acceptodds
Under review as a conference paper at ICLR 2027

AdaProteina: An Adapter Is All You Need for Developable Antibody Generation

Abstract

Modern flow-matching generators for antibody design sample from an unconditional prior and offer no control over the physicochemical properties (binding, humanness, aggregation) that determine therapeutic viability. Retraining the 160M-parameter denoiser under a conditional objective is expensive and must be repeated per property specification. Classifier-free guidance trades sample diversity for conditioning strength and doubles per-step inference cost. We propose AdaProteina, a lightweight noise-space adapter that leaves the denoiser and VAE decoder of La-Proteina entirely frozen and instead learns a refiner mapping base Gaussian noise and a class label to a developability-aware initial latent, with learnable parameters and no change in per-sample inference cost. Training combines cross-entropy against a frozen developability classifier with a per-class MMD loss that anchors refined latents to the data manifold and prevents the mode collapse, e.g., tyrosine/tryptophan-dominated sequences induced by naïve property optimization. To compose multiple developability properties without training a combinatorial number of joint classifiers, we introduce a Kernel Perron–Frobenius transfer operator that maps between per-property noise spaces in closed form, enabling training-free generation of, e.g., human-like low non-specific binders. AdaProteina improves substantially over La-Proteina on sequence diversity, developability separability, and every structural-validity metric evaluated and augmenting downstream training. To the best of our knowledge, this is the first large scale exploration of noise-space refinement for developability-aware antibody generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
diffusion, antibody sequence generation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_3nHzOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes AdaProteina, a noise-space adapter for property-conditioned antibody generation with a fixed backbone. The method freezes the La-Proteina denoiser and decoder and learns an adapter using a property-prediction loss and class-conditional MMD regularization. A kernel Perron-Frobenius (kPF) operator combines separately learned property spaces. Experiments examine non-specific binding, humanness, germline conditioning, structural quality, and downstream data augmentation.

Critical strengths
  1. Conditioning through an initial-noise adapter while freezing the base generator is a practical approach that avoids updating the full generative model.
  2. The ablations illustrate how property optimization can produce repetitive sequences and how MMD regularization mitigates this failure mode.
Critical weaknesses
  1. The baseline comparisons do not sufficiently establish the method's advantage. The principal conditional baseline shifts Gaussian noise from mean 0 to mean3 according to the class label, without establishing this as an effective conditioning strategy. Comparisons against generation followed by predictor-based filtering or direct noise optimization are missing. How does AdaProteina compare with these alternatives under matched computational budgets? Such comparisons are particularly important given the claimed efficiency advantage.
  2. Predictor-based improvements are not sufficiently connected to independently validated properties. Binding scores are computed using the same predictor optimized during training (Appendix A.1.1), leaving open whether gains reflect actual property improvement or predictor-specific biases. Downstream classification and EquiFold geometry provide complementary evidence but do not directly validate the generated antibodies' non-specific binding. Can the authors provide evaluation using an independent property predictor or other indepenedent measurements? For kPF composition, please also clarify why transferring between property spaces should preserve the first property and provide joint success rates demonstrating that both conditions are retained.
Questions for authors

Please see the questions included in Critical Weaknesses 1 and 2.

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

This paper proposes a new antibody design method, AdaProteina, based on La-Proteina. Inspired by Noise-Refine, the paper introduces a new adapter to inject the condition for decoding. Specifically, the method employs a new combined loss function for this task. Experiments demonstrate the effectiveness of the proposed method.

Critical strengths
  • This paper addresses an interesting question, with strong insight.
  • The proposed method is practical.
  • The experiments and the discovery are interesting.
Critical weaknesses
  • The contribution is generally incremental.
  • The writing in the methodology and introduction might be improved.
  • In the experiments, current results demonstrate that AdaProteina cannot consistently perform better than both baselines and its ablations.

Minor issues: \citet --> \citep.

Note: I am not an expert in this field.

Questions for authors

Please refer to the weaknesses.

Reviewer_ajNbOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

AdaProteina conditions a frozen La-Proteina generator by learning a residual adapter on its initial Gaussian latent noise. Given a fixed antibody backbone, the adapter uses a developability label to influence the generated sequence and side chains. Training combines a frozen property predictor's cross-entropy with class-conditional MMD against real terminal latents. Separate adapters target nonspecific binding and humanness, and a kernel Perron–Frobenius construction is proposed to combine them without joint training. Experiments examine sequence diversity, humanness, downstream prediction, EquiFold geometry, scaffold-label dependence, and six-way heavy-chain germline conditioning. The main practical contribution is adapting an existing generator through its input noise rather than updating its denoiser.

Critical strengths

A useful adaptation mechanism supported by informative collapse ablations. Freezing both the denoiser and decoder while training an initial-noise adapter is a practically appealing way to reuse a large generator. The comparison in Table 1 makes a concrete failure visible. Property loss alone produces repetitive aromatic sequences, with class-1 heavy-chain edit distance 28.4 and FD 61.23; adding MMD raises the distance to 98.3 and reduces FD to 10.22. Table 4 also shows FD improving substantially between MMD weights 0 and 0.25, with smaller changes thereafter. Although the metric definitions need clarification, the reported pattern is useful evidence that distributional anchoring matters when optimizing a frozen property model.

Critical weaknesses
  1. The evidence does not sufficiently separate biological property control from optimization of a surrogate. Appendix A.1.1 explicitly uses the training predictor h again to calculate binding scores, including the 0.704 reported for kPF in Table 5. This measures agreement with the optimization target. More directly, Section 3.3 reports that requesting the opposite binding class from a scaffold succeeds no better than chance, while the proposed label-swapping fix is illustrated through hydropathy rather than quantitative conditional success. The independent ESM2 experiment is valuable, but Figure 3 reports only 0.56 AUC on 96 Emibetuzumab variants, despite recall 0.94 and F1 0.69. High threshold-dependent scores with near-chance ranking do not establish robust discrimination. Table 2 improves AUC to 62.5 and 57.8, but lacks sufficient split, augmentation-size, selection, and uncertainty details to assess generalization. Furthermore, the main “La-Proteina” baseline assigns class 1 the arbitrary prior N(3, I), so it does not establish an advantage over competitive conditioning methods. An independent property evaluation with explicit scaffold/sequence-separated splits and a meaningful conditioning baseline would substantially change my assessment.

  2. The kPF composition mechanism is insufficiently specified and its joint-property claim is not demonstrated. In Algorithm 1, the reference Gram matrix uses gφ(zᵢ⁰; y), whereas the query kernel compares z* with raw zᵢ⁰. The algorithm never defines the Zᵢ used in its final reconstruction or how the humanness adapter supplies them. The correspondence between samples in the two transfer stages, selection of the human-like mode, and kernel/sample-size settings are also unclear. Mapping one marginal distribution to another does not by itself establish preservation of the original binding condition. Table 5 gives average Sapiens scores and the reused binding predictor, omits binding for direct composition, and provides neither joint success rates nor results for both requested binding labels. The central example of human-like low-nonspecific binders therefore remains unverified. The scalability and efficiency claims also need qualification. A joint multi-output predictor need not require one classifier per property combination, and no adapter parameter count, elapsed training cost, or end-to-end kPF sampling cost is reported.

Questions for authors
  1. What are the exact training and evaluation splits for h, the adapters, and the downstream classifiers, including scaffold/sequence overlap and augmentation selection? Are there independent property-control results, especially for label-swapped scaffolds, that can explain the low Figure 3 AUC and distinguish predictor exploitation from generalization?
  2. Can you provide a consistent definition of both kPF stages, including the query reference points, Zᵢ, sample correspondence, and human-mode selection, and report the fraction of samples satisfying both properties for each requested binding class using an independent evaluator?

Comments on the paper

1 comment
Area_Chair_Sy6SOpenReview ↗

Reviewers please respond

Hi reviewers,

Thanks for your work so far on this paper. Please respond to the authors, even if to acknowledge their response.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.