acceptodds
Under review as a conference paper at ICLR 2027

NanoIF: Interaction-Aware Nanobody Inverse Folding

Abstract

Nanobody inverse folding is limited by scarce VHH structures and the underuse of supervision provided by VHH–antigen complexes. We introduce NanoIF, an interaction-aware model for nanobody inverse folding that learns from VHH–antigen complexes during training and requires only the VHH backbone at inference. It first adapts to the VHH sequence–structure distribution and then incorporates antigen information through a dedicated antigen encoder, layer-wise cross-attention, and complex-derived geometric and contrastive supervision. On a SAbDab-nano-derived benchmark, NanoIF reaches state-of-the-art amino-acid recovery, improving paratope recovery by 3.2 points over the strongest baseline, with larger gains on CDR1 and CDR2 (+9.7 and +8.6 points). Controlled training shows that matched VHH-only continued training does not reproduce these gains, while the learned improvement is retained at antigen-free inference. These results establish complex-guided training as an effective strategy for fixed-backbone nanobody sequence design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
inverse folding, nanobody, protein language models, antibody–antigen complexes

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_cgGPOpenReview ↗
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
  1. 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.
  2. 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?
  3. Do you not think that substantially more than a few thousand training steps would have further improved the model?
3 comments
AuthorsOpenReview ↗

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.

ModelFine-tuning configurationGlobalCDR1CDR2CDR3
NanoIF1,500 steps, BS≈6081.3167.6555.5439.41
Official AntiFold → VHH1,500 steps, BS6479.6863.0847.9533.06
Official AbMPNN → VHH1,500 steps, BS6475.4361.3446.2035.32
ESM-IF1 + AntiFold-style recipe1,500 steps, BS6477.7361.0754.4633.39
ESM-IF1 + our improved recipe10 epochs79.6861.5152.0938.55

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.

AuthorsOpenReview ↗

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:

StepsFolding TM-score ↑Cα RMSD (Å) ↓CDR3 RMSD (Å) ↓
5000.9052.676.65
1,0000.9082.535.95
2,0000.9042.676.59
3,0000.8762.896.72
5,0000.8462.966.23

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.

StepsGlobalCDR1CDR2CDR3
1,50081.3167.6555.5439.41
3,00081.2865.9955.4739.65

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 (λNCE=0.1\lambda_{\mathrm{NCE}}=0.1), 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.

Reviewer_5XQiOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

NanoIF is an interaction-aware nanobody inverse-folding model that uses VHH–antigen complexes during training but requires only the VHH backbone at inference. It adapts DPLM-2 to VHH data, then adds antigen cross-attention, geometric epitope supervision, and antigen–CDR contrastive alignment.

Its main contribution is showing that complex-derived interaction information improves fixed-backbone nanobody sequence design beyond simply training longer on isolated VHH structures.

On the 43-complex nano43 benchmark, NanoIF achieves 81.3% global and 47.1% paratope recovery, with particularly strong gains on CDR1/2, and the best refolding RMSDs among compared methods. Ablations and matched VHH-only training support that the gains come from complex-derived supervision, though validation is limited to a small benchmark and in-silico refolding.

Critical strengths
  1. Strong empirical performance on the core task. NanoIF achieves the best amino-acid recovery across all reported regions, including 81.3% global AAR and 47.1% paratope AAR, with particularly large gains on CDR1 and CDR2. It also achieves the best refolding RMSDs among the compared methods. This is consequential because the improvements are both quantitatively meaningful and concentrated in functionally important regions.

  2. Convincing controlled evidence for the value of complex-derived supervision. The matched VHH-only continuation fails to reproduce the Stage-2 gains, and removing cross-attention, epitope loss, or contrastive alignment degrades performance. This strengthens the paper substantially because it supports the central claim that VHH–antigen complexes provide useful interaction-specific training signal rather than merely additional data or optimization.

Critical weaknesses
  1. Limited evaluation scale and lack of experimental validation. The main benchmark contains only 43 complexes, and structural quality is assessed using a single in-silico refolding pipeline rather than experimental measurements of binding, stability, or expression. This limits confidence that the reported gains will generalize to broader nanobody design settings.

  2. The mechanistic attribution remains somewhat incomplete. Although ablations show that cross-attention, epitope supervision, and contrastive alignment each contribute, the shuffled-antigen experiment retains most of the CDR3 gain, suggesting that some improvement may come from generic complex-context regularization rather than truly pair-specific interaction learning. This weakens the strongest interpretation of the paper’s central interaction-aware claim.

Questions for authors
  1. The shuffled-antigen experiment retains most of the CDR3 improvement. Can the authors clarify what evidence supports attributing the Stage-2 gains specifically to pair-specific interaction information, rather than to a more generic regularization effect from exposure to complex context?

  2. How sensitive are the main conclusions to the small nano43 test set? In particular, can the authors provide uncertainty estimates or per-target variability for the reported AAR improvements to show that the gains are not driven by a small subset of complexes?

Suggestion for future work Experimental validation of designed nanobodies, especially binding and stability measurements, would substantially strengthen the practical significance of the method, but I would view this as future work rather than a clarification required for the present claims.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 5XQi

We thank the reviewer for the encouraging assessment and constructive suggestions. We conducted additional multi-seed experiments and per-target uncertainty analyses to address both concerns.

Q1. Pair-specific information versus general complex-guided learning

We believe the strong performance under shuffled pairing reflects an important distinction between learning general complex-aware knowledge and learning antigen-specific interactions. Our hypothesis is that complex-guided training enables NanoIF to acquire transferable sequence–structure representations that capture general knowledge of antigen-binding environments, which can improve inverse folding even without precise antigen correspondence. Learning the additional constraints imposed by a specific antigen, however, is more challenging, especially with limited paired-complex data.

Importantly, correct antigen correspondence can provide additional benefits by offering more faithful information about the binding environment. In our additional three-seed experiments without auxiliary losses, correctly paired training outperforms shuffled pairing, improving Global AAR from 79.90% to 80.60%, CDR2 from 52.52% to 55.31%, and CDR3 from 34.98% to 36.19%. In contrast, random antigen embeddings achieve only 80.03% Global and 34.74% CDR3 recovery, close to VHH-only continuation (79.65% and 34.53%). These results suggest that the improvements cannot be explained simply by introducing arbitrary conditioning noise.

We further hypothesize that this behavior is related to the nature of fixed-backbone inverse folding. A VHH backbone extracted from a bound complex already encodes part of the structural constraints associated with antigen binding, while 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.

Overall, our results suggest that complex-guided training helps NanoIF learn transferable knowledge of antigen-binding environments, while correctly paired antigens provide additional benefits beyond this general knowledge.

Q2. Robustness on the small nano43 benchmark

To assess whether our improvements are driven by a small subset of complexes, we performed paired bootstrap analysis across all 43 test complexes, comparing NanoIF with nanoFOLD-ft:

MetricImprovement95% Bootstrap CI
Global+3.98 pp[+1.99, +6.32]
CDR1+9.63 pp[+4.65, +15.28]
CDR2+8.64 pp[+2.29, +15.16]
CDR3+1.02 pp[−2.73, +4.97]

NanoIF achieves higher Global AAR on 31 of 43 complexes (9 worse, 3 tied). Even after excluding the five targets with the largest improvements, the mean Global advantage remains +2.03 percentage points, indicating that the overall improvement is not driven solely by a few favorable targets.

We additionally evaluated three independent training seeds to assess training stability. NanoIF achieves 81.12 ± 0.23% Global AAR and 38.65 ± 0.94% CDR3 recovery, compared with 79.65 ± 0.03% and 34.53 ± 0.79% for matched VHH-only continuation.

Together, these analyses provide complementary evidence of robustness: the paired bootstrap supports the Global and CDR1/CDR2 improvements across test complexes, while the three-seed experiments demonstrate reproducible gains from Stage-2 training, including in CDR3.

Thank you for the response, the robustness test with bootstrapping is nice, and partly addresses my second question. The answer to question 1 is not convincing, however, and I share the same concern as Reviewer Vhtp. I retain my score

Reviewer_VhtpOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

The paper aims to improve nanobody (VHH) inverse folding by leveraging large-scale VHH sequence data and VHH–antigen complexes. Specifically, the authors first adapt DPLM-2 to VHH sequence and structure data, and then further train the model on VHH–antigen complexes. The second stage introduces antigen cross-attention, geometric epitope supervision, and antigen–CDR contrastive learning to incorporate interaction information. At inference, the model requires only the VHH backbone without antigen input. NanoIF achieves state-of-the-art sequence recovery on the nano43 benchmark, with particularly strong improvements in CDR1 and CDR2.

Critical strengths
  1. Important research question. The paper addresses an important and underexplored question: whether VHH–antigen complex information can improve nanobody inverse folding. This is particularly relevant given the limited availability of experimentally resolved VHH structures.

  2. Strong performance improvements. NanoIF achieves state-of-the-art performance on the nano43 benchmark, improving global sequence recovery by 4.0 percentage points over the strongest baseline. The improvements are especially notable for CDR1 (+9.7) and CDR2 (+8.6). The model also achieves better refolding accuracy, supporting its effectiveness for nanobody sequence design.

Critical weaknesses
  1. The strong performance of shuffled antigen pairing raises concerns about the paper's central claim. Shuffled pairing achieves identical CDR2 recovery (55.5%) and nearly identical CDR3 recovery (38.8% vs. 39.4%) compared with correct pairing. Given the high specificity of antibody–antigen interactions, why does random pairing work almost as well? This suggests that the improvements may not primarily come from learning antigen-specific interactions. The authors should clarify what actually drives these gains.

  2. DPLM-2 uses an ESM-IF-based GVP encoder to extract structure tokens, which may introduce antigen geometry and docking information through cross-chain neighbors if the full complex is processed. This raises a potential information leakage issue, especially for antigen-free inference. The authors should clarify how VHH structure tokens are generated and compare two settings: encoding the VHH together with its antigen versus encoding the isolated VHH. This comparison should cover the two training stages and evaluation.

  3. It is unclear whether the performance gains come from the proposed method or simply from additional training data. NanoIF uses approximately 9.76 million VHH sequences, while the baselines are evaluated using their original pretrained checkpoints. Although the VHH-only continuation experiment partially addresses this concern, the comparison with existing methods remains unfair in terms of training data. The authors should provide a comparison using matched training data and computational budgets.

Questions for authors

Please see the Weakness.

2 comments
AuthorsOpenReview ↗

Response to Reviewer Vhtp

We thank the reviewer for recognizing the importance of our research question and the improvements in nanobody inverse folding. We conducted additional experiments and audited the structure-tokenization pipeline to address the three concerns.

W1. Why does shuffled antigen pairing retain most of the gains?

We believe the strong performance under shuffled pairing reflects a distinction between learning general complex-aware knowledge and antigen-specific interactions. Our hypothesis is that complex-guided training helps NanoIF transferable knowledge of sequence–structure relationships in antigen-binding environments, which can improve inverse folding without precise antigen correspondence. Learning the additional sequence constraints imposed by a particular antigen is more challenging with limited paired-complex data.

Importantly, this does not mean that correct antigen correspondence is uninformative. Correctly paired antigens provide a more faithful representation of the binding environment, potentially offering additional constraints beyond the general knowledge learned from complex-guided training. In our additional three-seed experiments without auxiliary losses, correctly paired training outperforms shuffled pairing, improving Global AAR from 79.90% to 80.60%, CDR2 from 52.52% to 55.31%, and CDR3 from 34.98% to 36.19%. In contrast, replacing antigen representations with random embeddings yields only 80.03% Global and 34.74% CDR3 recovery, close to VHH-only continuation (79.65% and 34.53%). This suggests that the benefits are not simply caused by introducing arbitrary conditioning noise.

We further hypothesize that this behavior is related to the nature of fixed-backbone inverse folding. A VHH backbone extracted from a bound complex already encodes part of the structural constraints associated with antigen binding, while 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.

W2. Potential antigen information leakage through the GVP structure tokenizer

We confirm that all VHH structure tokens in Stage 1b, Stage 2, and evaluation are generated from isolated VHH chains; Stage 1a uses sequence-only training. We verified this on 20 complexes (10 training and 10 nano43) by comparing isolated-chain, full-complex, and stored VHH tokens. Across 2,432 residues, full-complex tokens agreed with isolated-chain tokens at only 2.55%, while the stored training and evaluation tokens matched isolated-chain tokens at 100%. Thus, cross-chain antigen information does not leak through the GVP tokenizer in our implementation.

W3. Fair comparison of training data and computational budgets

We fine-tuned the officially released AntiFold and AbMPNN checkpoints on the MFDesign VHH training subset, approximately matching NanoIF Stage-2 fine-tuning exposure (1,500 optimizer updates; effective batch size 64 versus approximately 60 for NanoIF). Since AntiFold itself is fine-tuned from ESM-IF1, we also evaluated ESM-IF1-initialized baselines using both an AntiFold-style training recipe and an improved recipe to provide stronger comparisons.

ModelFine-tuning configurationGlobalCDR1CDR2CDR3
NanoIF1,500 steps, BS≈6081.3167.6555.5439.41
Official AntiFold → VHH1,500 steps, BS6479.6863.0847.9533.06
Official AbMPNN → VHH1,500 steps, BS6475.4361.3446.2035.32
ESM-IF1 + AntiFold-style recipe1,500 steps, BS6477.7361.0754.4633.39
ESM-IF1 + our improved recipe10 epochs79.6861.5152.0938.55

VHH-specific fine-tuning improves the official AntiFold baseline from 71.91% to 79.68% Global AAR, and AbMPNN from 66.13% to 75.43%. The improved ESM-IF1 configuration reaches 79.68% Global and 38.55% CDR3 recovery, but remains below NanoIF (81.31% and 39.41%).

We also clarify that 9.76M VHH sequences constitute a sampling pool, not the number actually used for training. Stage 1a and Stage 1b involved approximately 65,985 and 49,086 sequence exposures, respectively.

These results demonstrate that VHH-specific fine-tuning substantially improves existing methods but does not eliminate NanoIF's advantage.

Thank you for the response. It remains unclear why incorrectly paired antigens, which do not represent the VHH’s actual binding environment, improve performance. The explanation of “general complex-aware knowledge” does not resolve this concern. I am therefore inclined to maintain my current score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.