19 samples failed to reproduce in PDB timesplit set

[BlacksheepJane]

Hi, thank you for your great work on SurfDock!

While reproducing the evaluation on the PDB timesplit (363) dataset using eval_samples.sh, I encountered failures in docking for 19 molecular complexes.

The PDB IDs along with the failure reasons are as follows:

PDB ID	Error Message
6qlp	index 284 is out of bounds for axis 0 with size 284
6d08	index 272 is out of bounds for axis 0 with size 272
6qtr	Sizes of tensors must match except in dimension 1
6cf7	Sizes of tensors must match except in dimension 1
6jam	index 165 is out of bounds for axis 0 with size 165
6pnn	Sizes of tensors must match except in dimension 1
6oim	Sizes of tensors must match except in dimension 1
6k2n	index 257 is out of bounds for axis 0 with size 257
6qi7	Sizes of tensors must match except in dimension 1
6olx	Sizes of tensors must match except in dimension 1
6pya	Sizes of tensors must match except in dimension 1
6jag	Sizes of tensors must match except in dimension 1
6qr7	Sizes of tensors must match except in dimension 1
6cjj	Sizes of tensors must match except in dimension 1
6h14	index 168 is out of bounds for axis 0 with size 168
6i78	index 105 is out of bounds for axis 0 with size 105
6jao	index 160 is out of bounds for axis 0 with size 160
6nd3	Sizes of tensors must match except in dimension 1
6kjf	Sizes of tensors must match except in dimension 1

From the above, there appear to be two main categories of errors:

1. index XXX is out of bounds for axis 0 with size XXX
2. Sizes of tensors must match except in dimension 1.

After preliminary investigation, I found that both issues originate in the get_complex function within SurfDock/score_in_place_dataset/score_dataset.py.

• The first issue occurs at the line:

  mda_rec_model = mda.Universe(pure_pocket_path)

• The second issue occurs during the call:

  get_rec_graph(mda_rec_model, rec_coords, c_alpha_coords, n_coords, c_coords, complex_graph, rec_radius=self.receptor_radius,
              c_alpha_max_neighbors=self.c_alpha_max_neighbors, all_atoms=self.all_atoms,
              atom_radius=self.atom_radius, atom_max_neighbors=self.atom_max_neighbors, remove_hs=self.remove_hs, lm_embeddings=lm_embeddings)

May I ask if you have encountered similar issues when testing on this dataset? If so, could you share how you addressed them?

I would really appreciate any help or guidance you can provide. Thanks again for the great work!

[BlacksheepJane]

I conducted tests on the above issue and found that it was caused by the presence of two END entries in the XXX_pure.pdb file. After removing the content after the first END, the first issue was resolved: three docking cases succeeded, while the remaining ones failed, leading to what I now refer to as the second issue.

The second issue arises in the get_calpha_graph function in SurfDock/datasets/process_mols.py, where the shapes of node_feat and lm_embeddings do not match in the first dimension. For example:

2025-07-23 17:25:37.918 | INFO | datasets.process_mols:get_calpha_graph:515 - [node_feat.shape] torch.Size([13, 10])
2025-07-23 17:25:37.918 | INFO | datasets.process_mols:get_calpha_graph:516 - [lm_embeddings.shape] torch.Size([36, 1280])

This mismatch directly causes the following line to raise an error:

complex_graph['receptor'].x = torch.cat([node_feat, lm_embeddings], axis=1) if lm_embeddings is not None else node_feat

Error:

Sizes of tensors must match except in dimension 1.

Do you have any ideas on how to resolve this issue?

[CAODH]

I conducted tests on the above issue and found that it was caused by the presence of two END entries in the XXX_pure.pdb file. After removing the content after the first END, the first issue was resolved: three docking cases succeeded, while the remaining ones failed, leading to what I now refer to as the second issue.

The second issue arises in the get_calpha_graph function in SurfDock/datasets/process_mols.py, where the shapes of node_feat and lm_embeddings do not match in the first dimension. For example:

2025-07-23 17:25:37.918 | INFO | datasets.process_mols:get_calpha_graph:515 - [node_feat.shape] torch.Size([13, 10])
2025-07-23 17:25:37.918 | INFO | datasets.process_mols:get_calpha_graph:516 - [lm_embeddings.shape] torch.Size([36, 1280])
This mismatch directly causes the following line to raise an error:

complex_graph['receptor'].x = torch.cat([node_feat, lm_embeddings], axis=1) if lm_embeddings is not None else node_feat
Error:

Sizes of tensors must match except in dimension 1.
Do you have any ideas on how to resolve this issue?

I did encounter this issue before, but I can't recall the exact solution. Could you provide specific error details or logs? I'll conduct immediate tests and follow up promptly

[BlacksheepJane]

I conducted tests on the above issue and found that it was caused by the presence of two END entries in the XXX_pure.pdb file. After removing the content after the first END, the first issue was resolved: three docking cases succeeded, while the remaining ones failed, leading to what I now refer to as the second issue.
The second issue arises in the get_calpha_graph function in SurfDock/datasets/process_mols.py, where the shapes of node_feat and lm_embeddings do not match in the first dimension. For example:
2025-07-23 17:25:37.918 | INFO | datasets.process_mols:get_calpha_graph:515 - [node_feat.shape] torch.Size([13, 10])
2025-07-23 17:25:37.918 | INFO | datasets.process_mols:get_calpha_graph:516 - [lm_embeddings.shape] torch.Size([36, 1280])
This mismatch directly causes the following line to raise an error:
complex_graph['receptor'].x = torch.cat([node_feat, lm_embeddings], axis=1) if lm_embeddings is not None else node_feat
Error:
Sizes of tensors must match except in dimension 1.
Do you have any ideas on how to resolve this issue?

I did encounter this issue before, but I can't recall the exact solution. Could you provide specific error details or logs? I'll conduct immediate tests and follow up promptly

Apologies for the delayed response, I had some matters to attend to over the past few days.

The error occurs duringget_rec_graph in the get_complex function, using ligand from pdb id 6i78 as example.

Here’s a relevant excerpt from the log output:

2025-08-07 21:19:45.621 | INFO     | score_in_place_dataset.score_dataset:get_complex:171 - Processing 6i78
2025-08-07 21:19:45.695 | INFO     | datasets.process_mols:get_calpha_graph:515 - [node_feat.shape] torch.Size([13, 10])
2025-08-07 21:19:45.695 | INFO     | datasets.process_mols:get_calpha_graph:516 - [lm_embeddings.shape] torch.Size([36, 1280])
2025-08-07 21:19:45.695 | INFO     | datasets.process_mols:get_calpha_graph:517 - [num residues] <ResidueGroup [<Residue VAL, 116>, <Residue PRO, 117>, <Residue ASN, 143>, ..., <Residue ASN, 160>, <Residue PRO, 161>]>
2025-08-07 21:19:45.699 | INFO     | score_in_place_dataset.score_dataset:get_complex:197 - [6i78] get_rec_graph failed: Sizes of tensors must match except in dimension 1. Expected size 13 but got size 36 for tensor number 1 in the list.
2025-08-07 21:19:45.699 | INFO     | score_in_place_dataset.score_dataset:get_complex:204 - Skipping 6i78 because of the error:'NoneType' object has no attribute 'shape'

As you can see, I added three logging lines in get_calpha_graph for better insight:

logger.info(f"[node_feat.shape] {node_feat.shape}")
logger.info(f"[lm_embeddings.shape] {lm_embeddings.shape}")
logger.info(f"[num residues] {mdn_rec.residues}")

It seems like there is a mismatch between the number of residues used to compute node_feat (13) and the number of residues represented in the ESM embedding (lm_embeddings.shape[0] = 36), which leads to the downstream tensor mismatch and failure.

Please let me know if you need further details, or if there’s anything I can help with in terms of debugging or checking preprocessing assumptions.