Single- and multidomain classification results on 2549 test proteins from the FUpred benchmark
| Method | Multidomain | Single-domain | All | |||
|---|---|---|---|---|---|---|
| Precision | Recall | Precision | Recall | ACC | MCC | |
| ResPRE-Fupreda | 0.860 | 0.873 | 0.936 | 0.929 | 0.910 | 0.799 |
| ESM2-Fupred | 0.631 | 0.974 | 0.982 | 0.716 | 0.802 | 0.651 |
| ESM2-RecCut | 0.663 | 0.941 | 0.963 | 0.761 | 0.821 | 0.663 |
[i] aResPRE-FUpred results are taken from Zheng et al. (2020). In principle, ESM2-FUpred and ESM2-RecCut should generate the same results as they use the same scoring scheme (so-called FUscore). However, there are certain technical details that we cannot replicate in ESM2-RecCut so the results vary slightly. “ACC” and “MCC” are the accuracy and Matthew's correlation coefficient, respectively. The results shown here were based on contact map predictions using the ESM-2 t30 model. Refer to Supplemental Table S1 for the results based on contact map predictions using the ESM-2 t33 model, which gave more accurate domain predictions but performed worse for our task of homology detection.