acceptodds
Under review as a conference paper at ICLR 2027

19 Degrees of Freedom Beat Full-Parameter Fine-Tuning in Fixed-Backbone Antibody Design

Abstract

Inverse-folding models generate antibody CDR sequences compatible with antibody–antigen structures, but are trained for structural compatibility rather than binding. We introduce PACT (Position-Additive Composition Tilt), which steers a frozen decoder with a shared amino-acid bias of only 19 effective degrees of freedom. Its coefficients are estimated in closed form from within-backbone reward differences, requiring no gradient training. Across three antigens, PACT improves mean Rosetta interface reward by 0.32–0.63 and increases shape complementarity on every target. In contrast, advantage regression yields no significant reward gain over the frozen reference, while filtered SFT reduces reward and falls behind PACT by 0.99–1.83 units. To investigate the limited benefit of full-parameter training, we examine how it reweights candidate designs. Across six tested objectives, fine-tuning produces log-likelihood ratios relative to the frozen reference that are nearly uncorrelated with reward differences among candidates for the same backbone. Checkpoint analysis of advantage regression further shows that these candidates retain highly aligned gradients throughout training. The fitted PACT biases also transfer across targets and improve ESM-IF and AntiFold without refitting. Together, these findings support direct steering of residue preferences as an effective and transferable alternative to the tested full-parameter updates.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
fixed-backbone antibody design

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_QbfKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces PACT, a lightweight reward-steering method for fixed-backbone antibody design. PACT freezes the inverse-folding decoder and adds a shared amino-acid logit bias with 19 effective degrees of freedom, estimated from within-backbone reward differences without gradient-based training. The paper also analyzes why several tested full-parameter fine-tuning objectives fail to effectively re-rank candidates within the same backbone. Experiments on three antigens show improvements in computational interface reward and shape complementarity, with additional transfer experiments across antigens and inverse-folding decoders.

Critical strengths
  1. Simple and practical intervention. PACT is lightweight, requires no full-model optimization, and can be applied directly to frozen inverse-folding models.
  2. Interesting empirical observation. The paper provides evidence that the tested full-parameter fine-tuning methods do not effectively achieve within-backbone reward re-ranking, while a much simpler output-space intervention performs well on held-out backbones.
Critical weaknesses
  1. Limited methodological novelty. PACT is essentially a frozen decoder with a shared amino-acid logit bias, closely resembling a learned composition prior or simple logit adjustment. The coefficient estimator is also a straightforward association between residue identity and centered reward. Overall, the method feels more like an effective engineering trick than a new modeling or optimization framework.
  2. The technical contribution is relatively thin. In my view, the most interesting contribution is the empirical observation that simple low-dimensional steering can outperform some full-parameter fine-tuning setups, rather than the PACT algorithm itself. Once the extensive fine-tuning diagnostics are separated from the method, the core technical contribution is quite limited for an ICLR full paper.
  3. A crucial low-dimensional baseline is missing. A natural comparison would freeze the decoder and train exactly the same shared amino-acid output bias using gradient-based optimization. Without this baseline, it is unclear whether the improvement comes from the proposed estimator or simply from using a low-dimensional output-space parameterization. This is the most important missing control in my assessment.
  4. Part of the gain may reflect a generic composition effect. The side-chain volume prior performs similarly to PACT on HER2 and IL7R, with no clear advantage for the learned bias on these targets. This suggests that some of the improvement may arise from generic changes in residue size, interface area, or overall composition rather than learned target-specific preferences. This weakens the novelty claim further.
Questions for authors
  1. Can the authors compare PACT against a frozen decoder with the same shared output bias trained directly by gradient descent? This comparison would materially affect my assessment of whether the proposed estimator itself contributes beyond the low-dimensional parameterization.
  2. How do the authors interpret the fact that the side-chain volume prior is competitive with PACT on HER2 and IL7R? Can the authors provide evidence that the gains are not primarily explained by a generic composition or residue-size effect?
1 comment

Dear Reviewer Qbfk, thank you for reviewing the paper and taking the time to consider our response and the additional experiments we conducted during the rebuttal period. Since the rebuttal deadline is approaching, we just wanted to let you know that if you have any remaining questions or concerns, we would be very happy to address them. If our response addressed your comments properly, we would greatly appreciate your consideration of a higher score. We appreciate your time and fully understand that you may need more time to review the results!

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

The authors argue that inverse-folding models are trained for structural compatibility rather than binding, so their highest-probability designs need not be the best binders, and that reward-based full-parameter fine-tuning does not deliver the reweighting its own objective predicts: across six objectives, log-likelihood ratios come out nearly uncorrelated with reward differences among candidates for the same backbone, which the authors attribute to those candidates’ gradients being highly aligned, so that updates shift their likelihoods together rather than separating them by reward. To this end, the authors propose PACT, which freezes the decoder and adds a shared amino-acid bias of 19 effective degrees of freedom to its output logits at every designable CDR position, with each coefficient estimated in closed form as the mean within-backbone centered reward over the designs in which that residue appears at a variable position. Through experiments on three antigens with independent refolding at evaluation, the authors demonstrate that PACT improves interface reward and shape complementarity on held-out backbones where the tested full-parameter objectives do not, and that the fitted coefficients transfer to other antigens and to other inverse-folding decoders without refitting.

Critical strengths
  1. The premise that inverse-folding likelihood is not binding affinity is reasonable, and the authors turn it into something testable rather than asserting it: the KL-regularized optimum predicts that higher-reward candidates for the same backbone should gain likelihood relative to the reference, and measuring that correlation across six objectives and 155 recorded epochs gives a concrete diagnostic rather than an end-to-end performance claim.
Critical weaknesses
  1. Full-parameter training here fits 1.66M weights on 2,048 designs from 128 backbones, and since designs sharing a backbone differ at a mean Hamming distance of 4.1, the effective sample size is closer to the backbone count — a regime where updating every weight is unlikely to generalize regardless of the objective. The comparison therefore spans only two points on the capacity axis, 19 coefficients against 1.66M weights, with nothing between them: a LoRA or adapter arm would show whether the benefit comes from the particular form of a composition bias or simply from reducing capacity to match the data. DPO is also left out of the refolded comparison despite being one of the six objectives in the diagnostic, even though within-backbone reward orderings supply preference pairs directly and a pairwise objective avoids the baseline-estimation issue the paper raises for advantage regression. With the strongest full-parameter comparator absent and the capacity axis unexplored, the conclusion that full-parameter updates are a poor mechanism for expressing reward preferences is hard to separate from the weaker one that this data regime does not support fitting 1.66M weights.

  2. Sharing one bias across every designable position is biologically hard to justify. CDR positions differ in structural environment and in how much contact they make with the antigen — H3 apex residues sit in a different context from framework-proximal H1 positions — so the same substitution should not carry the same preference everywhere, and the paper’s own per-loop extension finds exactly that, with leucine favored in H1 (+0.050) and disfavored in H3 (−0.042). That extension is the only test of the restriction, run once on HER2 and reported as unresolved (+0.026, p = 0.41), which does not show that position-dependence is unimportant so much as that this particular comparison was underpowered. What would settle it is running the per-loop variant on all three targets with the same budget and correction as the primary comparison, and ideally a variant conditioned on something structural rather than loop identity — burial or distance to the epitope at each position — since loop labels are a coarse proxy for the environment that should drive residue preference in the first place.

Questions for authors
  1. Appendix D.4 notes that the bias-strength sweep and subsequent protocol development used RF2 outcomes, which is the held-out evaluation pipeline, and that the total model-selection cost therefore exceeds coefficient fitting alone. The pilot in Table 18 is described as a separate vector and design set, but the choice of a maximum logit bias near 1.9 still traces back to measurements on that pipeline. How would the strength be set for a new target without consulting held-out scores, and how sensitive are the primary results to that choice?
1 comment

Dear Reviewer FeKv, thank you for reviewing the paper and taking the time to consider our response and the additional experiments we conducted during the rebuttal period. Since the rebuttal deadline is approaching, we just wanted to let you know that if you have any remaining questions or concerns, we would be very happy to address them. If our response addressed your comments properly, we would greatly appreciate your consideration of a higher score. We appreciate your time and fully understand that you may need more time to review the results!

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

This paper introduces PACT, a method for steering frozen inverse-folding models for fixed backbone antibody design. PACT estimates a shared amino-acid bias from within-backbone reward differences and adds it to the decoder logits, requiring only 19 effective degrees of freedom and no gradient-based training. Across three antigens, it improves a Rosetta-based interface reward and shape complementarity after independent RF2 complex refolding. The paper also compares PACT with full-prameter adaptation, analyzes the limited reward alignment observed during fine-tuning, and demonstrates transfer of the fitted biases across antigens and decoders.

Critical strengths
  1. Simple and effective steering. PACT requires only a shared amino acid bias without gradient training, yet improves both interface reward and shape complementarity across three antigens after independent RF2 refolding.
  2. Informative controls and transfer results. Shuffled coefficient and volume prior controls help characterize the learned preferences, while transfer across antigens and decoders demonstrates their reuse. Backbone-level paired comparisons appropriately account for dependence among designs.
Critical weaknesses
  1. Limited support for superiority over full-parameter fine-tuning. Only two objectives receive refolded evaluations, and filtered SFT lacks validation-based checkpoint selection. The gradient analysis does not establish the cause of training failure. Stronger baseline tuning and repeated runs are needed to support the central comparison.
  2. Uncertain significance beyond composition effects. The surrogate reward lacks held-out predictive validation. PACT does not significantly outperform the volume prior on two targets, and decoder-transfer gains mainly reflect increased interface area without significant shape-complementarity improvements. These findings limit conclusions about improved binding or interface quality.
Questions for authors

Please address the concerns and requested clarification in the Critical Weakness section.

Reviewer_1qTUOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors present PACT as a closed form based on a 19-degree-of-freedom shared amino-acid logit bias from within-backbone reward differences, and show that this approach improves held-out interface reward and shape complementarity across three antigens while transferring across targets and decoders. This is based on the key insights that standard full-parameter fine-tuning of inverse-folding models (AbMPNN) fail to reweight same-backbone candidate sequences according to reward. The diagnostic contribution is genuinely interesting and the paper is among the most carefully controlled. However, the method is a position-independent amino-acid frequency bias and cannot distinguish buried vs. exposed amino acid residues and the reward it optimizes has no held-out validation against experimental binding. The limitations presented are direct and are insightful.

Critical strengths

The gradient-kernel diagnostic is a neat contribution to understanding fine-tuning failure in this setting. The decomposition in Eq. 5 showing the separation of the residual response which drives reward-aligned reweighting from the mean-coupling term is rigorous and an empirical evaluation showing that it persists even after 60 epochs underscores the importance of their findings.

I felt that the controls were well designed and presented to report both the strengths and weaknesses in the paper. I also felt that the separation of the fitting and evaluation oracles was a good design choice. Using IgFold for coefficient fitting and RF2 for held-out evaluation directly addresses the concern that fitting-time improvements might not survive independent structural prediction.

PACT is genuinely simple and reproducible. The method requires no gradient training, no learning rate, no convergence criterion.

Critical weaknesses

The cross-decoder transfer gains are driven by increased buried surface area with no statistically resolved shape-complementarity improvement on any target-decoder pair. Further, the side-chain volume control is unresolved against PACT on two of three targets, leaving the question of whether the transferable signal is a learned binding preference or a size prior.

The reward model in Eq. 3 has no held-out predictive validation against experimental binding and is trained on single-residue SKEMPI mutations and applied to full CDR redesigns on generated backbones. The per-design signal-to-noise ratio against RF2 repeat variability is under 1 and therefore it seems like every result in the paper inherits this unquantified surrogate gap.

The title claims PACT beats full-parameter fine-tuning, but only two of six tested objectives receive held-out refolding evaluation. Moreover the filtered SFT is a weak comparator run without KL penalty or reward-based checkpoint selection. I also felt that the authors some how left the other natural approaches such as LoRA or properly tuned DPO with refolding evaluation from the evaluation, which could have strengthened their results.

The 19-DoF bias is position-independent and cannot distinguish a residue at a contact position from one in an adjacent turn (surface vs. buried vs. exposed amino acid residues), which are important for epitope constrained design.

Questions for authors

I'd recommend that the authors report the held-out predictive performance of the SKEMPI reward model against the expected free-energy changes corresponding to a unit of reward improvement. Without calibration against experimental binding, it it impossible to assess whether the detected gains are biologically meaningful or within the noise floor of the surrogate. Similarly, I am curious to see what happens when the fitted PACT bias is replaced with a pure side-chain-volume bias matched in both L1 norm and support, and then evaluated separately from the aggregate reward.

I would also suggest the authors present some evidence for natural baselines such as LoRA, or DPO with within-backbone pairs and reward-based early stopping through the same RF2 refolding evaluation. The current comparison tests PACT against regression and filtered SFT which seems to overfit by design. Approaches such as LoRA or DPO may help clarify whether the gradient-alignment issue is fundamental or an artifact of the tested training procedures.

Does the bias change which residues appear at interface-contact versus non-contact CDR positions, or does it shift composition uniformly? I feel that decomposing the composition shift by contact classification would distinguish a binding-relevant signal from a generic size effect.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.