GLM/PLM Embeddings

TEM-1 Activity (ESM-2)

Single-amino-acid mutants of TEM-1 beta-lactamase are represented by ESM-2 embeddings averaged across the protein sequence. The regression target is each mutant's scaled experimental activity value from the PEER benchmark.

RegressionLicence: PEER/TorchDrug source terms; ESM-2 model MIT; redistribution authorized by dataset curatorlocal-PEER-BetaLactamase-ESM2Target: activitylocal ↗

At this operating point

Model comparison

p=10,000, n=100 · Lower RMSE 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	RMSE	MAE	R²
Kumo Tabular (M)	0.2804	0.2008	0.196
Causilo	0.2815	0.2173	0.1901
TabPFN 3.5	0.2824	0.2201	0.1868
RealTabPFN v2	0.2828	0.2174	0.1844
AutoGluon	0.284	0.2041	0.178
TabFM	0.2861	0.2028	0.1619
RealTabPFN 2.5	0.2866	0.2143	0.1584
MLP (tuned)	0.2908	0.2118	0.1371
MLP	0.2924	0.2138	0.1252
TabDPT	0.2925	0.2226	0.1302
Extra Trees (tuned)	0.2937	0.2388	0.1247
MITRA	0.2947	0.215	0.1114
TabM	0.296	0.2235	0.1091
LimiX2	0.2962	0.2013	0.1099
GeneICL	0.2991	0.2414	0.08352
Extra Trees	0.3004	0.2464	0.08113
Random Forest (tuned)	0.3018	0.2471	0.07175
Logistic Regression (tuned)	0.3034	0.2362	0.05738
Random Forest	0.3045	0.2471	0.05386
RealMLP	0.3058	0.2327	0.04459
XGBoost (tuned)	0.3067	0.2407	0.04708
XGBoost	0.3113	0.2514	0.01494
Logistic Regression	0.3327	0.2582	-0.1407