Figure 8. Improving k_cat and K_M prediction for ADK sequences with semi-supervised learning.

The mean test spearman rho for different models across 30 samplings is plotted against training dataset size for (A) k_cat and (B) K_M. Models include Random Forest (RF), Support Vector Regression (SVR), ProteinNPT(PNPT) (107), and Convolutional Linear Regression (CLR) (107). For embeddings, SVR used a One-hot encoded MSA, RF used ESM-2 embeddings, and CLR and PNPT used Tranception(112) embeddings. Embeddings and other model hyperparameters were selected based on aggregate (mean) performance for both k_cat and K_M prediction. Shaded regions represent 95% confidence intervals across 30 training/test set samplings at each dataset size (Methods). A zero-shot evaluation of the Tranception PLM (112) is plotted as a dashed orange line. DLKcat (109) performance evaluated on all 175 sequences is plotted in (A) as a dashed black line.