PPDiff: Diffusing in Hybrid Sequence-Structure Space for Protein-Protein Complex Design

We appreciate the reviewer's dedication in providing detailed feedback. We have clarified all the reviewer’s concerns and conducted additional experiments accordingly. All ablation studies were conducted by generating one candidate and we evaluated the resulting complexes using AF3 with the same seed. Responses to specific points are provided below:

Q1: What is the co-designability score?

Ans: We evaluated three consistency metrics for the top-1candidate: (1) Seq RMSD: RMSD between the folded structure of the designed sequence and ground truth, (2) Struct RMSD: RMSD between the designed structure and ground truth, (3) Design scRMSD: RMSD between the folded structure of the generated sequence and the designed structure. Although Struct RMSD and Design scRMSD are higher than Seq RMSD, incorporating structural information significantly enhances binder sequence design. As we can see, PPDiff achieves a substantially lower Seq RMSD (1.13) than a sequence-only model (5.79, Finetuned ESM2 on PPBench), validating the importance of co-design in producing sequences that closely align with their intended structures.

Q2: It would be appropriate to run RFDiffusion with ProteinMPNN.

Ans: Please refer to the response of reviewer zYHd Q1.

Q3: Essential References Not Discussed

Ans: We appreciate the reviewer's valuable suggestions. We will cite and discuss MultiFlow, D3PM, and BindCraft in the revised version. We kindly note that our discrete sequence diffusion part is based on [1] as introduced in Sec. 3.3 line 144-148 in our manuscript, which precedes and is referenced by D3PM.

[1] Argmax flows and multinomial diffusion: Learning categorical distributions. Hoogeboom et al. NeurIPS 2021.

Q4: About dataset release

Ans: We will release all datasets shortly.

Q5: Can the authors clarify how the target protein is fed to the model as conditioning?

Ans: We concatenate the sequences of the target and binder proteins, and represent all residues in target and binder proteins jointly as a unified point set.

Q6: In Eq (7), $s_{t-1}$ should depend on $s_t$ and $x_t$, and similarly $x_{t-1}$.

Ans: We appreciate the reviewer's suggestion and agree with the correction. We will update our manuscript in the revised version.

Q7:Why does the layer need to be causal?

Ans: Without causal attention layers, sequences displayed repetitive residue types (e.g., "EEEE"), resembling multimodality issues observed in non-autoregressive MT [2]. We introduced causal attention layers for autoregressive dependency management, significantly improving performance, as shown in our ablation study below.

[2] Non-Autoregressive Neural Machine Translation. ICLR 2018.

Q8: Ho et al. suggested $\lambda_t=1$ for the noise prediction setting, not for $x_0$ prediction

Ans: We apologize for any confusion. We set $\lambda_t=1$ following previous work [3]. We will accordingly update our manuscript.

[3] 3D Equivariant Diffusion for Target-Aware Molecule Generation and Affinity Prediction. ICLR 2023.

Q9: The authors propose a more informative prior and sequence guidance, which should be supported by ablation studies.

Ans: Ablation studies, as presented below, clearly indicate that removing either the informative prior or sequence guidance decreases performance.

Q10: Why exactly do the mini-binder generation experiments use the AF2 pAE_interaction score for binder evaluation

Ans: Following established practices in previous studies [4], we utilized the AF2 pAE_interaction score to evaluate binder designs due to its proven effectiveness in distinguishing experimentally validated binders from non-binders. Previous research has shown that binder candidates selected using AF2 pAE_interaction scores yield experimental success rates ranging from 1.5% to 7% across various target proteins [4].

[4] Improving de novo protein binder design with deep learning. Nature Communication. 2023.

Q11: How important is the initialization of self-attention layers from ESM2?

Ans: We conducted an ablation study on removing ESM2 initialization below. Results indicate performance degradation upon removal, demonstrating the significance of leveraging ESM2's pretrained knowledge as an effective initialization for PPDiff.

We greatly appreciate the reviewer for the insightful and constructive feedback. We have conducted additional experiments as suggested. Detailed responses to specific questions are provided below:

Q1: To strengthen the in silico validation, could the authors provide docking scores for the designed binders, particularly for well-known targets?

Ans: Thank you for this valuable suggestion. Because our PPDiff designs provide only alpha-carbon backbones, they are not directly applicable for docking. Therefore, we first used ESMFold to generate the structures of our designed binder sequences. We then performed docking simulations using HDOCK [1], pairing each predicted binder with its corresponding target protein. For each target, we produced five binder candidates and reported the average docking score for the top-1 candidate in Table 1. As noted in the original HDOCK paper, a more negative docking score means a more possible binding model; notably, this score should not be treated as the absolute measures of binding affinity because it has not been calibrated to the experimental data. The results show that PPDiff consistently achieves more negative docking scores across all ten target proteins compared to baseline methods, indicating that binders designed by our model exhibit stronger potential affinities. Additionally, Table 2 compares these scores with ground truth docking values, demonstrating that PPDiff even outperforms experimentally confirmed binders (ground truth) in five categories — a finding that highlights its potential for designing high-affinity protein binders.

[1] The HDOCK server for integrated protein–protein docking. Yan et al. Nature Protocols. 2020.

Table1: Docking Scores Comparing with Baselines

Table2: Docking Scores Comparing with Ground Truth

We sincerely thank the reviewer for the valuable comments and insightful suggestions. We have addressed all raised concerns and conducted additional experiments as recommended. Please find detailed responses to the specific points below:

W1:Essential References Not Discussed:I think [1] should be discussed. It is a diffusion method based on framed.

Ans: We appreciate the reviewer for highlighting this relevant reference. Proteus is an innovative frame-based diffusion approach, integrating a graph-based triangle technique and a multi-track interaction network, which shows robust capabilities in designing protein backbone structures. We will include a citation and thorough discussion of this method ([1]) in the revised manuscript.

[1] Proteus: pioneering protein structure generation for enhanced designability and efficiency.

Comment: The authors did not cite framediff but framflow

Ans: Thank you for pointing out this oversight. We will appropriately cite the correct reference, Framediff ([2]), in our revised manuscript.

[2] SE(3) diffusion model with application to protein backbone generation. Yim et al. ICML 2023.

Q1: Can you finetune your method for the binder task and evaluate it against RFDiffusion? This would make the paper much stronger in my opinion.

Ans: Thank you for this valuable suggestion. We initially followed the binder design procedure outlined in the RFDiffusion paper to generate the backbone structures. Subsequently, ProteinMPNN was employed to design corresponding binder sequences based on these backbone structures. To ensure a fair comparison, we pretrained ProteinMPNN on our curated PPBench and then fine-tuned the pretrained model specifically for the downstream binder design task. As RFDiffusion does not provide a training script, we utilized their publicly available model weights directly. Evaluation of success rates was conducted consistent with the methodology described in our manuscript. The average success rates across each target protein category are presented below. Our results indicate that our model outperforms RFDiffusion + ProteinMPNN for 5 out of 10 target proteins, and achieves a higher overall average success rate across all tested target proteins. Notably, our PPDiff achieves much higher novelty and diversity scores, demonstrating our model's superior capability in designing high-affinity, novel and diverse binders.

Table: Success Rate, Novelty and Diversity on Target Protein-Mini Binder Design Task