- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper addresses nanobody inverse folding. The authors take DPLM-2, a multimodal discrete diffusion model for proteins, and finetune it on nanobody data. The most substantial part of the work is Stage 2 of the training process, where the model is conditioned on antigen context: a cross-attention module between the VHH and antigen representations, plus several loss functions intended to enforce that the model respects the epitope and that the CDR loops are specific to the antigen. The model is evaluated on a small nanobody benchmark against established reference models, using sequence recovery and refolding with NanoBodyBuilder2. It outperforms the baselines on both metrics and gains substantially from Stage 2 training. However, supplying the antigen during validation makes the model worse, so antigen conditioning does not appear to be necessary at inference.
Critical strengths
I consider the results of this work interesting and believe they are worth being published. My disagreement is with the framing, and with analyses that need more careful interpretation. I would rather see a paper that investigates the effect of antigen conditioning showing when and why it fails than just a paper about a better nanobody inverse folder.
- Leveraging both nanobody sequence and structure data is a good idea.
- Enforcing compatibility with the antigen is a good idea.
- Recovery rates and refolding experiments are the standard and most important evaluations, and both are reported.
- The most relevant references are mentioned and compared against: ProteinMPNN, ESM-IF, AbMPNN, AntiFold, nanoFOLD.
- The reported results are good and better than the compared models, although the test set is small — hard to avoid in this field.
- Considerable effort has gone into antigen conditioning: several new losses and a two-stage training scheme are proposed.
Critical weaknesses
Framing (most important). The paper claims antigen conditioning makes the model better, which the results do not support. The random-pairing experiment shows that correct antigen context adds almost nothing (38.8 vs. 39.4 CDR3 AAR, with no reported seed variance), and the gain looks like a regularization effect from including complex data as a training signal rather than conditioning information. Worse, supplying the antigen at inference does not merely fail to help, it hurts, by 1.1 points Global AAR (81.3% antigen-free vs. 80.2% with antigen, Table A.1). A model built around epitope conditioning that degrades when conditioned has a mechanism failure, not a neutral design property, yet §3.4 presents the antigen-free bypass as a convenient option rather than a necessity. The ablations are substantial but do not explain where the improvement comes from, and no concrete mechanism is offered.
Insufficient information on the finetuning runs. Training stops after 1,000 steps with no justification, and without the batch size it is impossible to know how many sequences the model has seen. 1,000 steps also seems small. This is suspicious as reported.
Small errors in the related-work description. As far as I am aware, AntiFold and AbMPNN are trained on the same data, which is stated differently here. Though this is a minor issue.
Methods structure. It is hard to tell what is already provided by DPLM-2, the base model. Model and backbone adaptation is §2.2, where DPLM-2 is introduced, but §2.1 already discusses training the model. The intent to define the problem generally, then the implementation, is reasonable, but §2.1 is too specific in places (Two-stage training etc.) for that split to work.
Problems with the losses. The antigen–CDR contrastive loss is not epitope-specific. It could also encourage a model that already knows the antigen to encode antigen specificity by picking an arbitrary residue in the CDR loop, rather than designing a compatible loop. Finding a balance where this loss genuinely drives compatible CDR design will be difficult. On top of that, the cross-entropy loss will probably overpower the contrastive term in most cases, since CE already pushes toward a one-to-one mapping between cdr loops and epitopde. The idea is nice but the loss needs more careful investigation.
Unexplained design choices. Why DPLM-2 as the base model? Why compare all models without antigen context when with-antigen is actually better for most of the baselines, even if only slightly? These are examples rather than an exhaustive list.
Questions for authors
- Is your model not better than the reference models simply because it is trained only on nanobody structures, and additionally on 9.76M nanobody sequences that no other inverse folder here uses? Without for example AbMPNN or AntiFold baseline retrained on the same nanobody subset, the complex-supervision claim cannot be separated from nanobody specialization, regardless of what the ablations show.
- It has been shown previously that inverse folders for antibodies and nanobodies condition poorly on the target, but you do not treat this as a problem. Why do you believe antigen context is irrelevant at validation? What could the model have learned that makes seeing the antigen unnecessary at validation time?
- Do you not think that substantially more than a few thousand training steps would have further improved the model?
3 comments
Response to Reviewer cgGP (Part 2/2: Q3 and other comments)
Q3. Would longer training further improve NanoIF?
We conducted additional experiments to examine the effect of training duration in both stages.
Stage 1: VHH adaptation. We selected the Stage-1a checkpoint based on folding performance on an independent validation set of 65 VHH chains. We evaluated checkpoints from 500 to 5,000 steps at 500-step intervals, with representative results shown below:
| Steps | Folding TM-score ↑ | Cα RMSD (Å) ↓ | CDR3 RMSD (Å) ↓ |
|---|---|---|---|
| 500 | 0.905 | 2.67 | 6.65 |
| 1,000 | 0.908 | 2.53 | 5.95 |
| 2,000 | 0.904 | 2.67 | 6.59 |
| 3,000 | 0.876 | 2.89 | 6.72 |
| 5,000 | 0.846 | 2.96 | 6.23 |
The 1,000-step checkpoint achieves the best validation performance across these structural metrics, while longer training generally degrades folding accuracy. This supports our checkpoint selection rather than suggesting that Stage 1 was stopped arbitrarily. Stage 1b subsequently performs another 1,000 updates for sequence–structure adaptation.
Stage 2: Complex-guided training. We further extended Stage-2 training from 1,500 to 3,000 optimizer updates, continuing from the original checkpoint under the same learning-rate schedule.
| Steps | Global | CDR1 | CDR2 | CDR3 |
|---|---|---|---|---|
| 1,500 | 81.31 | 67.65 | 55.54 | 39.41 |
| 3,000 | 81.28 | 65.99 | 55.47 | 39.65 |
Despite doubling the training steps, Global AAR remains nearly unchanged (81.31% vs. 81.28%), with only a marginal CDR3 improvement (+0.24 percentage points).
We hypothesize that the limited Stage-2 training data (1,231 VHH–antigen complexes) may constrain the benefits of prolonged optimization, potentially leading to diminishing returns or overfitting. Although validation performance improved, this did not translate into better overall test recovery. These single-seed results suggest that simply extending training under the current configuration offers limited additional benefit, while larger paired datasets or alternative training schedules may provide further improvements.
Other comments for weaknesses
Methods structure. We thank the reviewer for this suggestion and will reorganize the Methods section to more clearly distinguish the pretrained DPLM-2 components from our proposed adaptations and training procedures.
Problems with the losses. We agree that our antigen–CDR InfoNCE loss encourages representation-level alignment rather than directly enforcing epitope-specific biochemical compatibility. Its purpose is to complement the diffusion objective by encouraging the CDR representations to retain antigen-associated information.
To investigate whether InfoNCE is overwhelmed by the diffusion loss, we additionally analyzed their gradients on five training minibatches (40 samples) at the Stage-2 checkpoint. After applying the training loss weight (), the mean gradient-norm ratios of InfoNCE to diffusion were 0.958 for shared LoRA and 0.940 for cross-attention, with mean cosine similarities of 0.061 and 0.029, respectively. These results indicate that InfoNCE provides a gradient signal comparable in magnitude to the main objective, with a largely distinct direction on the sampled minibatches.
Consistently, our three-seed ablation shows that removing InfoNCE reduces Global AAR from 81.12% to 80.61% and CDR3 recovery from 38.65% to 37.83%. Together, these results support a meaningful training contribution from InfoNCE, although they do not establish antigen-specific biochemical compatibility.
Unexplained design choices. We selected DPLM-2 as a recent and competitive multimodal protein foundation model with pretrained sequence–structure modeling capabilities. More importantly, its discrete diffusion framework directly supports structure-conditioned sequence generation, making it a natural starting point for our inverse-folding task. Its unified sequence–structure representation also allows us to progressively adapt the same backbone to VHH sequences and structures, and subsequently incorporate antigen information through cross-attention during complex-guided training. These architectural properties make DPLM-2 particularly suitable for our proposed training framework.
Regarding baseline comparisons, we adopted a common backbone-only inference setting to evaluate sequence recovery without requiring antigen input. We also report antigen-conditioned results in Table A.1, where several baselines benefit from antigen information. Importantly, even with antigen input, these baselines remain below NanoIF's antigen-free performance in Global AAR. Thus, NanoIF's advantage is not simply a consequence of evaluating competing methods without antigen context.
I thank the authors for the careful rebuttal including reasonable changes to the model and additional results. However, it leave me somewhat unsatisfied, I think the main concern is not addressed well, and for that reason I will maintain my score. The authors originally argued their model performs well because of antigen-aware training, but then they show that random antigen context during training has nearly the same effect as correct antigen context, and that no antigen context at inference is the best setting of all. Now, in their rebuttal, they argue that an important distinction lies between learning general complex-aware knowledge and learning antigen-specific interactions, and that their model mainly learns the first. But in my opinion, they are not able to clearly explain what that general complex-aware knowledge would be, or why their possible explanation would be valid.
Response to Reviewer cgGP (Part 1/2: Q1–Q2)
We thank the reviewer for recognizing the value of our results and for the constructive suggestions. We agree that the distinction between complex-guided training and effective antigen-specific conditioning deserves greater emphasis. We conducted additional controlled experiments and will clarify this distinction in the revision.
Q1. Are the improvements mainly due to VHH-specific training?
To disentangle the effects of VHH-specific training from our proposed method, we conducted additional fine-tuning experiments using the officially released AntiFold and AbMPNN checkpoints on the MFDesign VHH training subset, with approximately matched additional training budgets (1,500 optimizer updates, effective batch size 64, compared with approximately 60 for NanoIF Stage 2).
We also evaluated an ESM-IF1-initialized baseline using an AntiFold-style training recipe, as well as an improved recipe developed to provide a stronger comparison.
VHH-specific fine-tuning substantially improves both officially released baselines (AntiFold: 71.91% → 79.68%; AbMPNN: 66.13% → 75.43% Global AAR). Furthermore, even our improved ESM-IF1-based configuration, which achieves 79.68% Global and 38.55% CDR3 recovery, remains below NanoIF (81.31% and 39.41%, respectively).
Also, we apologize for the lack of clarity in our description of the VHH sequence data. The 9.76M sequences constitute a large sampling pool, rather than the number of sequences actually used for training. Specifically, Stage 1a and Stage 1b involved approximately 65,985 and 49,086 sequence exposures, respectively, rather than training over the entire corpus.
Overall, these additional experiments demonstrate that VHH-specific fine-tuning alone does not eliminate NanoIF's performance advantage. We acknowledge that the 1,500-step experiments approximately match additional fine-tuning exposure rather than total pretraining compute, while the improved 10-epoch configuration is included as a stronger, non-budget-matched comparison.
Q2. Why does antigen conditioning fail at inference?
We agree that the performance degradation when providing the antigen at inference (80.16% vs. 81.31% Global AAR without antigen) reflects a limitation in NanoIF's current antigen-specific conditioning ability, rather than an advantage of omitting antigen information.
We believe an important distinction lies between learning general complex-aware knowledge and learning antigen-specific interactions. Our hypothesis is that complex-guided training primarily helps NanoIF acquire transferable sequence–structure knowledge associated with antigen-binding environments, which can improve inverse folding even without explicit antigen input. However, the model has not yet learned to reliably exploit the additional information provided by a specific antigen at inference.
One possible explanation lies in the nature of fixed-backbone inverse folding. Since the VHH backbones are extracted from bound complexes, they already encode part of the structural constraints associated with antigen binding. The remaining antigen-specific information may involve subtler sequence-level dependencies, such as interface chemical complementarity, which are difficult to learn from only 1,231 paired complexes. Interestingly, replacing the correct antigen with a mean antigen representation improves CDR3 recovery from 35.79% to 39.22%, approaching the antigen-free result of 39.41%. This observation is consistent with the hypothesis that the model benefits more from general complex-aware knowledge than from precise antigen-specific conditioning.
Overall, our results suggest that complex-guided training improves backbone-conditioned inverse folding, while effective antigen-specific inference remains an important limitation of the current model.