Tabular learning × biomedicine

A benchmark for high-dimensional biomedical tables.

TabBench-Bio compares classical models, neural networks, and tabular foundation models across transcriptomic, metagenomic, molecular, sequence-derived, and genomic-prediction datasets on a controlled feature-by-sample grid.

Read on arXiv View on GitHub Request dataset / model

—biomedical datasets

—model configurations

—evaluation points per model

5×cross-validation

Interactive results

Model rankings depend on the operating point.

Leaderboard

Bradley–Terry Elo

Metric Dataset modality Feature cap Training samples

Random Forest is anchored at 1,000. Error bars are target-bootstrap 95% intervals.

Budget response

Performance across sample budgets

Dataset modality Feature cap Models Uncertainty95% CI

Budget response

Performance across feature budgets

Efficiency

Performance vs. prediction cost

Metric Dataset modality Feature cap Training samples

Efficiency

Performance vs. fitting cost

Stability

Rank stability across operating points

Reference cell

Coverage

Datasets by modality

Snapshot scope. The live counter covers the configured benchmark run; performance charts remain pinned to the last status-complete aggregation until the cells finish.

Tuning protocol. Tuned models use a small predefined grid under a one-hour budget. Candidates are selected on a training-only holdout, then the winner is refitted on the full training fold.

System configuration. NVIDIA L40S with 48 GB VRAM, 16 CPU cores and 80 GB RAM.

Strict website default. Performance charts and rankings open in Strict (primary) mode, using results at the requested feature and sample budgets. Failures are scored with the constant predictor. Adaptive remains available as a sensitivity view and may reuse a verified smaller sample budget after a memory failure. Dataset-specific exports and the downloadable LaTeX leaderboard also use Strict results.