Here is the SOD1 protein, mutated A4V:
MATKVVCVLKGDGPVQGIINFEQKESNGPVKVWGSIKGLTEGLHGFHVHEFGDNTAGCTSAGPHFNPLSRKHGGPKDEERHVGDLGNVTADKDGVADVSIEDSVISLSGDHCIIGRTLVVHEKADDLGKGGNEESTKTGNAGSRLACGVIGIAQ
Using PepMLM, I generated 4 candidate binders of length 12:
| Binder | Pseudo Perplexity | |
|---|---|---|
| 0 | WLYYVAAARHKX | 15.758440 |
| 1 | HHYYAAVAAHWK | 14.330523 |
| 2 | HRYPAVAAAWKX | 7.247497 |
| 3 | WLYGPTAVAWKX | 9.601791 |
Plus one more known binder: FLYRWLPSRRGG.
Folding the protein with AlphaFold-Multimer with the alphafold2_multimer_v3 model:
2025-03-11 05:00:50,486 reranking models by 'multimer' metric
2025-03-11 05:00:50,486 rank_001_alphafold2_multimer_v3_model_3_seed_000 pLDDT=91.1 pTM=0.888 ipTM=0.306
2025-03-11 05:00:50,487 rank_002_alphafold2_multimer_v3_model_5_seed_000 pLDDT=91 pTM=0.889 ipTM=0.302
2025-03-11 05:00:50,487 rank_003_alphafold2_multimer_v3_model_1_seed_000 pLDDT=91.2 pTM=0.89 ipTM=0.291
2025-03-11 05:00:50,487 rank_004_alphafold2_multimer_v3_model_2_seed_000 pLDDT=91.2 pTM=0.89 ipTM=0.288
2025-03-11 05:00:50,487 rank_005_alphafold2_multimer_v3_model_4_seed_000 pLDDT=89.3 pTM=0.874 ipTM=0.145
2025-03-11 05:00:51,925 Done
Here is the highest-ranking model 3:

What does it mean that the pLDDT of the binder is very low? Can AlphaFold not fold it?
2025-03-11 05:15:35,344 reranking models by 'multimer' metric
2025-03-11 05:15:35,344 rank_001_alphafold2_multimer_v3_model_2_seed_000 pLDDT=91.4 pTM=0.883 ipTM=0.213
2025-03-11 05:15:35,345 rank_002_alphafold2_multimer_v3_model_5_seed_000 pLDDT=91.2 pTM=0.88 ipTM=0.209
2025-03-11 05:15:35,345 rank_003_alphafold2_multimer_v3_model_1_seed_000 pLDDT=91.2 pTM=0.881 ipTM=0.189
2025-03-11 05:15:35,345 rank_004_alphafold2_multimer_v3_model_3_seed_000 pLDDT=90.7 pTM=0.877 ipTM=0.168
2025-03-11 05:15:35,345 rank_005_alphafold2_multimer_v3_model_4_seed_000 pLDDT=90 pTM=0.874 ipTM=0.156
2025-03-11 05:15:36,400 Done
Here is the highest-ranking model 2:

For both my generated binder and the known binder, the ipTM scores were rather low (0.306 and 0.213). I think > 0.8 is considered a high-quality prediction, so does that mean that AlphaFold has no idea if the binders would bind to the protein?
Also, in both cases, the pLDDT is rather low — so AlphaFold has no idea what its structure is either?