acceptodds
Under review as a conference paper at ICLR 2027

AbFUSE: Joint Antibody Optimization for Discriminating Similar Targets

Abstract

Advances in protein structure prediction enable model-guided antibody sequence optimization, yet suppressing binding to related negative targets can also reduce predicted binding to the intended target. We present AbFUSE, a training-free framework that optimizes one antibody sequence against a positive target and multiple negative targets using frozen structure and affinity predictors. A reference-score penalty limits decreases in predicted positive-target binding; an optional static local relation term compares mapped target neighborhoods. For chimeric antigen receptor (CAR) design, the objective can additionally include a sequence-level positively charged patch (PCP) term from CAR-Tonic Signal Tuner (CAR-Toner). We evaluate candidate quality on a computational benchmark and explore multi-target predictions in a case study involving STEAP family members. No updated AbFUSE benchmark candidate passed the prespecified interface-confidence gate. In an exploratory AlphaFold Server assessment, designs targeting STEAP family member 1 (STEAP1) showed lower predicted binder–target pairwise alignment error against STEAP1 than against STEAP2/3; this pattern also appeared in the baseline designs. These findings characterize the framework and its current validation limits without establishing experimental affinity or selectivity.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Antibody design, protein sequence optimization, binding selectivity, multi-target optimization, structure-guided design

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_kJsyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AbFUSE, a training-free framework for jointly optimizing antibody sequences against a positive target and multiple negative targets using frozen structure and affinity predictors. The method combines a positive-target binding preservation penalty, an optional local relation term, and a charged-patch objective for CAR-related applications. The authors evaluate AbFUSE on a 37-target computational benchmark and explore multi-target predictions in a STEAP family case study. However, none of the evaluated AbFUSE benchmark runs pass the predefined interface-confidence gate. The exploratory STEAP results show some predicted target-preference signals, but similar patterns are observed in baseline methods, and no experimental binding or selectivity measurements are provided.

Critical strengths
  1. Well-motivated multi-objective formulation. The paper addresses a relevant challenge in antibody optimization by considering positive-target binding preservation, negative-target suppression, and optional CAR-related sequence properties within a unified framework. The use of frozen predictors and a shared sequence representation provides a flexible approach to coordinating these objectives.

  2. Transparent evaluation and discussion of limitations. The authors clearly distinguish between accepted candidates and exploratory predictions, report unsuccessful screening outcomes, and avoid interpreting computational scores as evidence of experimental binding or selectivity. This transparency helps readers understand the current capabilities and limitations of the proposed framework.

Critical weaknesses
  1. Limited evidence of successful antibody optimization. Although the proposed framework is well motivated, none of the 140 evaluable benchmark runs pass the predefined interface-confidence gate, and the STEAP experiments also fail to produce accepted candidates. Consequently, the experiments do not demonstrate that AbFUSE can reliably generate candidates satisfying its intended design constraints. This substantially limits the evidence for the practical effectiveness of the framework.

  2. Insufficient validation of multi-target selectivity. The STEAP case study provides exploratory predictions, but the observed preference for STEAP1 over STEAP2/3 also appears in the baseline methods. Moreover, the reported predictions do not establish improved selectivity relative to the original antibody or demonstrate the benefit of joint positive- and negative-target optimization. As a result, the central advantage of the multi-target formulation remains insufficiently supported.

Questions for authors
  1. Could the authors further clarify the main factors responsible for these failures? In particular, is there evidence that the optimization improves positive-target interface quality relative to the initial sequences, even when the final candidates fail the predefined threshold?

  2. Could the authors clarify what evidence specifically supports the benefit of joint positive- and negative-target optimization in AbFUSE, rather than reflecting an existing preference captured by the structure predictors?

Reviewer_xYp7OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces AbFUSE, a framework for optimizing a shared antibody sequence against one positive target and multiple negative targets while optionally controlling a CAR-related sequence property. The method represents editable amino-acid positions using continuous sequence logits and performs gradient-based optimization through frozen structure and affinity prediction models. AlphaFold-based complex predictions provide structural information, while PPAP supplies predicted binding-affinity scores. The main optimization objective consists of several components:

  • Positive-target binding: Maximizing predicted affinity for the desired target.
  • Negative-target suppression: Minimizing an aggregation of predicted affinities for undesired targets.
  • Positive-target preservation: Penalizing decreases in predicted positive-target affinity relative to a reference score recorded during optimization.
  • Local relation term: Encouraging differences in local interaction support between mapped regions of positive and negative targets.
  • CAR-related property optimization: Guiding predicted PCP toward a specified interval using CAR-Toner, or using an alternative analytic surface-charge proxy. The method follows a four-stage optimization schedule. After optimization, candidates must pass structural confidence gates before inverse-folding-based sequence reconstruction and final screening. An optional mutation-context two-site proposal module is implemented but not evaluated. The experiments comprise three settings. First, a 37-target positive-target-only benchmark evaluates candidate quality under two charged-patch objectives. Of 148 scheduled slots, 140 are evaluable, and none passes the first quality gate, which requires binder pLDDT ≥ 50 and actifpTM ≥ 0.5. Second, two positive-target-only case studies, D-IL3 and D-PDL1, evaluate 50 additional slots. Again, none passes Gate 1. Third, a multi-target STEAP case study optimizes a 244-residue scFv with 14 editable H3 positions, targeting STEAP1 while suppressing predicted interactions with STEAP2 and STEAP3. The authors conduct an exploratory AlphaFold Server evaluation of 25 designed sequences from AbFUSE, Germinal, RFantibody, and BoltzGen. All five method groups exhibit positive model-derived PAE margins favoring STEAP1. However, Germinal achieves the highest mean STEAP1 interface confidence and selectivity margin in this comparison. The CAR-Toner analysis shows that AbFUSE's learned-objective condition produces a higher mean predicted PCP score than its analytic-proxy condition (39.58 versus 29.95), but only one of five learned-objective designs falls within the desired interval of 46–56. The authors conclude that the framework provides a formulation for joint antibody optimization, while acknowledging that the present experiments do not establish experimental binding affinity, selectivity, or CAR functionality.
Critical strengths
  1. The paper addresses a practically important multi-objective optimization problem. Antibody engineering requires balancing multiple potentially conflicting objectives. Improving specificity against closely related proteins may reduce affinity for the intended target, while modifications intended to regulate CAR signaling may also alter antigen recognition. The paper appropriately recognizes that these requirements should be optimized jointly rather than independently. This is a meaningful problem formulation with potential applications in antibody therapeutics and CAR-T engineering.
  2. The positive-target preservation mechanism is conceptually well motivated. The proposed reference-score penalty directly addresses a common difficulty in selectivity optimization: suppressing undesired binding without sacrificing desired binding. Rather than relying exclusively on a weighted combination of positive- and negative-target scores, the method introduces a reference-based penalty and final-screening condition intended to prevent deterioration below the initial predicted positive-target score. This provides a clear and interpretable design objective, although its practical effectiveness is not demonstrated by the reported experiments.
  3. The framework integrates multiple prediction objectives without retraining the underlying models. AbFUSE combines structure prediction, affinity prediction, local target-relation information, and CAR-related sequence-property prediction within a single optimization framework. The use of frozen predictors avoids the need to train a new end-to-end model and could make the approach adaptable to different antibody-design tasks. The distinction between the learned CAR-Toner objective and the analytic charge proxy is also useful, although the experiments do not isolate their individual contributions.
Critical weaknesses
  1. The proposed framework fails to produce a single accepted candidate in the main experiments. This is the most significant weakness. According to Table 1 (page 6), none of the 140 evaluable benchmark runs passes Gate 1. The mean actifpTM scores are only 0.127 and 0.124 for the learned and analytic-proxy conditions, respectively, substantially below the required threshold of 0.5. The additional D-IL3 and D-PDL1 case studies also yield zero passing candidates across 50 runs. As a result, the framework does not demonstrate its ability to generate antibody sequences satisfying its own candidate-quality criteria. Moreover, the failure occurs before candidate reconstruction and final screening. Therefore, a substantial portion of the proposed pipeline remains untested in the reported experiments. The authors should investigate the underlying cause of this failure and establish a functioning baseline before claiming practical utility.
  2. The paper lacks convincing evidence that the proposed optimization formulation improves selectivity. The central contribution is joint optimization of positive-target affinity and negative-target suppression. However, the main 37-target benchmark and the D-IL3/D-PDL1 case studies are positive-target-only tasks. Consequently, they do not evaluate the proposed multi-target selectivity objective. The only reported multi-target setting is the STEAP case study, where none of the native AbFUSE attempts passes Gate 1. Although the exploratory AF3 results show positive STEAP1-versus-STEAP2/3 PAE margins, the same pattern appears for Germinal, RFantibody, and BoltzGen. Therefore, the experiments do not establish that AbFUSE improves selectivity compared with existing methods or even relative to its initial antibody sequence.
Questions for authors

Q1. What is the primary cause of the zero Gate-1 success rate? None of the 140 evaluable benchmark runs passes Gate 1, and the mean interface-confidence scores are substantially below the required threshold. Could the authors provide an analysis identifying whether these failures originate from the initial sequences, structure prediction, gradient optimization, objective conflicts, or the chosen thresholds? How do the confidence metrics evolve over the four optimization stages? Q2. Does AbFUSE improve positive-target affinity or selectivity relative to the initial sequence? The paper introduces a reference-score preservation mechanism and a multi-target selectivity objective. Could the authors report predicted positive- and negative-target affinity scores before and after optimization, including for designs that fail Gate 1? This would help determine whether the optimization objective improves its intended scores even when the resulting structures do not satisfy the confidence requirements. Q3. What is the contribution of the positive-target preservation penalty? Could the authors compare AbFUSE with and without the reference-score penalty under identical settings? Does the penalty prevent predicted positive-target affinity deterioration during negative-target suppression? How sensitive are the results to the penalty weight and tolerance parameter? Q4. What is the benefit of the static local relation term? The relation term is one of the main mechanisms distinguishing AbFUSE from existing multi-target optimization methods. Could the authors provide an ablation comparing optimization with and without this term on the STEAP task? Does it improve the positive-versus-negative predicted affinity margin, or does it introduce additional optimization difficulties?

Reviewer_bdN1OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper utilizes a sequence optimization framework to design binders for CAR-Toner. Its main method is multi-objective learning so that the binder is specifically binding to positive targets, not to negative targets.

Critical strengths
  • Good motivation based on an important real-world antibody engineering problem.
Critical weaknesses
  • This paper mainly integrates existing predictors and optimization components into an application-specific pipeline, without establishing a new ML principle or demonstrating that the proposed optimization actually improves performance. A generalizable machine learning methodology is not introduced beyond an application-specific pipeline built on existing methods.

  • It's unclear what this paper contributes. The experimental results seem not successful and do not support the proposed ideas.

  • I appreciate the authors being honest, but the conclusion is somewhat open.

Questions for authors

None

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.