Fig. 2: Results from AGBT framework and feature analysis.
From: Algebraic graph-assisted bidirectional transformers for molecular property prediction
a, b Illustrate the comparison of the R2 by various methods for the IGC50 set and the LC50DM set, respectively. AGBTs-FP means a supervised fine-tuning process is applied to AGBT-FP. The other results were taken from refs. 2,13,17,23,40,41. c The bar charts illustrate the R2 of AGBT-FPs and BT-FPs with three machine learning algorithms for the IGC50 dataset. d The bar charts illustrate the consensus R2 of AGBT-FPs and AGBTs-FPs with three machine learning algorithms for the LC50DM dataset. e Visualization of the LD50 set. The axes are the top three important features of AGBT-FPs. f Predicted results of AGBT-FPs with MT-DNN model for the IGC50 set(left) and the LC50DM set(right), respectively. The box plots statistic R2 values for n = 358 (left), and 70 (right) independent samples examined over 20 independent machine learning experiments, and the detailed statistic values are listed in Supplementary Table 5. g The AGBT-FPs of the IGC50 and LC50DM datasets were ranked by their feature importance. For both datasets, 188/512 of the AGBT features are from AG-FPs and the remaining 348/512 are from BT-FPs. h Based on AGBT-FPs and AGBTs-FPs, the variance ratios in the first two components from the principal component analysis (PCA) are used to visualize the IGC50 and LC50DM datasets.
