchmark of Geometry-Aware Causal Discovery for Genome-Scale Perturb-seq.
- Ndubuisi, C.W. and Olaoye, D. (2026). Geometry-Sensitive Persistent Representatives for Omics: A Sparse Extension of Harmonic Persistent Homology with Multi-Backend Geometry. Code
- Ndubuisi, C.W. and Kazic, T. (2026). Gradient-Free Stochastic Optimization of Navier-Stokes Active Contours for Unsupervised High-Throughput Phenotyping.
- Ndubuisi, C.W. and Kazic, T. (2026). Selective Semi-Amortized Inference for Physics-Based Lesion Phenotyping: Uncertainty-Guided Runtime-Quality Tradeoffs.
- Ndubuisi, C.W. (2026). Trajectory Analysis of Alveolar Epithelial State Coherence in Lethal COVID-19. Code
HPC Segmentation Suite
Containerized wavelet + NS-diffusion + active contours + SSL/SAM pipeline. Docker/Apptainer, reproducible SLURM on Hellbender & Lewis.
Ricci-Curvature Classification
Ollivier and Forman Ricci curvature as geometric features for supervised classification of omics and imaging data.
ARC-AGI Hybrid Solver
Symbolic-first DSL planner with LLM fallback and ViT perception. ~65% pixel accuracy on unseen test data. Top-6 finalist, MU Hackathon.
GOPA
Deterministic GO-projected pathway representations for cancer drug response prediction via ontology-guided attention.
Pathology Atlas
Geometric decoding of pathology foundation models to reveal morphology-encoded gene programmes in breast cancer.
CC-SPR
Geometry-sensitive sparse persistent representatives for interpretable omics feature selection via persistent homology.
- Honorable Mention, Nexus Informatics Conference (2026)
- Wolfram Summer Research Institute (2026)
- Vice-Programming Director, Graduate Professional Council (2024-25)
- Finalist (Top 6), ARC-AGI-2 Hackathon (2025)
- Second-Best Poster, DAREC Symposium (2025)
- EECS Departmental Fellowship (2021)
- Gold Medal & Best African Student, Intl. Junior Science Olympiad (2010)