Further Response to Reviewer Y45d
**Comment 5**: In the evaluation using Vina dock metrics, it remains unclear which protein structures were utilized for benchmarking. Could you specify whether the benchmarks employed the initial holo structures or those updated by FlexSBDD?
**Response 5**: For the Vina scores of FlexSBDD, we use the updated protein structure. As for the other baselines, we follow previous works to use the target structure from the test set for evaluation.
**Comment 6**: Could you elaborate on the statement made in lines 216-217? “We note that it is fair to compare FlexSBDD with other baseline methods as the additional 217 apo structures contain no ligand molecules and cannot be used by baselines for training.” I do not fully understand the points here.
**Response 6**: Thanks for the detailed question! In FlexSBDD, we associate holo structures from training dataset (CrossDocked and Binding MOAD) with apo conformations from Apobind [3] to create apo-holo pairs for training. We want to note that the additional Apobind dataset does not contain protein-ligand structures (i.e., only protein structures) and cannot be used by baseline methods for training. The Apobind dataset employed by FlexSBDD will not bring data leakage or additional advantage. Therefore, it is fair to compare FlexSBDD with baseline methods. We will make the statement clearer in our revised paper.
**Comment 7**: The code provided with the submission appears to lack both training and inference components, which undermines the credibility of the reported results. Could you address this omission?
**Response 7**: Thanks for the valuable comment! We have uploaded the training and inference codes. We will open-source all the codes upon paper acceptance.
**Comment 8**: Regarding the prediction of side-chain conformations, it appears that the analysis is limited to the mean squared error (MSE) of chi angles without considering the orientation within the residue frame prediction. Could you discuss the rationale behind this methodological choice?
**Response 8**: Thanks for the detailed comment! We use the mean squared error (MSE) of chi angles to evaluate the prediction of sidechain conformations following previous works [9-11]. Generally, lower MSE indicate more precise sidechain structure prediction. In table 6 of the paper, we can observe that FlexSBDD achieves better performance in generating valid sidechain structures.
During rebuttal, we perform additional analysis of the side chain prediction. For example, we follow DynamicBind to conduct a comprehensive analysis six distinct conformational changes across the picosecond level to millisecond level (molecular dynamics), each exemplified by a case from PDBbind. In the following table, we report Δpocket RMSD (including side chain and backbone) of DynamicBind and FlexSBDD, which measures the relative decrease in pocket RMSD (crystal structure as reference) compared with the AlphaFold structures. **A negative Δpocket RMSD indicates that the predicted aligns more closely with the crystal structure compared with the AlphaFold prediction.** We observe that FlexSBDD achieves competitive performance that improves the AlphaFold prediction to have lower pocket RMSD, even though it is not specifically designed for dynamic docking.
| Methods | 6QGF | 6PGO | 6N8X | 6UWV | 6ROT | 6S9X |
|------------|--------------------|------|------------------------|------|---------------|------|
| DynamicBind | -0.669 | -1.140| -2.297 | -0.465| -2.327 | -5.245 |
| FlexSBDD| -0.680 | -0.976| -1.159 | -0.504| -1.148 | -3.083 |
[9] Zhang Y, Zhang Z, Zhong B, et al. Diffpack: A torsional diffusion model for autoregressive protein side-chain packing. Advances in Neural Information Processing Systems, 2023.
[10] McPartlon M, Xu J. An end-to-end deep learning method for protein side-chain packing and inverse folding[J]. Proceedings of the National Academy of Sciences, 2023, 120(23): e2216438120.
[11] Dong T, Yang Z, Zhou J, et al. Equivariant flexible modeling of the protein–ligand binding pose with geometric deep learning[J]. Journal of Chemical Theory and Computation, 2023, 19(22): 8446-8459.