Fig. 3: Effectiveness of DeepTernary designs on PROTAC and MG(D) validation benchmarks.

a Comparison of decoder types: Our proposed pocket points decoder significantly outperforms IEGMN in predicting more medium- to high-quality binding poses (DockQ >0.49). Statistical analysis was performed using the two-sided independent t-test. b Impact of multi-head attention on coordination prediction: A single attention head achieves performance comparable to an eight-head configuration, while reducing computational demands. c Influence of latent embedding dimension on model performance: VPS results indicate that larger dimensions enhance learning learning, especially for MG(D) complexes. d Effect of noise level on model robustness: Increasing the noise level from 1 to 2 improves performance on both PROTAC and MG(D) benchmarks, as reflected by VPS. e Impact of the number of sampled random conformations on test results: An increased number of sampled conformations correlates with higher DockQ scores and acceptance rates (DockQ >0.23) for PROTACs, while MG(D) performance remains largely unchanged. Data are presented as mean values with error bars representing 95% confidence intervals estimated from 22 PROTAC and 94 MG(D) samples. (All box plots in this figure extend from the first quartile (Q1) to the third quartile (Q3) of the data, with a line at the median. The whiskers extend from the box to the farthest data point lying within 1.5x the inter-quartile range (IQR) from the box. Flier points are those past the end of the whiskers. The sample number (n) is annotated under each plot. Source data are provided as a Source Data file.