TabOmics reads every feature in one pass with 7.3M parameters: mean AUROC 0.943 across 46 wide omics tasks, and one member matches a model 30 times larger

Prediction on wide omics tables no longer requires shrinking the feature space first: a 7.3M-parameter model can read every retained feature in one forward pass and output class labels for the test rows. Earlier tabular foundation models hit computational and generalization limits on very wide tables and commonly resorted to feature subsampling, grouping, internal chunking or compression into dense row representations, with each TabPFN-3.5 ensemble member handling at most 768 features. On 46 high-dimensional, low-sample-size classification tasks, eight-member TabOmics v2.5 reaches a mean AUROC of 0.943 and beats eight-member TabPFN-3.5 on 32 of the 46 tasks (median difference +0.0011), while that model has about 30 times as many parameters.

The model encodes each table cell as a token, lets features within a row attend to each other, then lets each column attend only to training rows, so test rows never influence one another. This structure follows nanoTabPFN's lightweight reimplementation of TabPFNv2, extended with missing-value indicators, random column embeddings and a test-row flag in the label column. Pretraining uses synthetic tables only: narrow tables first, then batches widened with probability 0.3 to at most 3,350 features, then a lower-learning-rate stage; v2.5 totals 307.5k optimizer steps on 2.46M synthetic tables, about 2.4 A100 GPU-days.

The same forward pass also returns attention-derived feature-importance scores and sample embeddings, so gene rankings and embedding-based classification are available for downstream analysis. Perturbation-based attribution becomes expensive with thousands of features, and TabPFN-Wide averages attention over training and test rows together. On synthetic tables widening from 200 to 30,000 features, feature-recovery AUROC stays between 0.826 and 0.859 with no significant difference from TabPFN-Wide at any width; on eight mRNA tasks, logistic regression on the embeddings reaches a mean AUROC of 0.967 versus 0.972 for direct prediction.

The evaluated scope is high-dimensional, low-sample-size tables: 46 tasks from the HDLSS, MLOmics and Shamir benchmarks, predicting tumor subtypes or histological types rather than clinical outcomes such as survival, disease progression or treatment response. The models are for classification with up to 10 classes; the architecture also supports regression and fine-tuning, but no regression model is released and fine-tuning is not evaluated.

Attention weights are not necessarily faithful explanations, so the scores should be read as feature importance rather than causal or mechanistic importance; the high ranking of PAM50 genes in breast cancer tasks reflects concordance with the genes used to define the labels and is not independent biological validation, and the top-ranked genes were not experimentally validated. Whether the gains come from the TabICLv2 prior or the feature-widening procedure has not been separated by an equal-compute control, and the 46 tasks are not 46 independent cohorts.