T-SMmOTE: tweaked synthetic majority minority oversampling technique for data scarcity issue in multi omics studies

Multiomics data offer a rich data mine for modeling complex as well as day-to-day diseases, but their practical deployment is constrained by the limited sample availability. To this end, generating synthetic samples is a viable remedy. Extant schemes operating along this line, however, are mostly limited to augmenting the minority class in imbalanced datasets and often produce synthetic samples that lack sufficient diversity and fail to faithfully capture the underlying data distribution. As a result, the full potential of synthetic augmentation in multi-omics learning remains underexplored. The aim is to address the data scarcity problem in multi-omics domain. We propose a synthetic oversampling framework, which is dedicated to addressing overall data scarcity in multi-omics datasets and the lack of diversity in synthetic samples. Contrary to conventional methods that restrict augmentation to minority classes and rely on interpolation of two neighbors, our method generates diverse yet distribution-aligned synthetic samples by interpolating three neighbors and extends this augmentation paradigm to the majority class. The framework first balances the dataset by generating synthetic minority samples, and subsequently augments the balanced dataset by oversampling both majority and minority classes.

Empirical evaluation on multi-omics data obtained from three heterogeneous health scenarios-inflammatory bowel disease, multi-organ dysfunction syndrome, and colorectal cancer-substantiates the utility of the proposed scheme in improving the predictive performance. The models trained on T-SMmOTE-augmented data achieve higher Matthews correlation coefficient values, along with improved $F_1$ scores for both majority and minority classes. Notably, oversampling of the majority class improves the cognition of the minority class as well. We also explore the consistency of the class distributions between the original and augmented class-specific datasets. These findings confirm the capability of our scheme to learn from small, high-dimensional multi-omics datasets and highlight its potential for non-invasive disease detection.

https://github.com/payelu/TSMm.