GLM/PLM Embeddings

Gene Fusion Breakpoints (NT-v2)

Each row pairs DNA sequences around two candidate gene-fusion breakpoints, represented by concatenated Nucleotide Transformer v2 embeddings. The task distinguishes fusion-positive pairs from constructed pseudo-fusions, whose partner genes were not known to participate in fusions.

Binary classificationLicence: CC BY 4.0FusionAI-NTv2-GeneFusionTarget: fusionfusionai ↗

At this operating point

Model comparison

p=10,000, n=100 · Higher Macro-F1 is better

Random Forest is marked with a diamond. Models are ordered by the selected metric.

Across sample budgets

Learning curves

Models

All metric values at this operating point

Model	Macro-F1	MCC	Balanced accuracy	ROC-AUC
Kumo Tabular (M)	0.9273	0.8549	0.9273	0.9792
MLP (tuned)	0.9257	0.8519	0.9257	0.9793
MLP	0.9254	0.851	0.9254	0.9784
Logistic Regression (tuned)	0.925	0.8512	0.925	0.9791
RealMLP	0.9249	0.8511	0.9249	0.9738
TabPFN Wide 5k (ne3)	0.9245	0.8498	0.9245	0.9794
Causilo	0.9244	0.8493	0.9244	0.9791
Logistic Regression	0.9239	0.8493	0.9239	0.9794
GeneICL	0.9237	0.8482	0.9237	0.9784
TabPFN 3.5	0.9235	0.8474	0.9235	0.9786
MITRA	0.9232	0.8472	0.9233	0.9784
Extra Trees	0.9214	0.8436	0.9214	0.9783
Random Forest (tuned)	0.9212	0.8428	0.9212	0.9767
TabDPT	0.921	0.8423	0.921	0.9726
Extra Trees (tuned)	0.92	0.8414	0.92	0.9778
RealTabPFN 2.5	0.9199	0.8406	0.92	0.9753
Random Forest	0.9189	0.8383	0.9189	0.9763
AutoGluon	0.9188	0.8389	0.9188	0.9721
TabFM	0.9187	0.839	0.9187	0.9777
RealTabPFN v2	0.9185	0.8383	0.9185	0.9757
TabM	0.9158	0.8348	0.916	0.9775
XGBoost (tuned)	0.8378	0.6775	0.8381	0.8891
XGBoost	0.8233	0.6492	0.8237	0.8841
LimiX2	0.3333	0	0.5	0.5