Fig. 5: Results under the multi-task learning framework.

a The Help-Harm matrix for different combinations of tasks. The values indicate the percentage (unit: %) of improvement using multi-task learning compared to single-task learning (STL) defined by the tasks in rows. The columns represent the paired tasks. We boldfaced blocks with increments larger than 0.5%, which is a threshold reported in ref. ⁹⁶ as an acceptable improvement, and half of the natural threshold 1%. b Comparisons for the results under MTL and STL. The metric for regression tasks, including Loewe and RI_row, is PCC. The metric for the classification task, including Classification, is ROCAUC. c Comparisons for the results under different training settings. Data are presented in boxplots (n = 5 per group; center line, median; box limits, upper and lower quartiles; whiskers, up to 1.5× interquartile range; points, outliers). Here, BAITSAO-FT represents that we fine-tuned the pre-trained model, BAITSAO-ZS represents that we applied the pre-trained model for these tasks under a zero-shot learning framework, and BAITSAO-FS represents that we did not use the pre-trained weights for these tasks. Here, FT means fine-tuning, ZS means zero-shot learning, and FS means from scratch. We included four metrics across three datasets for comparisons. d The first example of tri-drug cases for drug synergy prediction with BAITSAO. e The second example of tri-drug cases for drug synergy prediction with BAITSAO. Logos of cell lines are created with BioRender.com. Source data are provided as a Source Data file.