TABLE 1.
T1044: (A) Confidence scores (pLDDT) for different prediction systems that were considered and (B) accuracy (GDT_TS) for predictions of domains in T1044
| (A) Confidence (pLDDT) a | |||||
|---|---|---|---|---|---|
| Original domain | Original full sequence | Crop‐then‐fold domain | Submitted full sequence | Final system full sequence | |
| T1031 | 80.6 | 64.4 | 84.1 | 71.9 | 69.4 |
| T1033 | 54.7 | 69.0 | 85.0 | 73.9 | 77.2 |
| T1035 | 81.7 | 78.8 | 83.2 | 77.9 | 82.6 |
| T1037 | 83.9 | 78.7 | 89.7 | 77.7 | 82.8 |
| T1039 | 75.0 | 68.2 | 89.3 | 67.9 | 71.6 |
| T1040 | 73.1 | 51.6 | 87.2 | 82.0 | 74.7 |
| T1041 | 83.7 | 78.4 | 86.6 | 77.8 | 80.0 |
| T1042 | 79.7 | 68.1 | 81.5 | 80.6 | 73.5 |
| T1043 | 37.5 | 47.4 | 79.7 | 82.8 | 64.1 |
| Average | 72.2 | 67.2 | 85.2 | 76.9 | 75.1 |
| Full sequence | N/A | 71.3 | N/A | 77.1 | 77.4 |
| (B) Accuracy (GDT_TS) b | |||||
|---|---|---|---|---|---|
| Original domain | Original full sequence | Crop‐then‐fold domain | Submitted full sequence | Final system full sequence | |
| T1031 | 87.6 | 86.6 | 86.8 | 87.1 | 88.4 |
| T1033 | 44.0 | 85.7 | 89.0 | 87.7 | 87.0 |
| T1035 | 92.6 | 94.1 | 93.1 | 94.9 | 95.8 |
| T1037 | 82.9 | 88.4 | 85.0 | 87.3 | 92.6 |
| T1039 | 79.2 | 81.4 | 78.9 | 82.3 | 82.9 |
| T1040 | 55.0 | 30.8 | 70.0 | 71.7 | 69.2 |
| T1041 | 86.6 | 89.7 | 85.8 | 90.5 | 89.6 |
| T1042 | 62.1 | 80.2 | 69.9 | 83.8 | 90.4 |
| T1043 | 16.6 | 53.2 | 76.2 | 83.3 | 82.3 |
| Average | 67.4 | 76.7 | 81.6 | 85.4 | 86.5 |
| Full sequence (TM score) | N/A | 0.807 | N/A | 0.878 | 0.960 |
T1044: Confidence scores (pLDDT) for different prediction systems that were considered. The mean full‐sequence pLDDT over a given domain cannot be directly compared to the mean pLDDT found just by folding that domain, as pLDDT will consider the effect of mispredicting inter‐domain distances as well as intra‐domain distances, which penalizes longer predictions. However, it can be seen that using “crop‐then‐fold” led to an improvement, often substantial, in confidence across all domains. The full sequence confidences of predictions made with the submitted (template‐patched) system were also superior to the original system. The final improved system gives an equivalent level of confidence to the submitted prediction.
Accuracy (GDT_TS) for predictions of domains in T1044. It can be seen that, using the original system, T1033 was predicted more accurately as part of the full chain, but T1040 was predicted more accurately when folded as an independent domain. Both crop‐then‐fold domains and submitted (template‐patched) full‐sequence predictions get the best of both worlds and give better mean domain accuracy. The final system gets equivalent performance with no complex interventions, and better chain‐level TM score.