How can I see cyclic peptide structure prediction?

[kimdn]

I installed localcolabfold and copied HighFold git repo into my localcolabfold.

So that my localcolabfold git repo looks as

Then I ran

input_fasta="/people/kimd999/script/python/predict_structure/protein_only/HighFold/input.fasta"
colabfold_batch --model-type alphafold2 $input_fasta path_output

where my input.fasta looks

>fasta
XXXXXXXXXX

(here 'X' means any AA of mine. My input sequence has only 1 cysteine. Therefore, I don't expect internal S-S bond)

However, all 5 predicted peptide structures are linear (i.e., not expected cyclic peptide).

I wonder whether I run HighFold correctly.
Or I ran HighFold well, just my input AA sequence is indeed predicted to be linear according to HighFold.

[kimdn]

Can you please clarify "and then copy the source codes of HighFold into the installed ColabFold preject." ?

Even if I remove HighFold folder from my localcolabfold folder,

input_fasta="/people/kimd999/script/python/predict_structure/protein_only/HighFold/input.fasta"
colabfold_batch --model-type alphafold2 $input_fasta path_output

still runs.

However, I cannot reproduce your paper's HighFold results with

and

I could reproduce your

and

HighFold results only.

[NumnumnumM]

I installed localcolabfold and copied HighFold git repo into my localcolabfold.

So that my localcolabfold git repo looks as

Then I ran

input_fasta="/people/kimd999/script/python/predict_structure/protein_only/HighFold/input.fasta"
colabfold_batch --model-type alphafold2 $input_fasta path_output

where my input.fasta looks

>fasta
XXXXXXXXXX

(here 'X' means any AA of mine. My input sequence has only 1 cysteine. Therefore, I don't expect internal S-S bond)

However, all 5 predicted peptide structures are linear (i.e., not expected cyclic peptide).

I wonder whether I run HighFold correctly. Or I ran HighFold well, just my input AA sequence is indeed predicted to be linear according to HighFold.

Can you please clarify "and then copy the source codes of HighFold into the installed ColabFold preject." ?

Even if I remove HighFold folder from my localcolabfold folder,

input_fasta="/people/kimd999/script/python/predict_structure/protein_only/HighFold/input.fasta"
colabfold_batch --model-type alphafold2 $input_fasta path_output

still runs.

However, I cannot reproduce your paper's HighFold results with and

I could reproduce your

and

HighFold results only.

You should replace the colabfold and alphafold files provided by the author into the localfold directory. The specific location is:
/your/path/localcolabfold/colabfold-conda/lib/python3.xx/site-packages/colabfold & alphafold.
Remember to backup the files that will be replaced before making the changes. After the replacement, some package imports might show errors. You can modify these based on the original files.

[coliva92]

You should replace the colabfold and alphafold files provided by the author into the localfold directory. The specific location is:
/your/path/localcolabfold/colabfold-conda/lib/python3.xx/site-packages/colabfold & alphafold.

I suggest adding that bit of information to the readme file in the repository to make the installation instructions clearer.

[rcossio]

Just in case anyone goes through the same, here is how I solved it:

I installed localcolabfold and pasted colabfold and alphafold folders in the python3.10 site-packages folder.

I get an import error from the line

from Bio.Data import SCOPData

because Bio.Data has no SCOPData library. The original colabfold has

from Bio.Data import PDBData

So I replaced the line with

from Bio.Data import PDBData as SCOPData

With this change the software will run but in the relaxation of structures another error will be raised, it is a problem importing OpenMM, I changed

from simtk import openmm
from simtk import unit
from simtk.openmm import app as openmm_app
from simtk.openmm.app.internal.pdbstructure import PdbStructure

with

import openmm
from openmm import unit
from openmm import app as openmm_app
from openmm.app.internal.pdbstructure import PdbStructure

[ahaldane]

Hi, I'm trying to confirm I've installed Highfold correctly by reproducing the result for 7m3u shown above and in the paper. But it's not working, can anyone see what I'm missing?

I copied the alphafold folder to "localcolabfold/colabfold-conda/lib/python3.9/site-packages/alphafold", and the colabfold folder to "localcolabfold/colabfold-conda/lib/python3.9/site-packages/colabfold".

Then I run

colabfold_batch --amber --model-type alphafold2 7m3u.fa out_7m3u_notempl

where 7m3u.fa contains:

>7M3U_1
GFCWQHTCLPSGCADFPWPVGHQCFPD

It looks like it is using the right code because I see output about "offset_ss_monomer", here's my log:

run log

2025-02-02 15:24:29,007 Running on GPU
2025-02-02 15:24:29,091 Found 5 citations for tools or databases
2025-02-02 15:24:29,091 Query 1/1: 7M3U_1 (length 27)
2025-02-02 15:24:29,092 Setting max_seq=2, max_extra_seq=1
dict_keys(['aatype', 'between_segment_residues', 'domain_name', 'residue_index', 'seq_length', 'sequence', 'deletion_matrix_int', 'msa', 'num_alignments', 'msa_species_identifiers', 'template_all_atom_positions', 'template_all_atom_masks', 'template_sequence', 'template_aatype', 'template_confidence_scores', 'template_domain_names', 'template_release_date', 'template_sum_probs', 'asym_id'])
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1738527885.500319 3316591 mlir_graph_optimization_pass.cc:401] MLIR V1 optimization pass is not enabled
dict_keys(['aatype', 'residue_index', 'seq_length', 'is_distillation', 'seq_mask', 'msa_mask', 'msa_row_mask', 'random_crop_to_size_seed', 'atom14_atom_exists', 'residx_atom14_to_atom37', 'residx_atom37_to_atom14', 'atom37_atom_exists', 'extra_msa', 'extra_msa_mask', 'extra_msa_row_mask', 'bert_mask', 'true_msa', 'extra_has_deletion', 'extra_deletion_value', 'msa_feat', 'target_feat'])
[]
[]
[]
[]
offset_ss_monomer
[[[ 0. -1. -2. ... -3. -2. -1.]
  [ 1.  0. -1. ... -4. -3. -2.]
  [ 2.  1.  0. ... -5. -4. -3.]
  ...
  [ 3.  4.  5. ...  0. -1. -2.]
  [ 2.  3.  4. ...  1.  0. -1.]
  [ 1.  2.  3. ...  2.  1.  0.]]

 [[ 0. -1. -2. ... -3. -2. -1.]
  [ 1.  0. -1. ... -4. -3. -2.]
  [ 2.  1.  0. ... -5. -4. -3.]
  ...
  [ 3.  4.  5. ...  0. -1. -2.]
  [ 2.  3.  4. ...  1.  0. -1.]
  [ 1.  2.  3. ...  2.  1.  0.]]

 [[ 0. -1. -2. ... -3. -2. -1.]
  [ 1.  0. -1. ... -4. -3. -2.]
  [ 2.  1.  0. ... -5. -4. -3.]
  ...
  [ 3.  4.  5. ...  0. -1. -2.]
  [ 2.  3.  4. ...  1.  0. -1.]
  [ 1.  2.  3. ...  2.  1.  0.]]

 [[ 0. -1. -2. ... -3. -2. -1.]
  [ 1.  0. -1. ... -4. -3. -2.]
  [ 2.  1.  0. ... -5. -4. -3.]
  ...
  [ 3.  4.  5. ...  0. -1. -2.]
  [ 2.  3.  4. ...  1.  0. -1.]
  [ 1.  2.  3. ...  2.  1.  0.]]]
Traced<ShapedArray(float32[27,27])>with<DynamicJaxprTrace(level=0/1)>
2025-02-02 15:25:00,740 alphafold2_model_1_seed_000 recycle=0 pLDDT=59.8
2025-02-02 15:25:00,875 alphafold2_model_1_seed_000 recycle=1 pLDDT=63.2 tol=1.46
2025-02-02 15:25:00,975 alphafold2_model_1_seed_000 recycle=2 pLDDT=64.4 tol=0.766
2025-02-02 15:25:01,075 alphafold2_model_1_seed_000 recycle=3 pLDDT=65.7 tol=0.222
2025-02-02 15:25:01,075 alphafold2_model_1_seed_000 took 15.2s (3 recycles)
2025-02-02 15:25:01,185 alphafold2_model_2_seed_000 recycle=0 pLDDT=60.7
2025-02-02 15:25:01,285 alphafold2_model_2_seed_000 recycle=1 pLDDT=64.8 tol=1.46
2025-02-02 15:25:01,385 alphafold2_model_2_seed_000 recycle=2 pLDDT=67.8 tol=0.467
2025-02-02 15:25:01,484 alphafold2_model_2_seed_000 recycle=3 pLDDT=69 tol=0.204
2025-02-02 15:25:01,484 alphafold2_model_2_seed_000 took 0.4s (3 recycles)
2025-02-02 15:25:01,594 alphafold2_model_3_seed_000 recycle=0 pLDDT=66.3
2025-02-02 15:25:01,694 alphafold2_model_3_seed_000 recycle=1 pLDDT=62.9 tol=0.821
2025-02-02 15:25:01,794 alphafold2_model_3_seed_000 recycle=2 pLDDT=62.9 tol=0.47
2025-02-02 15:25:01,893 alphafold2_model_3_seed_000 recycle=3 pLDDT=64.1 tol=0.511
2025-02-02 15:25:01,893 alphafold2_model_3_seed_000 took 0.4s (3 recycles)
2025-02-02 15:25:02,003 alphafold2_model_4_seed_000 recycle=0 pLDDT=67.2
2025-02-02 15:25:02,103 alphafold2_model_4_seed_000 recycle=1 pLDDT=65.9 tol=0.874
2025-02-02 15:25:02,203 alphafold2_model_4_seed_000 recycle=2 pLDDT=65.9 tol=0.257
2025-02-02 15:25:02,302 alphafold2_model_4_seed_000 recycle=3 pLDDT=66.3 tol=0.132
2025-02-02 15:25:02,302 alphafold2_model_4_seed_000 took 0.4s (3 recycles)
2025-02-02 15:25:02,413 alphafold2_model_5_seed_000 recycle=0 pLDDT=68.2
2025-02-02 15:25:02,512 alphafold2_model_5_seed_000 recycle=1 pLDDT=62.5 tol=0.764
2025-02-02 15:25:02,612 alphafold2_model_5_seed_000 recycle=2 pLDDT=61.7 tol=0.353
2025-02-02 15:25:02,711 alphafold2_model_5_seed_000 recycle=3 pLDDT=62.2 tol=0.223
2025-02-02 15:25:02,712 alphafold2_model_5_seed_000 took 0.4s (3 recycles)
2025-02-02 15:25:02,721 reranking models by 'plddt' metric
2025-02-02 15:25:07,271 Relaxation took 4.5s
2025-02-02 15:25:07,271 rank_001_alphafold2_model_2_seed_000 pLDDT=69
2025-02-02 15:25:09,935 Relaxation took 2.7s
2025-02-02 15:25:09,935 rank_002_alphafold2_model_4_seed_000 pLDDT=66.3
2025-02-02 15:25:12,796 Relaxation took 2.9s
2025-02-02 15:25:12,796 rank_003_alphafold2_model_1_seed_000 pLDDT=65.7
2025-02-02 15:25:15,400 Relaxation took 2.6s
2025-02-02 15:25:15,400 rank_004_alphafold2_model_3_seed_000 pLDDT=64.1
2025-02-02 15:25:18,168 Relaxation took 2.8s
2025-02-02 15:25:18,168 rank_005_alphafold2_model_5_seed_000 pLDDT=62.2
dict_keys(['rank', 'metric', 'result_files'])
2025-02-02 15:25:18,629 Done

But when I compare the top-rank result "7M3U_1_relaxed_rank_001_alphafold2_model_2_seed_000.pdb" to the experimental 7m3u.pdb, it does not match as shown above or in the paper, and only has plddt of 0.69.

If I run with --templates it matches very well and has plddt of 0.92, of course since it uses 7m3u.pdb as a template, but I assume the result reported in the paper is without templates

I've tried both by first installing the current (2024) colabfold and replacing files, and also by installing all packages as they existed in May 2023 and replacing the files.

Any help much appreciated.

[ahaldane]

I just took another look at this. I believe the command-line to reproduce the results for 7m3u, if no templates or MSA are used, should be more like:

colabfold_batch --msa-mode single_sequence --data=$HOME/alphafold/colabfold_data  --model-type alphafold2 7m3u.fa out --disulfide-bond-pairs 2 7 12 23

Is that close to the command used in the paper?

Then I still don't get the result shown in the paper. I get a cyclic peptide with all four cysteines nearby in space with disulfide bonds between the wrong pairs of cysteines. I checked that the offset_ss matrix looks right (it sets a distance of 1 between residue 2 and 7 in 0-based indexing, and residues 12 and 23, which have disulfide bonds in experiment). I tried various combos, eg using alphafold2_ptm, alternate indexing scheme, etc. What might I be be missing?

Thanks

[alejandromontesa-unizar]

I have copied the /alphafold /colabfold and /utils folders to the path that is specified in this and other issues. I also did the change in some lines that implied the simkt package as was also suggested (I am not sure if I did this good). When I run the colabfold_batch for this .fa file without the --amber flag:

fasta
SEAKYPLGPGPG:
GLGKDHEILRRRIENGAKELWFFLQSELKKLKNLEGNELQRHADEFLLDLGHHERSIMTDLYYLSQTDGAGDWREKEAKDLTELVQRRITYLQNPKDCSKAKKLVCNINKGCGYGCQLHHVVYCFMIAYGTQRTLILESQNWRYATGGWETVFRPVSETCTDRSGISTGHWSGEVKDKNVQVVELPIVDSLHPRPPYLPLAVPEDLADRLVRVHGDPAVWWVSQFVKYLIRPQPWLEKEIEEATKKLGFKHPVIGVHVRRTDKVGTEAAFHPIEEYMVHVEEHFQLLARRMQVDKKRVYLATDDPSLLKEAKTKYPNYEFISDNSISWSAGLHNRYTENSLRGVILDIHFLSQADFLVCTFSSQVCRVAYEIMQTLHPDASANFHSLDDIYYFGGQNAHNQIAIYAHQPRTADEIPMEPGDIIGVAGNHWDGYSKGVNRKLGRTGLYPSYKVREKIETVKYPTYPEAE

I obtain linear peptides instead of cyclic ones. When I use the --amber flag I obtain the following error:

File "/media/SSD2/localcolabfold/localcolabfold/colabfold-conda/bin/colabfold_batch", line 7, in <module> sys.exit(main()) File "/media/SSD2/localcolabfold/localcolabfold/colabfold-conda/lib/python3.10/site-packages/colabfold/batch.py", line 2044, in main run( File "/media/SSD2/localcolabfold/localcolabfold/colabfold-conda/lib/python3.10/site-packages/colabfold/batch.py", line 1716, in run results = predict_structure( File "/media/SSD2/localcolabfold/localcolabfold/colabfold-conda/lib/python3.10/site-packages/colabfold/batch.py", line 584, in predict_structure pdb_lines = relax_me(pdb_lines=unrelaxed_pdb_lines[key], use_gpu=use_gpu_relax) File "/media/SSD2/localcolabfold/localcolabfold/colabfold-conda/lib/python3.10/site-packages/colabfold/batch.py", line 359, in relax_me relaxed_pdb_lines, _, _ = amber_relaxer.process(prot=pdb_obj) File "/media/SSD2/localcolabfold/localcolabfold/colabfold-conda/lib/python3.10/site-packages/alphafold/relax/relax.py", line 62, in process out = amber_minimize.run_pipeline( File "/media/SSD2/localcolabfold/localcolabfold/colabfold-conda/lib/python3.10/site-packages/alphafold/relax/amber_minimize.py", line 476, in run_pipeline ret = _run_one_iteration( File "/media/SSD2/localcolabfold/localcolabfold/colabfold-conda/lib/python3.10/site-packages/alphafold/relax/amber_minimize.py", line 420, in _run_one_iteration raise ValueError(f"Minimization failed after {max_attempts} attempts.") ValueError: Minimization failed after 100 attempts.

I don't know how to proceed,I have tried to run relax with GPU and it didn't work neither