acceptodds
Under review as a conference paper at ICLR 2027

SPIN-BO: Source-Guided Protein–Protein Interaction Discovery for New Targets

Abstract

Protein engineering often requires optimizing a common mutational library against multiple interaction partners, but experimentally mapping a complete interaction landscape for every new partner is costly. Existing landscapes from related partners offer valuable prior knowledge, yet transferring this knowledge is challenging because shared mutational trends coexist with partner-specific effects. We introduce SPIN-BO, a source-guided Bayesian optimization framework for discovering high-response variants against a new interaction partner under a limited target-assay budget. SPIN-BO reuses previous interaction landscapes through two complementary transfer mechanisms. First, an adaptively calibrated source-response mean uses sparse target assays to learn how the source profiles should be weighted and combined for the new partner. Second, a source-supervised protein-pair representation, learned from pretrained protein language model embeddings, defines the geometry of a target-specific Gaussian-process residual, enabling sparse target measurements to generalize across the mutant library while capturing target-specific structure not explained by transferred source patterns. As target assays accumulate, SPIN-BO continually recalibrates how source landscapes contribute to prediction, updates predictive uncertainty, and selects the next variants to test. Controlled semi-synthetic experiments show that response-level and representation-level transfer provide complementary benefits across different source–target relationships. Across 54 retrospective discovery tasks from the human bZIP interaction network, SPIN-BO achieves a mean optimum-success rate of 95.0% at a budget of 200 target assays, compared with 73.3% for target-only Bayesian optimization. These results show that previously measured interaction landscapes can be systematically reused to accelerate discovery for new interaction partners.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Bayesian optimization, protein engineering, protein–protein interactions, knowledge transfer, protein language models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_2HDTOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SPIN-BO, a source-guided Bayesian optimization framework for discovering high-response protein variants against a new interaction partner under a limited target-assay budget. SPIN-BO transfers information from previously measured source interaction landscapes through two mechanisms: an adaptively calibrated source-response mean and a source-supervised protein-pair representation used by a target-specific GP residual. The resulting surrogate is updated as target assays accumulate and drives batch Thompson-sampling acquisition. The method is evaluated in controlled semi-synthetic experiments and on 54 retrospective discovery tasks from a human bZIP interaction dataset.

Critical strengths
  1. The cross-partner discovery setting is practically meaningful. Reusing existing interaction landscapes to reduce assays for a new partner is well motivated, and the formulation explicitly accounts for the fact that source partners may share broad mutational trends while retaining target-specific effects.

  2. The two transfer mechanisms are clearly separated and carefully analyzed. The controlled simulations independently vary source–target response association and representation-aligned smoothness, providing useful insight into when response-level and representation-level transfer help.

Critical weaknesses
  1. The setting assumes access to complete and perfectly aligned source interaction landscapes over the same mutant library. These historical measurements are treated as freely available and do not count toward the assay budget. While reasonable for retrospective studies, this is a strong practical assumption and may substantially limit applicability when source landscapes are incomplete, measured on different libraries, or themselves expensive to obtain.

  2. The empirical comparison with existing transfer Bayesian optimization methods is incomplete. The related-work section discusses RGPE, DeltaBO, and DynMeanBO as closely related transfer-BO approaches, but the experiments primarily compare against target-only BO, internal SPIN-BO ablations, ALDE-style, EVOLVEpro-style, and random search. The absence of matched transfer-BO baselines makes it difficult to assess how much improvement comes from the proposed formulation relative to existing source-transfer strategies.

  3. The additional benefit of representation-level transfer appears modest on the real bZIP evaluation. At 200 assays, SPIN-BO (Both) achieves 95.0% optimum success versus 93.9% for Response Transfer alone, while Response Transfer slightly outperforms the full model on Recall@20 (89.2% vs. 88.7%). The authors themselves note that response-level transfer accounts for most of the gain in this dataset. This weakens the empirical evidence that both proposed components are necessary in realistic data.

  4. The real-data validation remains narrow. The retrospective experiments use a single human bZIP interaction dataset, with the primary analysis covering ten panels across four mutant backgrounds. Although the number of retrospective tasks is reasonably large, all tasks come from the same biological system and experimental platform. It is therefore unclear whether the gains generalize to other PPI families, assay types, or mutational libraries.

  5. The response-level calibration may become difficult to scale to many source partners. With SS sources, the calibrated mean contains S+1S+1 coefficients, and the paper explicitly requires the initial target design matrix to have full column rank with n0>S+1n_0>S+1. Thus, increasing the number of available source landscapes also increases the minimum number of target assays needed before optimization can begin, which may conflict with the assay-efficient motivation.

Questions for authors
  1. How would SPIN-BO operate when source interaction landscapes are incomplete or defined over partially different mutant libraries? The current formulation appears to rely on candidate-wise aligned source responses for the same mutant library, which may be restrictive in practical protein-engineering campaigns.

  2. Could the authors include matched comparisons with existing transfer-BO approaches such as RGPE, DeltaBO, or DynMeanBO? These methods are closely related to the proposed setting, and such comparisons would better isolate the benefit of SPIN-BO’s adaptive source-response calibration and residual GP formulation.

  3. How does SPIN-BO scale as the number of source partners increases? Since the initial target set must satisfy n0>S+1n_0 > S+1 and a full-rank source-response design, does increasing the number of source landscapes substantially increase the required initial assay budget? Could regularization or source selection alleviate this dependence?

  4. Is the improvement from representation-level transfer statistically significant on the retrospective bZIP tasks? The full model is only slightly better than response-level transfer alone on several metrics and is slightly worse on Recall@20. Paired confidence intervals or bootstrap differences between these variants would help clarify the additional contribution of the learned representation.

  5. Can the authors evaluate SPIN-BO on an additional experimental PPI or protein-engineering dataset? The current real-data evaluation is concentrated on the human bZIP interaction system, so broader experimental validation would help establish generality across different proteins, interaction mechanisms, and assay platforms.

4 comments

W1 & Q1: Incomplete, unaligned, and potentially costly source measurements

We thank the reviewer for highlighting these practical concerns. We have extended SPIN-BO to accommodate incomplete source landscapes and clarified the scope of our assay-budget setting.

  • SPIN-BO accommodates incomplete source landscapes with different observed mutant subsets. We allow each source partner to have its own set of observed mutants. Missing source responses are predicted using source-specific models trained on available source measurements and ESM2 embeddings, while observed responses are preserved. The pair encoder is trained only on genuinely observed interactions (Section 3.1; Appendix D.1). We further evaluated SPIN-BO across all 54 bZIP tasks under 30–70% source missingness with MCAR and MNAR mechanisms. The results show that GP-based completion preserves substantial discovery gains over Target-only BO and generally outperforms source-mean completion and three source-aware transfer-BO baselines, although performance declines with increasing missingness (Appendix D.2–D.3, Tables 12–13).

  • Our assay budget focuses on new target measurements. Our setting assumes that historical source interaction data are already available, and the objective is to minimize the additional assays required for a new target. Accordingly, historical source measurements are treated as previously collected rather than cost-free, and only newly acquired target assays count toward the optimization budget.


W2 & Q2: Matched transfer-BO baselines

We thank the reviewer for highlighting the importance of comparisons with existing source-aware transfer-BO methods. Following your suggestion, we have expanded our experimental comparisons under matched source information and target-assay budgets.

  • Added three source-aware transfer-BO baselines. We have incorporated DeltaBO [1], DynMeanBO [2], and RGPE [3] into our experiments. As these methods were originally developed for general Bayesian optimization, we adapted their core transfer mechanisms to our finite mutant-library setting using pretrained ESM2 embeddings and historical source-response profiles. Details are provided in Appendix B.3, with updated comparisons in Sections 4–5.

  • SPIN-BO outperforms all three transfer-BO baselines in the primary bZIP benchmark. Across 54 retrospective tasks, SPIN-BO (Both) achieves 82.2% and 95.0% optimum success at 100 and 200 assays, respectively, compared with 73.2% and 91.1% for DeltaBO, the strongest transfer-BO baseline in optimum success. SPIN-BO also outperforms all three baselines in normalized best score, Recall@5, and Recall@20 at both budgets (Section 5, Figure 3 and Table 2). The key results are reproduced below.

    Revised Table 2 (selected methods): Multi-source bZIP discovery at 100 and 200 assays; higher is better.

    Method100: NB100: Opt100: R@5100: R@20200: NB200: Opt200: R@5200: R@20
    SPIN-BO (Both)0.9900.8220.7500.7230.9980.9500.8910.887
    DeltaBO-MS-TS0.9760.7320.7040.7020.9910.9110.8590.872
    DynMeanBO-MS-TS0.9780.6470.6820.6090.9840.7320.8170.803
    RGPE-2022-MS-TS0.9650.6320.6380.6120.9790.7340.7920.798
    Target-only BO0.9590.5360.5570.4880.9830.7330.7690.769

    NB: normalized best; Opt: optimum success; R@5/R@20: Recall@5/Recall@20.

  • Additional experiments support these findings across diverse settings. We further evaluated the three baselines in controlled simulations (Table 1), single-source and biology-agnostic bZIP experiments (Appendix C, Tables 7–9), incomplete source landscapes (Appendix D, Tables 12–13), and two additional PPI systems, IgG1 and KRAS (Appendix E, Tables 14–15). SPIN-BO generally achieves stronger discovery performance across these settings, although some metric-specific exceptions remain.


References

[1] Lin, H., Zhao, B., Kolar, M., & Liu, C. (2026). Provable Accelerated Bayesian Optimization with Knowledge Transfer. AISTATS 2026. https://proceedings.mlr.press/v300/lin26c.html

[2] Qu, C., Liu, M., Lan, J., Dong, S., & Liu, Z. (2026). Incorporating Expert Priors into Bayesian Optimization via Dynamic Mean Decay. ICLR 2026. https://openreview.net/forum?id=ylX5PWfK0v

[3] Feurer, M., Letham, B., Hutter, F., & Bakshy, E. (2022). Practical Transfer Learning for Bayesian Optimization. arXiv:1802.02219. https://arxiv.org/abs/1802.02219

W3 & Q4: Incremental representation benefit and paired uncertainty

We thank the reviewer for raising this important question. We agree that the modest differences between SPIN-BO (Both) and Response Transfer in the real bZIP benchmark warrant a careful paired uncertainty assessment. We therefore conducted additional paired analyses and examined the contribution of the learned representation more directly.

  • Paired analysis of incremental discovery gains. We compared Both and Response Transfer across all 54 bZIP tasks using ten matched campaign seeds, aggregating differences equally across the ten panels. At 200 assays, Both improves optimum success by 1.12 percentage points (descriptive 95% background-bootstrap interval: [0.28, 1.88] points) and Recall@5 by 1.56 points ([-0.23, 3.00]), while Recall@20 decreases by 0.52 points ([-1.98, 0.60]). Although the background-bootstrap interval for optimum success excludes zero, a sensitivity analysis jointly resampling mutant backgrounds and campaign seeds yields intervals containing zero for all four discovery metrics. We therefore do not claim a robust, statistically significant incremental advantage.

  • Clarifying the distinction between the two variants. Both variants retain the adaptive source-response mean and target-specific GP residual, with identical fitting and acquisition procedures. They differ only in residual GP geometry: Both uses the source-supervised pair encoder, whereas Response Transfer uses raw pretrained ESM2 embeddings (Section 4; Appendix B.3). Their similar performance is therefore plausible when the source-response mean already captures substantial transferable information. This motivates examining whether the learned encoder provides additional predictive value after accounting for the source-response mean.

  • New held-out analysis demonstrates improved residual prediction. We conducted an additional analysis across all 54 bZIP tasks, randomly selecting 100 or 200 target measurements for training, with identical training sets across representations. For each training set, we first fit a common OLS source-response mean and compute the residuals. We then fit separate Matérn-kernel ridge regressions to these residuals using either raw ESM2 embeddings or the source-supervised encoder, selecting hyperparameters by five-fold cross-validation on the training data. Each residual prediction is added to the same source-response mean. We evaluate the resulting predictions on the five highest-response mutants in the held-out pool, using their true responses only for evaluation, without tuning specifically for high-response prediction. As shown below, the pair encoder reduces held-out top-5 MSE by 6.14% at 100 assays and 5.35% at 200 assays relative to raw ESM.

    Additional rebuttal analysis: Held-out top-5 prediction MSE across 54 bZIP tasks (lower is better).

    Target assaysSource mean onlyRaw ESMPair encoderReduction vs. raw ESM
    1002.9162.7222.5556.14%
    2002.6272.3162.1925.35%

With the source-response mean held fixed, these results provide evidence that the learned encoder improves residual correction in high-response regions. While this does not establish statistically significant incremental gains in the full BO campaigns, it supports the value of source-supervised representations beyond the transferred source-response mean.

W4 & Q5: Additional experimental PPI systems

We thank the reviewer for suggesting broader experimental validation. In response, we have extended our evaluation to two additional experimental PPI systems with distinct protein families, assay readouts, and mutant-library sizes. In each dataset, one partner serves as the target and the remaining five as sources (Section 5; Appendix E, Figures 9–10 and Tables 14–15).

  • IgG1 Fc–Fc$\gamma$R binding. Using the dataset of Keith et al. [1], we evaluated 4,282 shared single-substitution mutants across six receptor conditions, with experimentally measured effective binding affinities (log KaK_a). At 560 target assays, SPIN-BO (Both) achieves 83.3% optimum success, compared with 66.7% for DeltaBO and 20.0% for Target-only BO. It also achieves 86.7% Recall@5 and 90.2% Recall@20, outperforming all three transfer-BO baselines on both recall metrics (Appendix E.3, Figure 9 and Table 14).

  • KRAS–effectors/DARPins. Using the dataset of Weng et al. [2], we evaluated 2,417 shared single-substitution mutants across six binding partners, with experimentally measured BindingPCA fitness scores. At 410 target assays, SPIN-BO (Both) achieves 82.0% Recall@5 and 73.3% Recall@20, outperforming Target-only BO and all three transfer-BO baselines on both metrics (Appendix E.4, Figure 10 and Table 15).

    Selected final-budget results from revised Tables 14–15; higher is better.

    MethodIgG1: OptIgG1: R@5IgG1: R@20KRAS: OptKRAS: R@5KRAS: R@20
    SPIN-BO (Both)0.8330.8670.9020.8000.8200.733
    DeltaBO-MS-TS0.6670.6270.6900.8000.7800.658
    DynMeanBO-MS-TS0.3000.4070.4650.7000.7670.640
    RGPE-2022-MS-TS0.5330.4530.5600.8000.7730.693
    Target-only BO0.2000.2070.2500.8000.6270.428

    Opt: optimum success; R@5/R@20: Recall@5/Recall@20.

Overall, these results provide additional evidence that SPIN-BO can improve discovery across different PPI systems and assay readouts, with performance gains varying across datasets and metrics.


References

[1] Keith, A. D., Xu, T., Chernova, T. A., et al. (2025). Mapping Affinity and Allostery in Human IgG Antibody Fc Region–Fcγ Receptor Interactions. bioRxiv. https://doi.org/10.1101/2025.03.28.645945

[2] Weng, C., Faure, A. J., Escobedo, A., & Lehner, B. (2024). The Energetic and Allosteric Landscape for KRAS Inhibition. Nature, 626, 643–652. https://doi.org/10.1038/s41586-023-06954-0

W5 & Q3: Many-source scaling, identifiability, and regularization

We thank the reviewer for highlighting this scalability concern. We acknowledge the increasing initial assay requirement of unregularized calibration and discuss two complementary strategies: biologically informed source selection and ridge regularization.

  • Limitations of unregularized calibration. With SS sources, the calibrated mean contains S+1S+1 coefficients. The GLS/REML formulation requires n0>S+1n_0>S+1 and a full-column-rank initial design, so the minimum initial assay budget grows linearly with SS. Correlated source profiles may further destabilize estimation. Our primary bZIP experiments use S=3S=3–99 with n0=20n_0=20, satisfying these requirements, but larger source banks can make this constraint restrictive (Section 3.5; Appendix A.1).

  • Biologically informed source selection. Selecting a smaller, informative source subset reduces the calibration dimension without requiring additional target assays. Our bZIP panels are constructed using subfamily relationships, sequence similarity, and known interaction biology (Appendix C.2, Table 6). Biologically grouped panels exhibit stronger source–target response correlations than biology-agnostic panels (mean Pearson: 0.791 versus 0.076–0.077; Table 10) and larger discovery gains. At 100 assays, their optimum-success improvement over Target-only BO is 28.5 percentage points, compared with 4.0–10.0 points for biology-agnostic panels (Appendix C.5, Table 11 and Figure 8). These descriptive findings support the practical motivation for biologically informed source selection.

  • Ridge regularization removes the full-rank constraint. We added a theoretical extension in Appendix A.5 that penalizes source coefficients while leaving the intercept unpenalized:

    β^λ=(H⊤Cη−1H+λR)−1H⊤Cη−1y,R=diag⁡(0,IS),λ>0.\widehat{\beta}_{\lambda}=\left(H^\top C_\eta^{-1}H+\lambda R\right)^{-1}H^\top C_\eta^{-1}y,\qquad R=\operatorname{diag}(0,I_S),\quad \lambda>0.

    For H=[1n,X]H=[\mathbf{1}_n,X] and any nonzero u=(a,b⊤)⊤u=(a,b^\top)^\top,

    u⊤Aη,λu=(a1n+Xb)⊤Cη−1(a1n+Xb)+λ∥b∥22>0.u^\top A_{\eta,\lambda}u=(a\mathbf{1}_n+Xb)^\top C_\eta^{-1}(a\mathbf{1}_n+Xb)+\lambda\lVert b\rVert_2^2>0.

    provided CηC_\eta is positive definite. Hence, Aη,λA_{\eta,\lambda} remains positive definite even when S≥nS\geq n or source columns are linearly dependent, removing the original full-rank and n0>S+1n_0>S+1 requirements. Appendix A.5 also derives the corresponding Gaussian posterior and restricted marginal likelihood, preserving uncertainty in the regularized source coefficients.

Together, these strategies address the initial-assay dependence in complementary ways: source selection limits the number of coefficients to be calibrated, while ridge regularization removes the algebraic identifiability constraint for larger source banks.

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

SPIN-BO addresses the problem of efficiently discovering high-performing protein variants for a new interaction partner. The key idea is to reuse previously measured interaction landscapes from related source partners through two transfer mechanisms: response-level transfer, which adaptively combines source interaction profiles, and representation-level transfer, which learns an interaction-aware protein-pair embedding from source data. These transferred signals are combined with a target-specific Gaussian process model to guide Bayesian optimization and select the next mutants to test.

Critical strengths
  1. The problem formulation is clear. It addresses a realistic bottleneck in protein engineering: expensive screening for every new interaction partner. The idea of reusing existing experimental landscapes instead of treating each target as an independent optimization problem is good.
  2. This paper proposes a well-motivated transfer strategy. Response-level transfer captures shared mutation trends across partners. Representation-level transfer captures functional similarity beyond directly measured variants.
  3. Controlled simulations clearly separate when different transfer mechanisms are useful. Retrospective bZIP experiments provide realistic validation under limited assay budgets.
Critical weaknesses
  1. This paper has a strong assumption of complete source landscapes which limits the usefulness. Real DMS datasets are often sparse and contain missing variants. Maybe introducing the soomthness to response-level transfer to interpolate the missing mutation will solve this problem.
  2. Limited comparison with modern protein foundation models. This paper does not directly compare against strong PLM-based mutation effect predictors or structure-based affinity prediction models like Boltz-2/Nesso-1.
Questions for authors
  • How does SPIN-BO perform when source interaction landscapes are incomplete or sparse?
  • Can mutation relationships learned from historical landscapes improve transfer beyond ESM representations?
  • How sensitive is the method to differences in assay normalization and measurement scales? Many DMS data can not dierctly compare between different target since they are acquired from different assay.
  • Can the authors provide a pre-screening criterion to determine whether a new target is suitable for source-guided optimization? In the Exp2 dataset, how strong is the correlation between source and target interaction landscapes, and how smooth is the underlying mutation landscape? Can the applicability of SPIN-BO be assessed based on biological or biochemical properties?
  • In Exp2, SPIN-BO (Both) performs similarly to SPIN-BO (Response Transfer), suggesting that the gain mainly comes from transferring historical interaction landscapes. Could the authors analyze the learned source weights and investigate whether they correspond to biological similarity between partners? Additionally, are there alternative transfer strategies (e.g., similarity-based transfer, multi-task GP, or matrix completion) that could serve as stronger baselines?
4 comments

W1 & Q1: Incomplete or sparse source landscapes

We thank the reviewer for raising this practical concern. We have extended SPIN-BO to accommodate incomplete source measurements and conducted additional retrospective experiments.

  • Extended the framework to accommodate incomplete source landscapes. Each source ss may be observed on its own mutant subset Os\mathcal O_s. For each incomplete source, a separate imputation model predicts missing responses from observed interactions while retaining measured values. The completed profiles support adaptive response transfer, whereas the encoder is trained only on observed source interactions. Different sources may have different observed mutant subsets, provided candidate sequences are available (Section 3.1 and Appendix D.1).

  • New retrospective experiments assess the impact of source missingness. Across all 54 bZIP tasks, we evaluated 30%, 50%, and 70% missingness under MCAR and response-dependent MNAR mechanisms, using three mask replicates and ten matched BO seeds. We applied GP-based completion with ESM2 mutant embeddings to SPIN-BO and three transfer-BO baselines (DeltaBO, DynMeanBO, and RGPE), alongside source-mean completion as an alternative (Appendix D.2).

  • SPIN-BO remains effective under substantial source missingness. SPIN-BO (GP) generally achieves stronger discovery performance at 30% and 50% missingness and the highest optimum success among GP-completed transfer-BO methods across all six settings. Even at 70% missingness, it outperforms Target-only BO in optimum success under both masking mechanisms, although performance declines with increasing sparsity and some recall metrics favor competing methods. These results support SPIN-BO's applicability to incomplete source landscapes while acknowledging the challenges of severe missingness (Appendix D.3, Tables 12–13).


W2: Limited comparison with modern protein foundation models

We thank the reviewer for suggesting comparisons with recent protein foundation models. We clarify how they relate to our research objective and expand our transfer-BO comparisons.

  • Differences in prediction targets. Boltz-2 [1] supports biomolecular complex structure prediction, but its affinity prediction module focuses on protein–small-molecule binding. Similarly, Nesso-1 [2] predicts binding affinity and binder likelihood from protein sequences and small-molecule ligands. While these models represent important advances, their affinity predictions do not correspond to our experimentally measured protein–protein interaction responses, such as bindingPCA log fold changes.

  • Focus on assay-efficient sequential discovery. Our objective is to identify high-response mutants through sequential experiments under limited target-assay budgets, rather than predict a specific molecular property. SPIN-BO accommodates different quantitative interaction responses. We therefore compare BO and active-learning methods supporting iterative candidate selection under matched budgets. Existing comparators include Target-only BO, ALDE-style, and EVOLVEpro-style; Target-only BO and EVOLVEpro-style both incorporate pretrained ESM2 representations (Section 4 and Appendix B.3).

  • Expanded comparisons with stronger source-aware transfer-BO baselines. Our original external comparators used target measurements without historical source interactions. We have now added DeltaBO [3], DynMeanBO [4], and RGPE [5], adapting their transfer mechanisms to our mutant-library setting with pretrained ESM2 embeddings and historical source profiles. Adaptations are detailed in Appendix B.3, with updated results in Sections 4–5. In the primary bZIP benchmark, SPIN-BO (Both) outperforms all three across all four discovery metrics at both 100 and 200 assays (Table 2, Figure 3). These comparisons strengthen the evidence for SPIN-BO in source-guided sequential discovery.


References

[1] Passaro, S., Corso, G., Wohlwend, J., et al. (2025). Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction. bioRxiv. https://doi.org/10.1101/2025.06.14.659707

[2] Shenoy, N., Errington, D., Bengio, E., et al. (2026). Nesso-1: Accelerating Open-Source Binding Affinity Predictions. bioRxiv. https://doi.org/10.64898/2026.08.01.742196

[3] Lin, H., Zhao, B., Kolar, M., & Liu, C. (2026). Provable Accelerated Bayesian Optimization with Knowledge Transfer. AISTATS 2026. https://proceedings.mlr.press/v300/lin26c.html

[4] Qu, C., Liu, M., Lan, J., Dong, S., & Liu, Z. (2026). Incorporating Expert Priors into Bayesian Optimization via Dynamic Mean Decay. ICLR 2026. https://openreview.net/forum?id=ylX5PWfK0v

[5] Feurer, M., Letham, B., Hutter, F., & Bakshy, E. (2022). Practical Transfer Learning for Bayesian Optimization. arXiv:1802.02219. https://arxiv.org/abs/1802.02219

Q2: Can historical interactions learn mutation relationships beyond ESM?

We thank the reviewer for this insightful question. Our matched ablations directly examine whether historical source interactions provide useful representation-level information beyond pretrained ESM embeddings.

  • A matched ablation isolates representation-level transfer. SPIN-BO (Representation Transfer) and Target-only BO share the same intercept-only mean, residual GP, fitting procedure, and acquisition strategy. The key difference is the GP representation: a pair encoder trained on historical source interactions versus raw pretrained ESM embeddings (Section 4; Appendix B.3, Table 4).

  • Controlled simulations demonstrate the benefits of source-supervised representations. At (ρ,γ)=(0.6,0.9)(\rho,\gamma)=(0.6,0.9), Representation Transfer reduces INR from 0.0831 to 0.0347 and improves Recall@20 at 200 assays from 0.035 to 0.350 relative to Target-only BO (Section 4, Table 1). Even when source–target response correlation is absent but shared smooth structure is strong, (ρ,γ)=(0,0.9)(\rho,\gamma)=(0,0.9), Representation Transfer reduces INR from 0.0831 to 0.0329 (Appendix B.5, Table 5). This suggests historical interactions can reveal useful geometric structure beyond direct source–target response association.

  • Retrospective discovery supports the value of learned representations. Across 54 real bZIP tasks, Representation Transfer improves optimum success from 53.6% to 62.0% at 100 assays and from 73.3% to 79.2% at 200 assays relative to Target-only BO. It also outperforms Target-only BO across all four discovery metrics at both budgets (Section 5, Table 2). Together, these findings show that historical interactions can improve discovery through representation-level transfer beyond pretrained ESM embeddings alone.


Q3: Assay normalization and measurement scales

We thank the reviewer for raising this practical concern. We agree that responses measured across different assays may have different numerical scales. We clarify how SPIN-BO handles such differences and report new full-pipeline sensitivity experiments.

  • Adaptive source calibration accommodates affine scale differences. Under source-specific transformations YsmS′=as+bsYsmSY^{S\prime}_{sm}=a_s+b_sY^S_{sm} (bs≠0b_s\neq 0), SPIN-BO's intercept and source coefficients can absorb shifts and rescalings while preserving the transferred mean exactly (Sections 3.3–3.5). With a fixed representation, the REML objective changes only by a constant independent of covariance parameters, preserving its optimum in exact arithmetic. This invariance need not extend to the source-supervised encoder, motivating full-pipeline experiments.

  • New full-pipeline experiments assess source-response transformations. We selected ten bZIP target tasks at random, one per biological panel, and used five matched BO seeds and three perturbation replicates. We considered:

    1. Affine perturbations: Independent source-specific shifts and rescalings,

      YsmS′=μs+σsds+bs(YsmS−μs).Y^{S\prime}_{sm}=\mu_s+\sigma_s d_s+b_s(Y^S_{sm}-\mu_s).

      Here μs\mu_s and σs\sigma_s are the source mean and SD, ds∼Unif(−1,1)d_s\sim\mathrm{Unif}(-1,1), and bs∼LogUniform(0.5,2.0)b_s\sim\mathrm{LogUniform}(0.5,2.0).

    2. Asinh perturbations: Source-specific nonlinear transformations,

      usm=asinh⁡(csYsmS−μsσs).u_{sm}=\operatorname{asinh}\left(c_s\frac{Y^S_{sm}-\mu_s}{\sigma_s}\right).

      YsmS′=μs+σsusm−uˉssd⁡(us).Y^{S\prime}_{sm}=\mu_s+\sigma_s\frac{u_{sm}-\bar{u}_s}{\operatorname{sd}(u_s)}.

      Here cs∼LogUniform(0.5,2.0)c_s\sim\mathrm{LogUniform}(0.5,2.0). These transformations preserve response ordering, means, and SDs while altering response spacing.

    Perturbations were sampled independently for each source and replicate. Transformed responses were used for both adaptive calibration and encoder retraining; target responses, candidate libraries, initial assays, and BO procedures were unchanged. Comparisons used the same tasks and matched seeds.

    The table reports performance at 200 target assays. Perturbed results were averaged over seeds and replicates within tasks, then equally across panels; the Original reference uses the same tasks and seeds.

    Source-response conditionNBOpt (%)R@5 (%)R@20 (%)
    Original SPIN-BO (Both)0.99981398.0091.6088.50
    Affine-perturbed SPIN-BO (Both)1.000000100.0092.6788.87
    Asinh-perturbed SPIN-BO (Both)0.99987598.6792.2789.17

    NB: normalized best; Opt: optimum success; R@5/R@20: Recall@5/Recall@20.

    At 200 assays, neither perturbation substantially changes any aggregate metric. Normalized Best differs by less than 0.0002; Optimum Success and Recall differences are at most 2.0 percentage points. These results support stability under the tested affine and nonlinear source-response transformations.

Q4: Pre-screening criteria and biological applicability

We thank the reviewer for this suggestion. We have clarified our biological source-selection criteria in Appendix C.2 and added analyses of source–target relationships, landscape smoothness, and discovery gains in Appendix C.5.

  • Biologically informed pre-screening criteria. We propose a practical source-selection strategy based on information available before target assays. Drawing on established biological classifications and expert input, we distinguish three levels of relatedness: (H) close subfamily relationships and high sequence homology, (M) broader family or conserved-domain relationships, and (L) known functional interaction networks. These criteria guide our ten biologically grouped bZIP panels, including C/EBP, Fos-related, and AP-1/Jun/ATF2 groups, and provide practical screening guidance without requiring target measurements (Appendix C.2, Table 6).

  • Quantitative characterization of source–target relationships and landscape smoothness. We added retrospective diagnostics comparing biologically grouped and biology-agnostic source panels (Appendix C.5, Table 10). We calculate sequence identity, ESM2 cosine similarity, Pearson and Spearman response correlations, and normalized local landscape roughness. Biologically grouped panels exhibit substantially higher source–target response correlations and smoother mutation landscapes. Selected results from Table 10 are summarized below.

    Revised Table 10 (selected metrics): Biological similarity and empirical landscape diagnostics (equal-panel means).

    DiagnosticGroupedAgnostic 9Agnostic 5
    Sequence identity0.5790.3410.345
    ESM2 cosine similarity0.9530.8800.881
    Signed Pearson correlation0.7910.0760.077
    Signed Spearman correlation0.7140.0730.074
    Target roughness R10R_{10} (lower is smoother)0.4820.8610.861

    R10R_{10} measures normalized response variation among ten nearest neighbors in raw ESM2 space; lower values indicate smoother landscapes.

  • Biologically grouped panels are associated with greater discovery gains. We compared SPIN-BO (Both) with Target-only BO across source-panel settings (Appendix C.5, Table 11 and Figure 8). At 100 assays, SPIN-BO gains 28.5 percentage points in optimum success and 23.5 points in Recall@20 with grouped panels. In contrast, the gains with biology-agnostic panels are 4.0–10.0 and 3.7–4.8 points, respectively. Similar advantages persist at 200 assays. These findings support biologically informed source selection, although cross-setting comparisons remain descriptive because benchmark tasks differ. Together, the biological criteria and retrospective diagnostics provide guidance for identifying promising source panels.


Q5: Learned source weights and alternative transfer strategies

We thank the reviewer for suggesting an analysis of learned source weights and comparisons with alternative transfer strategies.

  • Learned source weights show modest associations with biological similarity. To investigate whether adaptive source coefficients reflect biological relatedness, we analyzed their relationships with source–target sequence identity and ESM2 cosine similarity across all 54 bZIP tasks. We adjusted coefficients for source-response scale differences and calculated within-task Spearman correlations, aggregated equally across biological panels. At 200 assays, the correlations are 0.272 for sequence identity and 0.127 for ESM2 similarity. These results indicate modest positive associations, but biological similarity does not consistently determine the learned source weights. This is expected because coefficients are jointly calibrated from target measurements and reflect each source's predictive contribution conditional on the others, rather than independent biological importance.

  • Comparisons with alternative source-aware transfer strategies. As discussed in our response to W2, we expanded comparisons to DeltaBO, DynMeanBO, and RGPE, adapting their transfer mechanisms to our cross-partner discovery setting using historical source interaction profiles and pretrained ESM2 embeddings. Adaptations are detailed in Appendix B.3, with comparisons updated in Sections 4–5, Table 2 and Figure 3.

    While dedicated similarity-based weighting, multi-task GP, and matrix-completion baselines are not directly evaluated, we prioritized DeltaBO, DynMeanBO, and RGPE as established source-aware transfer-BO methods enabling consistent comparisons under matched source information and target-assay budgets.

Thanks for your responses, all my concerns are resolved. I raise my score to 3.

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

This paper studies how to leverage experimental data from source partners to efficiently identify protein mutants that exhibit strong interaction responses with a new target partner, all while operating under a limited experimental budget. The authors propose a framework called SPIN-BO, based on Bayesian optimization, which utilizes source interaction data through two complementary knowledge-transfer mechanisms.

The authors evaluated the framework using semi-synthetic N-TIMP2/MMP data and real-world human bZIP interaction data. The results show that the two transfer mechanisms are complementary under varying conditions. In a series of 54 retrospective discovery tasks involving human bZIP, SPIN-BO achieved a 95.0% success rate in identifying the optimal protein variants within a budget of 200 target assays, significantly outperforming the 73.3% success rate of a target-only Bayesian optimization approach.

Critical strengths
  1. The idea of reusing existing source interaction landscapes to guide discovery for a new target is new for the problem setting. Adaptively adjusting the source contributions using target observations provides a plausible way to account for differences between interaction partners.

  2. The paper provides detailed descriptions of the training settings and experimental protocols, with additional implementation details and mathematical derivations in the appendices. The additional single-source and biology-agnostic experiments further help characterize the conditions under which the method is effective.

Critical weaknesses
  1. The paper discusses DeltaBO (Lin et al., 2026) and DynMeanBO (Qu et al., 2026) in the related work, but neither is included in the experimental comparisons. Since the ALDE-style and EVOLVEpro-style baselines use only target measurements, comparisons with source-aware transfer BO methods would provide stronger evidence for the benefits of the proposed adaptive source calibration.

  2. The pair encoder is trained to predict source responses, but this objective does not explicitly ensure that the learned embedding geometry captures the covariance of target-specific residuals. In the multi-source bZIP experiment (Table 2), OptSuccess increases from 0.939 with response-only transfer to 0.950 with both mechanisms, so a more direct test of this assumption would be helpful.

Questions for authors
  1. Could the authors provide experimental comparisons with appropriately adapted DeltaBO (Lin et al., 2026) and DynMeanBO (Qu et al., 2026) under matched source information and target-assay budgets?

  2. Could the authors provide evidence that the source-supervised embedding geometry captures the relationships between target-specific residuals after accounting for the source-response mean?

3 comments

W1 & Q1: Comparisons with source-aware transfer BO

We thank the reviewer for highlighting the importance of comparisons with existing source-aware transfer-BO methods. Following your suggestion, we have substantially expanded our experimental comparisons to better assess the benefits of SPIN-BO.

  • Added three source-aware transfer-BO baselines. We have incorporated DeltaBO (Lin et al., 2026), DynMeanBO (Qu et al., 2026), and RGPE (Feurer et al., 2022) into our experiments. As these methods were originally developed for general Bayesian optimization rather than protein–protein interaction discovery, we adapted their core transfer mechanisms to our finite mutant-library setting using pretrained ESM2 embeddings and historical source interaction profiles. We have detailed these adaptations in Appendix B.3 and updated the comparisons in Sections 4–5.

  • SPIN-BO outperforms all three transfer-BO baselines in the primary bZIP benchmark. Across 54 retrospective tasks, SPIN-BO (Both) achieves 82.2% and 95.0% optimum success at 100 and 200 assays, respectively, compared with 73.2% and 91.1% for DeltaBO, the strongest transfer-BO baseline in optimum success. SPIN-BO also outperforms all three transfer-BO baselines in normalized best score, Recall@5, and Recall@20 at both budgets (Section 5, Table 2 and Figure 3). The key results are summarized below.

    Revised Table 2 (selected methods): Multi-source bZIP discovery at 100 and 200 target assays (equal-panel means; higher is better).

    Method100: NB100: Opt100: R@5100: R@20200: NB200: Opt200: R@5200: R@20
    SPIN-BO (Both)0.9900.8220.7500.7230.9980.9500.8910.887
    DeltaBO-MS-TS0.9760.7320.7040.7020.9910.9110.8590.872
    DynMeanBO-MS-TS0.9780.6470.6820.6090.9840.7320.8170.803
    RGPE-2022-MS-TS0.9650.6320.6380.6120.9790.7340.7920.798
    Target-only BO0.9590.5360.5570.4880.9830.7330.7690.769

    NB: normalized best; Opt: optimum success; R@5/R@20: Recall@5/Recall@20. Bold indicates the best value among the five displayed methods.

  • Additional experiments further support SPIN-BO's advantages across diverse settings. Beyond the primary multi-source benchmark, we have extended the comparisons to controlled simulations (Table 1), single-source bZIP discovery (Appendix C.3, Table 7), biology-agnostic source panels (Appendix C.4, Tables 8–9), incomplete source landscapes (Appendix D, Tables 12–13), and two additional PPI systems (Appendix E, Tables 14–15). Across these evaluations, SPIN-BO generally demonstrates stronger discovery performance than the three transfer-BO baselines, although some metric-specific exceptions remain. In particular, it achieves higher optimum success on IgG1 and higher top-variant recall on KRAS than all three baselines. Overall, these expanded experiments provide broader evidence for the effectiveness of SPIN-BO.


References

[1] Lin, H., Zhao, B., Kolar, M., & Liu, C. (2026). Provable Accelerated Bayesian Optimization with Knowledge Transfer. AISTATS 2026. https://proceedings.mlr.press/v300/lin26c.html

[2] Qu, C., Liu, M., Lan, J., Dong, S., & Liu, Z. (2026). Incorporating Expert Priors into Bayesian Optimization via Dynamic Mean Decay. ICLR 2026. https://openreview.net/forum?id=ylX5PWfK0v

[3] Feurer, M., Letham, B., Hutter, F., & Bakshy, E. (2022). Practical Transfer Learning for Bayesian Optimization. arXiv:1802.02219. https://arxiv.org/abs/1802.02219

W2 & Q2: Does source-supervised geometry capture target residual structure?

We thank the reviewer for highlighting the important distinction between source-response supervision and target-specific residual modeling. Following your suggestion, we conducted additional analyses across all 54 retrospective bZIP tasks to directly assess whether the learned representation improves target-residual prediction.

  • Motivation and limitations of representation-level transfer. We agree that source-response supervision does not mathematically guarantee alignment with the target-residual covariance. However, our setting assumes no target measurements are available before optimization, making it infeasible to directly supervise an encoder using target-specific residuals at initialization. Instead, we build upon pretrained ESM2 embeddings learned from large-scale protein sequences and further train a pair encoder using historical source interactions. Since practical protein-engineering campaigns often involve biologically related partners, we hypothesize that source supervision can refine these representations toward interaction-relevant features transferable to a new target. The resulting embedding provides an interaction-informed inductive bias, while the target-specific residual GP is fitted as new target measurements accumulate (Sections 3.4–3.5). We acknowledge that the effectiveness of this approach depends on the shared interaction structure between source and target partners.

  • Controlled simulations demonstrate when representation transfer is beneficial. At (ρ,γ)=(0.6,0.9)(\rho,\gamma)=(0.6,0.9), where source–target association and shared smooth structure are both informative, SPIN-BO (Both) reduces INR from 0.0138 to 0.0096 and improves Recall@20 at 200 assays from 0.655 to 0.810 relative to Response Transfer (Section 4, Table 1). Even without source–target response association, at (ρ,γ)=(0,0.9)(\rho,\gamma)=(0,0.9), Representation Transfer reduces INR from 0.0831 to 0.0329 relative to Target-only BO (Appendix B.5, Table 5). These results suggest that learned representations can improve discovery when source and target landscapes share informative smooth structure.

  • New held-out analyses provide evidence of improved residual prediction. To directly examine the learned geometry, we conducted an additional prediction analysis across all 54 bZIP tasks using matched, randomly sampled target training sets. For each split, both representations share the same estimated OLS source-response mean, while their residual corrections are modeled using Matérn-kernel ridge regression. Hyperparameters were selected using five-fold cross-validation, without further tuning for high-response prediction. We evaluated prediction errors on the five highest-response mutants in each held-out candidate pool, using their true responses only for evaluation. At 100 and 200 target assays, the pair encoder reduces held-out top-5 MSE by 6.14% and 5.35%, respectively, compared with raw ESM. At 100 assays, the improvement is also consistent across all four mutant-background averages. The results are summarized below.

    Additional rebuttal analysis: Held-out top-5 prediction MSE across 54 bZIP tasks (lower is better).

    Target assaysSource mean onlyRaw ESMPair encoderReduction vs. raw ESM
    1002.9162.7222.5556.14%
    2002.6272.3162.1925.35%
  • These findings provide empirical support for our inductive bias. Because the source-response mean is identical across representations, the improvements reflect more effective residual correction beyond the transferred source information, particularly among high-response mutants. Together with the controlled ablations (Section 4, Table 1) and retrospective discovery results (Section 5, Table 2), these findings support the practical value of source-supervised, interaction-informed geometry. Although this benefit is not uniform across all tasks and budgets, the new analysis provides encouraging evidence for its usefulness in discovery-relevant regions of the target landscape.

Thanks for the detailed response and the additional experiments. The comparisons with DeltaBO, DynMeanBO, and RGPE address the main concern about the missing source-aware baselines, and the additional prediction analysis provides further support for the representation-level transfer. These results strengthen the experimental evidence for SPIN-BO, and I will maintain my original positive score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

Dear Area Chair and Reviewers,

We sincerely thank the Area Chair and all three reviewers for their thoughtful assessments and constructive feedback. We are encouraged that the reviewers recognized the novelty and practical significance of our cross-partner discovery setting, particularly the idea of reusing historical protein–protein interaction landscapes to accelerate discovery for new targets. We also appreciate the positive assessments of our adaptive source-response calibration and complementary representation-level transfer, which together accommodate shared mutational patterns and target-specific effects. Furthermore, the reviewers highlighted the carefully designed controlled simulations, strong retrospective bZIP results, and detailed methodological descriptions and experimental protocols.

We have carefully addressed the reviewers' concerns in our individual responses. Guided by their suggestions, we have substantially strengthened our experimental evaluation, broadened the applicability of SPIN-BO, and clarified its methodological assumptions. The main revisions are summarized below.

  • Stronger source-aware transfer-BO comparisons. We incorporated three additional baselines—DeltaBO, DynMeanBO, and RGPE—adapted to our cross-partner discovery setting under matched source information and assay budgets. In the primary 54-task bZIP benchmark, SPIN-BO outperforms all three methods across all four discovery metrics, strengthening the evaluation of our transfer framework.

    See Sections 4–5, Table 2, and Appendix B.3.

  • Broader experimental validation across PPI systems. We extended our retrospective evaluation to two additional experimental systems: IgG1 Fc–Fc$\gamma$R interactions and KRAS interactions with effectors and DARPins. These datasets span different protein families, assay readouts, and larger mutant libraries, with SPIN-BO demonstrating strong discovery performance across both systems.

    See Section 5 and Appendix E.

  • Extension to incomplete source landscapes. We extended SPIN-BO to accommodate source-specific observed mutant subsets through source-only response completion. Additional experiments under 30–70% source missingness with MCAR and MNAR mechanisms show that substantial discovery gains can be retained despite incomplete historical measurements.

    See Sections 3.1 and 5, and Appendix D.

  • Biological source selection and transfer diagnostics. We clarified biological pre-screening criteria and conducted new analyses of source–target similarity, response correlations, and landscape smoothness. Biologically grouped panels exhibit stronger response correlations and larger discovery gains than biology-agnostic panels, supporting biologically informed source selection.

    See Section 5 and Appendices C.2 and C.5.

  • Scalability and regularized source calibration. We clarified the identifiability limitations of unregularized calibration and introduced a ridge-regularized theoretical extension that removes the full-rank requirement while retaining uncertainty-aware Gaussian prediction.

    See Section 3.5 and Appendix A.5.

We believe these additional experiments and methodological clarifications substantially strengthen our paper and provide a more comprehensive evaluation of SPIN-BO's effectiveness and practical applicability. We again thank the Area Chair and reviewers for their valuable feedback, which has greatly helped improve our work.

Thank you.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.