Tabular learning × biomedicine

A benchmark for high-dimensional biomedical tables.

TabBench-Bio benchmarks machine learning methods on high-dimensional biological and biomedical tables. It 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.

46biomedical datasets
34model configurations
6,440evaluation points per model
5×cross-validation

Interactive results

Model rankings depend on the operating point.

Primary analysis: each score comes from the requested feature- and sample-cap cell; no smaller cell is substituted. All failed fits, including tuned models, use the constant predictor.

Leaderboard

Bradley–Terry Elo

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

Budget response

Performance across sample budgets

Budget response

Performance across feature budgets

Efficiency

Performance vs. prediction cost

Efficiency

Performance vs. fitting cost

Stability

Rank stability map

Coverage

Datasets by modality

Methods and interpretation

About the benchmark

Biological tables often contain many more features than labelled samples, a setting known as high-dimensional, low-sample-size (HDLSS) learning. TabBench-Bio evaluates classification and regression across gene expression, DNA methylation, metagenomics, molecular properties, genomic prediction, and sequence-derived embeddings. Browse the biological dataset registry for individual tasks and data sources.

Which machine learning models perform best on my biomedical data?

The best-performing model depends on the modality, number of training samples, feature budget and evaluation metric. Use the interactive leaderboard to compare models at the operating point closest to your data (select modality, sample and feature count), and check uncertainty intervals and failure rates alongside the rankings. This should be a good starting point for model selection.

How are models compared?

Each result belongs to an operating point: a feature cap (2,000, 10,000, 25,000 or all features) and a training-sample budget (20 to 1,000 rows, or the full training split). Every model is fitted on the same frozen five-fold cross-validation splits, so all comparisons are paired by dataset and fold. The reference operating point is 10,000 features and 100 training samples.

How should I interpret the rankings?

Elo ratings come from a Bradley–Terry model fitted to fold-level comparisons, using macro-F1 for classification and RMSE for regression. Random Forest is anchored at 1,000. A 400-point difference corresponds to 10:1 odds that the higher-rated model wins a comparison. Elo is not an accuracy score. The 95% intervals resample whole targets; overlapping intervals do not establish a clear winner. Rankings change with the feature and sample budget, so compare models at the operating point closest to your own data. In the Strict view, failed fits are scored as the constant baseline and count against the model. The Adaptive and Failures-omitted views are sensitivity analyses. Compare classical baselines, neural networks, AutoML and tabular foundation models in the interactive leaderboard and the model comparisons.

How do I benchmark my own model?

Define create_model(task) in a Python file. It should return a cloneable scikit-learn classifier or regressor, with imputation, scaling and feature selection inside the pipeline. The evaluation downloads the published reference results and fits only your model on the frozen folds, then writes an HTML report with Elo rankings, plots and per-dataset metrics.

Command line

Install and run

Guide
pip install tabbench-bio
tabbench-bio evaluate --model my_model.py --task overall

Minimal example

my_model.py

Guide
from sklearn.impute import SimpleImputer
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.svm import SVC, SVR


def create_model(task):
    model = SVC() if task == "classification" else SVR()
    return make_pipeline(
        SimpleImputer(strategy="median", keep_empty_features=True),
        StandardScaler(),
        model,
    )

How can I reuse the benchmark results?

Explore results on the benchmark leaderboard and use the artifacts page to download the underlying evidence. Automated tools can read the benchmark guide, dashboard JSON, leaderboard JSON, and dataset metadata index without running browser scripts. Use the model-selection agent skill to match a comparison to your task and budget, and the citation page to cite the benchmark.