Fig. 2: Performance of ActFound on the NPs dataset.
a,b, Bar plots comparing ActFound and ActFound Transfer’s performance on the NPs dataset in terms of (a) r2 and (b) RMSE. Plots show the mean ± standard error of the mean (s.e.m.) performance values across the n = 14 assays used to fine-tune the ChEMBL pretrained models and the BindingDB pretrained models. Before plotting, the performance values for the assays were averaged across each shot setting used to fine-tune the models. c,d, Bar plots comparing the models’ performances in terms of r2 at the (c) 8- (n = 14), 16- (n = 14), 32- (n = 10), 64- (n = 4) and 128- (n = 1) shot settings and when (d) 20% (n = 14), 40% (n = 14), 60% (n = 14) and 80% (n = 14) of the assays were used for fine-tuning. Plots show the mean ± s.e.m. performance values across the n number of assays used to fine-tune the pretrained models. Statistical significance between the performance of ActFound and ActFound Transfer was assessed using a one-sided Wilcoxon test. e, t-SNE plot comparing the molecules in the NPs dataset and a random selection of 50% of the ChEMBL training dataset.
