Physics-Inspired Deep Learning for Genomics: Transforming High-Dimensional Tabular Data into Spatial Maps for Precision Medicine
Abstract
Purpose
High-dimensional genomic and clinical tabular data lack spatial structure, limiting the effectiveness and interpretability of modern deep learning. In medical physics, raw measurements are routinely transformed into structured physical representations (e.g., sinograms, k-space, dose maps) prior to inference. Building on prior work, GenoMap extends this paradigm to non-imaging biomedical data by transforming unordered feature vectors into spatially organized ‘genomic images,’ enabling convolutional learning and visualization. This work benchmarks GenoMap against TabPFN, a state-of-the-art transformer foundation model for tabular data, under clinically realistic conditions.
Methods
Twenty-two classification datasets, including large-scale and high-dimensional genetic cohorts, were evaluated using 5-fold cross-validation with four random seeds. Performance was assessed under (1) progressive subsampling (80%, 60%, 40%, 20%) to emulate limited clinical data and (2) adaptive Gaussian noise injection (5–20%) to simulate biological and measurement variability. GenoMap reshaped tabular features into dataset-specific spatial maps processed by CNNs, while TabPFN operated directly on tabular inputs. Mean classification accuracy and runtime were reported.
Results
GenoMap achieved superior accuracy on large and high-dimensional datasets, reaching up to 99.4% on genetic benchmarks and outperforming TabPFN when full feature spaces were preserved. TabPFN demonstrated strong robustness in low-sample and noisy regimes, maintaining competitive performance with minimal training. Under subsampling, GenoMap retained performance advantages when sufficient data were available, while TabPFN excelled in extreme low-resource conditions. Several full-scale genomic datasets exceeded TabPFN’s memory limits, whereas GenoMap scaled efficiently.
Conclusion
This study establishes GenoMap as a physics-inspired bridge between genomics and medical imaging. By transforming abstract biomedical vectors into spatial fields, GenoMap enables deep spatial learning, visualization, and scalability for precision medicine. The results define complementary AI strategies: GenoMap for high-fidelity, large-scale clinical genomics and TabPFN for rapid, low-resource inference. This work extends core medical physics principles, which are representation, transformation, and signal structure, into next-generation computational diagnostics.