acceptodds
Under review as a conference paper at ICLR 2027

AbGate: Length-Decoupled Antigen-Conditioned Antibody Design

Abstract

Antibodies are critical molecular therapeutics, but a new target is usually available only as a sequence, without the solved structure or mapped epitope that existing generators need to run. We propose AbGate, which conditions a frozen antibody language model on the antigen sequence alone. A learned pooler compresses the antigen into a fixed number of tokens whatever its length and a cross-attention module then injects these tokens into every decoder layer, so antigen length never changes how much of the decoder's context AbGate uses. Training only 1.09% of its parameters, AbGate performs on par with MAGE, a fully fine-tuned generator trained on the same corpus, across every structural metric we measure on a shared target. Across 502 held-out targets AbGate generates antibodies at 95.6% filter retention and 100% novelty against the training pairs, and they carry the 9-mer composition of natural human repertoires. Additionally, on an independent test set of 496 targets, antibodies generated by AbGate performed comparably to experimentally validated binding antibodies on interface confidence.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
antigen-conditioned generation, protein language models, parameter-efficient fine-tuning, cross-attention conditioning, de novo protein design, protein structure prediction

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_a794OpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2(+1)
Presentation
1
Soundness
2
Read review
Paper summary

This study presents AbGate, a parameter-efficient framework for generating antibody candidates conditioned solely on antigen sequence, without prior knowledge of the binding site or antibody–antigen interface geometry. AbGate uses a learned pooler to compress variable-length antigen sequences into 32 tokens, which are supplied to every layer of a pretrained antibody language model through cross-attention. With approximately 1.1% of the trainable parameters used by a fully fine-tuned baseline. Experimentally, AbGate achieves comparable mean RosettaFold3 ipTM and pLDDT scores on a shared target.

Critical strengths
  1. This paper addresses the design of antibody sequences directly from antigen sequences without annotated binding sites. To incorporate antigen information into the biological language model responsible for antibody design, the authors encode variable-length antigen sequences and compress their representations into a fixed number of tokens. This is an interesting and reasonable approach to conditioning antibody generation on antigen sequence information.

  2. Although the experimental section provides limited quantitative analysis, the authors include a range of statistical plots to support the strong performance of AbGate despite its relatively small number of trainable parameters.

Critical weaknesses
  1. Research scope. Designing antibodies from antigen sequences alone, without incorporating 3D structural information or binding-site annotations, represents a relatively unconstrained formulation of conditional antibody design. While scientifically relevant, the practical utility of this setting for future drug-design applications may be limited.

  2. Methodological innovation. Encoding variable-length antigen sequences into a fixed number of token representations is a reasonable and efficient way to provide antigen context to an antibody generation model. However, the generation component still relies on a pretrained foundation model, and cross-attention is a well-established mechanism for integrating conditioning features. The approach can therefore be viewed as a parameter-efficient fine-tuning (PEFT) adaptation to the certain setting, with limited methodological novelty.

  3. Experimental evaluation. The paper reports relatively few numerical performance metrics, relying more heavily on statistical visualizations to support its conclusions. Although these analyses are varied, they do not, on their own, provide sufficiently strong evidence to establish the superiority of AbGate.

Questions for authors

I suggest the authors will more clearly articulate the methodological novelty of their proposed approach and provide more comprehensive quantitative evaluations to substantiate AbGate’s claimed advantages.

3 comments

We appreciate your request for a clearer account of AbGate's contribution and stronger quantitative support. We have tried our best to revise the paper to improve the presentation accordingly and the revision explains the role of the conditioning interface and adds comparisons that assess its response to antigen input and its parameter efficiency.

Scope and practical role. Specifically, AbGate is a sequence-based only antibody design method and learns from antibody–antigen sequence pairs without the need to have resolved complex structures and generates candidates for subsequent structural screening. AbGate, compared to existing methods, is more scalable and efficient when considering conditioning on long antigen sequences for targeted antibody design with no structure information available. To clarify its practical role, the revised Discussion connects this stage to future evaluation against experimental binding data, with the aim of improving candidate selection for laboratory testing.

Methodological contribution. Besides focusing on sequence-based conditional antibody design, AbGate adopts a parameter-efficient fine-tuning (PEFT) adaptation strategy with the corresponding design of a fixed decoder interface for antigens of different lengths as well as customized cross-attention conditioning mechanism. Our hypothesis is that learned pooling can retain features relevant to antigen recognition within 32 tokens, while cross-attention lets each decoder layer retrieve features relevant to the antibody sequence being generated. The decoder receives the same number of antigen tokens for every target, without consuming autoregressive input positions.

The architectural length decoupling in AbGate is a unique feature. With the identified potential bias due to the length dependence of predictor scores in the tested baselines through our presented comparisons within each antigen and across length groups, we showed that a single trained AbGate model can produce candidates across 502 held-out antigens without separate finetuning for each antigen family, with mean confidence close to known binders across length groups and several families absent from training. This demonstrates the effectiveness of the learned interface across targets by AbGate while the pretrained decoder blocks remain frozen.

Quantitative evidence. To address your request for stronger quantitative support, we added the conditioning and architecture comparisons in Section 3.4. Below, we summarize these additions alongside the MAGE and held-out generation results already reported in Sections 3.2–3.3. Together, they show the parameter efficiency of AbGate, the quality of its generated candidates, and its response to antigen input.

  1. Similar mean confidence with fewer trainable parameters. On the shared RBD, AbGate uses 8.46M trainable parameters compared with MAGE’s 764M. Mean ipTM is 0.434 versus 0.435, and mean pLDDT is 0.787 versus 0.789.

  2. A smaller conditioning architecture with higher sequence retention. In Section 3.4, AbGate uses 8.46M trainable parameters, compared with 12.44M for mean pooling and 15.21M for K32 pooling with vector adapters. Their mean ipTM scores are nearly identical, ranging from 0.470 to 0.472. AbGate has the highest filter pass rate, at 95.2%, compared with 92.0% and 94.1%. Each architecture generates 6,000 candidates, with three passing designs per target evaluated across 99 targets.

  3. Generation across diverse held-out antigens. Across 502 targets, 95.6% of candidates pass the sequence filters. None of the pairs passing the numbering and length checks exactly duplicates a training pair. The interquartile ranges overlap with held-out antibodies across all seven predicted DeepSP properties. In the separate 496-target matched panel, mean ipTM is 0.460 for AbGate and 0.471 for known binders, while mean pLDDT is 0.787 and 0.788.

  4. A measurable response to antigen input. Across 100 targets, replacing the antigen changes generated CDR3 distributions more than repeated sampling does on average, with mean ΔMMD² = 5.69 × 10⁻⁴. CDR3 loops contribute to antigen recognition, so this result shows that antigen input affects a relevant part of the generated antibody sequence.

These results support a small trainable interface that responds to antigen input while maintaining candidate quality and similar mean structural confidence. The revised Discussion also identifies the need to strengthen computational evaluation against experimental outcomes, alongside future comparisons with unconditional generators and exploration of reinforcement learning.

We would appreciate your reconsideration of the quantitative support for AbGate’s contribution, taking these new comparisons together with the results already reported in Sections 3.2–3.3.

I thank the authors for their response. The additional experiments and analyses have helped clarify the positioning and effectiveness of the proposed method. However, the premise of excluding 3D structural information, together with the reliance on additional pretrained foundation models, limits its further practical applicability. Taking these considerations into account, I have decided to raise my rating score to 2.

Thank you for taking another look at our work. Your questions helped us explain more clearly what AbGate contributes and how our results support it.

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

This paper introduces AbGate, a parameter-efficient framework for generating antibody sequences conditioned on antigen sequences. The approach uses learned attention pooling and cross-attention to adapt a pretrained antibody language model while keeping most parameters frozen. The authors demonstrate generation across diverse held-out antigens and report structural confidence scores comparable to known binders.

Critical strengths

The authors present an interesting and computationally efficient approach to conditioning pretrained antibody language models on antigen sequences. The ability to accommodate variable-length antigens with a fixed conditioning representation is appealing. The model generates plausible antibody sequences across a diverse set of held-out targets while training relatively few parameters.

Critical weaknesses

My primary concern is that the results do not convincingly demonstrate antigen-specific antibody design. The evaluation relies heavily on predicted structural confidence, which is not necessarily indicative of binding. In fact, the authors’ own analysis of experimentally characterized antibodies finds that ipTM poorly distinguishes binders from nonbinders (AUC = 0.576). Furthermore, ipTM is strongly correlated with antigen length, complicating interpretation of the comparisons with known binders.

The paper also lacks convincing controls demonstrating that antigen conditioning meaningfully improves design. It is unclear whether the model learns target-specific interactions or primarily generates plausible antibody sequences using its pretrained representations.

Finally, the comparison with MAGE is limited to a single antigen, and although mean ipTM scores are similar, MAGE produces substantially higher-scoring individual candidates.

Questions for authors
  1. Can the authors demonstrate that antigen conditioning improves target-specific design relative to an unconditioned model or one conditioned on shuffled antigen sequences?
  2. Given the weak relationship between ipTM and experimental binding in Appendix F, how should the structural confidence results be interpreted as evidence of successful antibody design?
  3. Can the authors provide a more comprehensive comparison with MAGE, particularly regarding the ability to generate high-scoring candidates rather than simply matching average structural confidence?
2 comments

Thank you for asking how the conditioning and structural results support target-specific design. In response, we added an antigen-input control, clarified what each evaluation measures, and made the MAGE comparison settings explicit.

Conditioning controls. We focus on CDR3 loops because they contribute to antigen recognition. The new control asks whether their sequence distributions change when we replace the antigen, beyond the changes caused by repeated sampling. In Section 3.4 we use a fixed checkpoint to generate 20 pairs for each of 100 held-out antigens under three seeds, then repeat generation with replacement antigens from different groups selected for similar length. We compare CDR3 distributions after sequence filtering using MMD². After subtracting variation between runs with the same input, the mean ΔMMD² is 5.69 × 10⁻⁴. This positive value means that replacing the antigen changes the generated CDR3 distribution more than repeating generation with the same antigen does on average. Figure 8 shows the results across all 100 targets.

Our initial RF3 comparison did not establish a conditioning advantage, and further validation was not completed within the revision period. The Discussion now identifies comparisons with unconditional generators using evaluation tied to experimental binding as future work. The present sequence result establishes that the generator responds to antigen input.

Structural confidence and antigen length. We used ipTM to compare predicted complex confidence with MAGE and known binders, not as a criterion for successful binding. As you noted, the analysis in Appendix F showed that ipTM alone is a weak discriminator of binding in that panel. The submitted manuscript also evaluated sequence validity, novelty, and predicted antibody properties. We have revised the presentation to make the role of these assessments clearer alongside the structural comparisons.

For the comparison with known binders, we evaluate both groups against the same antigen for each of 496 targets. This holds antigen identity and length fixed within each comparison. The length-group analysis in Figure 5 further shows that AbGate and known binders have similar mean ipTM across all five antigen-length groups, even though scores decrease with length in both groups. These comparisons support similar predicted structural confidence across the evaluated targets. For future work, we aim to combine these metrics for candidate selection and test whether highly ranked candidates bind more often.

MAGE comparison. We have clarified the generation and selection settings in Section 3.3 to make the existing comparison easier to assess: all 20 AbGate samples are evaluated without structural preselection, whereas the 180 MAGE designs are selected by ESMFold pLDDT from 991 released designs. Mean ipTM is 0.434 versus 0.435 and mean pLDDT is 0.787 versus 0.789, with AbGate using approximately 1.1% of MAGE's trainable parameters.

As you noted, MAGE produces higher ipTM scores among its top candidates in this RBD comparison. Its maximum is 0.853, compared with 0.463 among our 20 designs. The 20-design subsampling analysis in Appendix E also favors MAGE on this metric. Our result is that AbGate achieves similar mean ipTM and pLDDT using approximately 1.1% of MAGE’s trainable parameters.

This comparison uses MAGE’s released RBD designs. A comparison under a common antigen split would require repeating MAGE’s full finetuning from its pretrained backbone, followed by candidate selection and structural evaluation. We have not completed that comparison for this revision.

We would appreciate your reassessment of the paper in light of the new conditioning control, the clarified MAGE comparison, and the existing evaluations across held-out antigens.

Your review raised an important question about what the conditioning control tells us about binding. If you have any remaining questions about how we interpret the CDR3 results alongside the structural scores, we would be happy to discuss them before the discussion closes.

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

This paper proposes AbGate, a parameter-efficient antibody generation framework conditioned on a target antigen. The main novelty is leveraging a pretrained antibody language model (LM) called ProGen2-OAS with minimal architectural modifications to enable antigen-conditional generation. These modifications include antigen embeddings from ESM-C, an attention pooling adapter, and cross-attention adapters, while the main backbone of the LM remains frozen during training. Standard supervised fine-tuning is performed using antigen–antibody paired data. Performance is evaluated from various aspects using the MAGE dataset. Its performance is comparable to that of MAGE on a shared target, while using only about 1% as many trainable parameters.

Critical strengths
  • The method is really simple and technically sound. Leveraging the LM's generative capability by training small cross-attention adapters for conditioning is clever.
Critical weaknesses
  • The performance may largely stem from the LM's generative capability; it is unclear how effectively antigen conditioning works. It is possible that even an unconditional antibody generative LM would perform well on the tasks benchmarked by the authors.
Questions for authors
  1. Could you provide more architectural ablation studies?

  2. Do the authors think this adapter training approach would also be effective when applying reinforcement learning to antibody design?

2 comments

Thank you for highlighting the role of antigen conditioning and suggesting architectural comparisons. We now present both comparisons in Section 3.4, with statistical and implementation details in Appendix A.4–A.5.

Effect of antigen conditioning. To address your concern about whether generation depends on antigen information, we test whether changing the antigen affects generated CDR3 sequences beyond random sampling. CDR3 loops are variable regions of the heavy and light chains that contribute to antigen recognition. With a fixed checkpoint, we generate 20 VH/VL pairs for each of 100 held-out antigens under three sampling seeds. We repeat this with each antigen replaced by an antigen from a different group selected for similar length. We compare the CDR3 distributions of candidates passing the sequence filters using MMD². We compute ΔMMD² by subtracting the variation between repeated runs with the same input from the difference after antigen replacement. Its mean is 5.69 × 10⁻⁴, indicating that antigen replacement changes the CDR3 distribution more than repeated sampling on average. Figure 8 shows the full distribution.

Architectural comparisons. We compare three ways of passing antigen information to the decoder. Mean pooling averages the ESM-C residue embeddings, and an MLP converts this average into one 256-dimensional conditioning vector for the entire antigen. This vector is supplied to an adapter at every decoder layer. K32 pooling first uses 32 learned queries to compute weighted summaries of the antigen residues, then combines these summaries into a single conditioning vector through an MLP. Our AbGate implementation keeps the 32 summaries separate so that each decoder layer can attend to them directly. The queries, projections, and adapters are trained jointly through the antibody next-token prediction loss, while the pretrained decoder blocks remain frozen. Mean pooling itself has no trainable parameters. The table counts trainable parameters across the full implementation, including the MLPs and decoder adapters.

ArchitectureTrainable parameters (M)Filter pass rate (%)Mean RF3 ipTM
AbGate8.4695.20.471
Mean pooling + vector adapters12.4492.00.472
K32 pooling + vector adapters15.2194.10.470

Pass rates cover 6,000 attempts per architecture. RF3 evaluates three randomly selected passing candidates per target across 99 targets. We use one checkpoint per architecture trained with the same frozen backbone, data split, and optimizer settings. AbGate achieves similar mean structural confidence with fewer trainable parameters and the highest filter pass rate among these alternatives.

Reinforcement learning. We appreciate the reviewer’s question on the potential extension to apply reinforcement learning in the AbGate framework. The pooler and adapters could be updated under a reward objective while the decoder blocks remain frozen. With the limited rebuttal time, we have not tested this extension for comprehensive evaluation on the advantages of RL over the current training strategy. Exploring the potential of this extension will be considered as future work, especially with the need of validating rewards against the intended biological outcome and evaluating improvements independently of the training reward.

We would welcome your feedback on the added conditioning and architecture comparisons.

Your question about the role of the pretrained antibody model motivated the comparison in Section 3.4. If there is anything in our response you would like us to clarify, we are available to follow up before the discussion closes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.