acceptodds
Under review as a conference paper at ICLR 2027

AlloShift: Finite-Dose Population-Directed Binder Design

Abstract

Protein function often depends on the populations of different conformational states. Binders can alter this balance by preferentially binding one state, but state preference alone does not determine the effect at a feasible concentration: a selective binder may be too weak to engage the target, while a potent binder may bind both states. We introduce **ALLOSHIFT**, linking state-resolved affinities, baseline populations, and free binder concentration to an exact finite-dose rule for maximizing a desired state or matching a target population. learns state-resolved affinities from sparse and censored measurements and supplies scores for binder screening and peptide design. Across three measured-affinity panels, affinity-only screening misses the optimum in 70.4% of separating one-binder tasks; the balance of potency and selectivity changes with dose. improves receptor-cold affinity and state-direction prediction, and its scores improve model-evaluated peptide guidance and reranking. Together, these results provide a population-aware design objective and a learned route to candidate prioritization.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Peptide binder design, Agonist, Allosteric design, Discrete flow matching

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_aieHOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

AlloShift proposes a finite-dose population-control framework for conformationally selective binder design. Instead of optimizing affinity alone, it integrates state-specific KdK_d, baseline conformational populations, and binder concentration to select binders that maximize desired state populations. The authors develop Qbind to predict state-resolved affinities from sparse measurements and use it for binder screening and peptide generation guidance. Evaluations on measured affinity landscapes show that population-aware ranking can differ from affinity-based selection and improve computational candidate prioritization.

Critical strengths
  1. Clear formulation of finite-dose allosteric binder selection.
  2. Connects affinity, selectivity, and concentration in a principled way.
  3. Shows simple but important counterexamples where affinity-only ranking fails.
Critical weaknesses
  1. The methodological novelty appears limited. The proposed framework largely consists of predicting state-resolved affinities with Qbind and applying a deterministic thermodynamic model to calculate the binder-induced population shift. While the finite-dose formulation provides a useful perspective, the downstream optimization is fully determined by the predicted affinities, making the contribution more of an objective reformulation.
  2. Limited benchmarking of Qbind. Qbind is the only learned component in the framework, yet its evaluation is limited to internal baselines. Comparisons with existing state-of-the-art affinity prediction methods (e.g., Boltz-2 and other modern affinity predictors) are missing
  3. The main contribution is unclear, as the paper focuses extensively on the finite-dose formulation, which is largely a deterministic optimization given state-resolved affinities, while the key learned component, Qbind, receives limited methodological development and evaluation.
Questions for authors

See weakness

2 comments

We thank the reviewer for recognizing the value of connecting affinity, selectivity, and concentration. We fully understand the request to clarify the contribution and broaden the learned-model comparisons. To address it, we completed external affinity benchmarks and organized the evidence around selecting a binder and feasible dose for a desired conformational population.

1. Methodological contribution

“The methodological novelty appears limited.”

We introduce joint binder-and-concentration selection for an explicit equilibrium population target. Starting from classical thermodynamic linkage (Wyman, 1964 and Monod et al., 1965), we derive the reachable population interval, identify the appropriate concentration, and compare candidate binders under a finite dose ceiling. We also quantify the consequences of affinity-estimation errors through decision regret.

This matters because the best selection rule changes with the available dose. Across 28 binders in three measured-affinity panels, affinity-only screening missed the optimum in 70.4% of 513 separating one-binder decisions. We obtained the following regrets under two dose ceilings. Lower is better.

Dose ceilingAffinity-only regretSelectivity-only regret
0.1 μM0.00000.0074
100 μM0.04530.0000

At low dose, absolute affinity determines whether sufficient engagement is possible. At higher dose, state preference becomes more consequential. We account for both when selecting the binder and concentration and identify whether the requested population is reachable.

Related work addresses multistate design (Leaver-Fay et al., 2011), equilibrium ensemble prediction (BioEmu, Lewis et al., 2025), and affinity prediction (Boltz-2, Passaro et al., 2025). Our contribution connects state-resolved affinities to a concentration-constrained intervention decision and evaluates that decision on measured-affinity landscapes.

2. New external affinity benchmarks

“Comparisons with existing state-of-the-art affinity prediction methods ... are missing.”

We added comparisons with ProAffinity-GNN, GraphDTA, and PRODIGY on the same 162 exact-affinity measurements from 107 receptors. We report the submitted state scorer and the affinity-focused variant separately.

MethodAffinity MAE ↓
ProAffinity-GNN1.2711
Qbind-Affinity1.2748
Cold constant1.3217
Qbind-State1.3279
GraphDTA1.3644
PRODIGY1.7672

We obtained closely matched point estimates with Qbind-Affinity and ProAffinity-GNN. Their paired MAE difference was +0.0037 [−0.1991, +0.1962]. We selected Qbind-Affinity after examining internal receptor-cold results and report this matched-subset analysis as exploratory.

On the full receptor-cold cohort of 1,663 measurements across 450 receptors, Qbind-Affinity achieved MAE 1.3244, compared with 1.4075 for the constant control. The paired difference was −0.0831 [−0.1566, −0.0013]. For the 28-candidate decision panel, both Qbind variants returned usable predictions for all 28 candidates. The evaluated external workflows had zero decision coverage under our eligibility protocol, which we will report alongside coverage and failure details.

We also examined Boltz-2. Its reported affinity endpoint is on an IC50 scale, so we keep those outputs separate from the measured dissociation-constant comparison. We use a common endpoint for the numerical benchmark.

3. Connecting prediction to the intervention decision

“The main contribution is unclear ...”

We connect three steps in AlloShift: estimating state-resolved affinities, determining reachable populations, and selecting a binder and free concentration. We now evaluate the complete decision using out-of-fold predictions for both binder and dose. On 3,956 target-matching tasks, population errors were 0.2418 for finite-dose Qbind and 0.2416 for affinity-only selection. With measured-dose optimization of the nominated binder, regret was 0.0376 versus 0.0474. We use these comparisons to quantify the effect of concentration-estimation error on the benefit from binder selection.

We further assessed computational generation across 24 targets, two directions, and three seeds. Evaluator-assigned gains were +0.113 [0.047, 0.186] and +0.058 [0.001, 0.119] with two reporting ensembles from the same model family, both used during development. We report population calculations and generation scores at their computational scope.

We thank the reviewer again for encouraging these additions. The concentration-dependent selection results establish the decision problem, and the new external benchmarks substantially broaden the evaluation of its learned implementation. We respectfully ask the reviewer to increase the score and support acceptance of AlloShift if you believe your concerns have been met.

Thank you for the additional experiments and results during the rebuttal. I still think that the paper’s contributions are not sufficiently supported. For the estimation and selection task, the selection rule follows from classical thermodynamic relationships, so it is unsurprising that this rule identifies the optimal choice. For guided generation, the in silico evaluations use scoring models from the same family as the guidance model. Also, the population shifts calculated using the thermodynamic formula depend directly on KdK_d. Without actual measurements of conformational populations, these results do not establish whether there is a real improvement. Therefore I keep my original score.

Reviewer_pb63OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents AlloShift, a finite-dose framework for state-selective binder design. The central idea is that binder selection should jointly consider state selectivity, absolute affinity, and achievable dose, because selectivity determines the limiting population shift while affinity determines whether that shift can be reached at a feasible concentration. The paper introduces an analytical decision layer for desired-state maximization and target-population matching, and proposes Qbind to predict state-resolved affinities from binder features, biochemical context, and receptor-state representations. Qbind predictions are then used for low-budget binder screening and for guiding/reranking peptide generation with PepTune. The validation includes measured-affinity panels, receptor-cold affinity/state-direction prediction, computational peptide generation, and a Boltz-2 structural case study.

Critical strengths

1.Clear and useful finite-dose formulation. The paper cleanly separates the roles of selectivity, absolute affinity, and dose. The reachable-interval formulation for target matching is elegant, interpretable, and more relevant to practical intervention design than optimizing affinity alone. 2.Decision-oriented and carefully scoped evaluation. The paper evaluates downstream binder selection rather than only prediction error, and shows that affinity-only selection can be suboptimal under finite-dose constraints. I also appreciate that the manuscript carefully distinguishes measured affinities from model-derived population shifts and evaluator-based generation outcomes.

Critical weaknesses
  1. The learned model has a structural limitation for binder-specific selectivity, and the empirical gain over a state-only model is modest. Qbind uses an additive readout. When measurement contexts are matched across states, the binder/context contribution cancels in Δ^\hat{\Delta}, so the model cannot rank different binders by state preference within that state pair. This is consequential because binder-specific state selectivity is central to the paper’s motivation. Consistent with this concern, state-direction balanced accuracy improves only from 0.6113 for the state-only control to 0.6195 for full Qbind, suggesting that much of the signal may come from receptor-state features rather than binder-specific interactions.
  2. The evaluation does not yet fully validate the claimed population-directed binder-design pipeline. In the learned finite-dose experiment, Qbind ranks binders, but the final dose is optimized using measured affinities. Thus, the experiment isolates binder selection rather than jointly testing predicted binder-and-dose decisions. A true end-to-end evaluation should select both binder and dose from predicted KdK_d values and evaluate the chosen intervention using measured affinities. In addition, the generation results are evaluated computationally, and the structural case study relies on static Boltz-2 predictions. Since the paper centers on conformational states and population shifts, the lack of MD or another dynamic analysis leaves the structural interpretation less convincing.
Questions for authors

Clarification questions

  1. How much variation does predicted Δ^\hat{\Delta} show across binders within matched-context state pairs? In particular, can the authors report the variance or ranking diversity of Δ^\hat{\Delta} within such pairs?
  2. How much state-direction performance remains if the structural state term is removed or fixed? This would help clarify whether Qbind learns binder-specific selectivity or mainly state-level bias.
  3. Why was no MD or other dynamic analysis performed for the structural case study? Given that the paper focuses on conformational states and population shifts, what supports the use of static Boltz-2 structures as sufficient structural evidence? Questions that could materially affect my assessment
  4. Did the authors test an explicit binder × state interaction model, for example a regularized or low-rank interaction baseline? A positive result here would directly address my main architectural concern.
  5. Can the authors provide an end-to-end evaluation in which both the binder and the dose are selected using predicted KdK_d values, and the resulting intervention is evaluated using measured affinities?
2 comments

We thank the reviewer for the encouraging assessment of the finite-dose formulation and the specific suggestions for strengthening the learned component. We completed the requested state-term interventions, interaction comparisons, and prediction-based binder-and-dose evaluation. We also clarify the structural example and describe the experimental validation now in preparation.

1. State dependence and within-pair variation

“How much variation does predicted ... Δ show across binders within matched-context state pairs?” “How much state-direction performance remains if the structural state term is removed or fixed?”

We examined 33 matched-context pools. The maximum range of predicted affinity differences was 1.5 × 10⁻⁷, and the maximum standard deviation was 3.8 x 10⁻⁸. All candidates tied at 10⁻⁴ precision. We therefore attribute within-pair ranking to potency and the predicted direction to receptor-state information.

To test this attribution, we intervened on the state term in the 25 submitted models across 117 comparisons and 26 receptors.

State-term interventionAll pairs ↑Same composition (36) ↑Different composition (81) ↑
Full Qbind0.61950.55240.6530
Removed / fixed0.50350.50000.4716
Isolated state term0.62050.55240.6534
Permuted row0.52450.53430.5197

We observed direction accuracy of 0.6195 with the full model versus 0.5035 after removal. The isolated state term accounted for this signal. These are inference-time interventions, separate from training a state-only baseline. The receptor-level interval for the full-minus-removed difference includes zero.

2. Explicit binder-by-state models

“Did the authors test an explicit binder × state interaction model ...?”

Yes. We trained the requested interaction models on identical receptor-cold folds.

ReadoutDirection macro ↑Within-transition ↑Exact-Δ MAE ↓Screening regret ↓
Additive0.6110.6090.9930.0397
Rank-16 interaction0.5380.5490.9910.1221
Attention0.4360.6141.0250.2024
Frozen-base + rank-80.6110.6020.9930.0397

We additionally tested bilinear, low-rank, contrastive, and ridge formulations with and without potency. The new readouts matched or trailed the state-free control on pooled ordering. The available 215 orderings span 13 groups, all with composition changes, and 147 depend on affinity bounds. We retain the additive readout for the present data and will pursue interaction learning with additional paired-state measurements.

3. End-to-end binder and dose selection

“Can the authors provide an end-to-end evaluation in which both the binder and the dose are selected using predicted ... values ...?”

We fixed both decisions from out-of-fold predictions before scoring with measured affinities. For desired-state maximization, all 28 candidates had the correct predicted affinity direction. Because the optimal dose is zero or the ceiling, prediction-based dose selection preserved the submitted five-model screening regret of 0.0397.

We also evaluated 3,956 separating target-matching tasks. Lower is better throughout.

PolicyPredicted-dose errorPredicted-dose regretMeasured-dose errorMeasured-dose regret
Affinity-only0.24160.08990.19900.0474
Finite-dose Qbind0.24180.09010.18930.0376

The prediction-based policies achieved nearly identical errors. When we optimized each nominated binder's dose using measured affinities, finite-dose selection reduced regret by 20.7%. We now quantify binder-selection and dose-estimation effects separately under the equilibrium model.

4. Structural interpretation and experimental validation

“Why was no MD or other dynamic analysis performed for the structural case study?”

We designed Figure 5 to illustrate the affinity-to-population calculation and use the static structures as geometric context. For the illustrated peptide, predicted affinities are approximately 50 and 20 μM for the free-toxin and HAP-bound-derived receptor frames, with HAP removed before prediction. At an illustrative baseline population of 0.10, we calculate populations of 0.12 at 10 μM and 0.18 at 100 μM free peptide under the equilibrium model.

For direct dynamic validation, we are preparing NRAS-binding macrocyclic peptides and prospective 15N NMR measurements of GTP-loaded NRAS with a collaborating team. These measurements will test the predicted population response experimentally, and we'll be sure to put this in the manuscript upon acceptance.

Overall, we greatly appreciate the reviewer's concrete suggestions and would be grateful if the reviewer would consider raising the score and supporting acceptance based on our finite-dose decision analysis and expanded evaluation.

Thank you for the thorough rebuttal and additional analyses. However, the results largely confirm my original concerns. In matched-context pools, predicted state-affinity differences show essentially no binder-dependent variation, and direction prediction is driven mainly by receptor-state features. More importantly, when both binder and dose are selected from predicted affinities, finite-dose Qbind does not outperform affinity-only selection; the gain appears only with measured-affinity dose optimization. The proposed NMR experiments are promising but do not constitute current validation. I encourage the authors to narrow their claims and report these limitations prominently. I will maintain my score.

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

The paper introduces AlloShift, a framework for designing and selecting protein binders that shift conformational state populations under finite-dose constraints. The authors formulate a two-state equilibrium model that combines state-specific binding affinities, baseline populations, and free binder concentration to optimize the desired state population or match a target population. They also propose Qbind, a model that predicts state-resolved binding affinities from peptide features, biochemical context, and receptor structures using sparse and censored affinity measurements. Experiments are conducted on an affinity corpus containing 2,919 measurements, 117 paired state comparisons, and three measured-affinity panels covering 28 binders. The authors show that affinity-only screening selects a suboptimal binder in 70.4% of 513 separating tasks, while the proposed finite-dose objective achieves optimal selection under the assumed equilibrium model. Qbind also improves state-direction prediction and candidate screening compared to its learned baselines. Finally, the authors demonstrate that Qbind-guided peptide generation and reranking improve model-evaluated direction success from 0.500 to 0.613 using PepTune.

Critical strengths
  • The paper addresses an important problem in protein binder design by considering both binding affinity and conformational selectivity under finite-dose constraints.
  • The finite-dose population formulation is well motivated and provides a clear mathematical framework for selecting binders and concentrations to achieve desired conformational populations.
  • The authors provide detailed analyses of how dose constraints, binding potency, and selectivity affect candidate selection, including robustness analyses under affinity prediction errors.
  • The work integrates affinity prediction with peptide generation and reranking, demonstrating the potential utility of Qbind for candidate prioritization.
Critical weaknesses
  • I am not sure how effectively Qbind can learn binder-specific conformational selectivity. Its additive architecture makes the predicted affinity difference independent of binder identity when the measurement contexts match. Moreover, the state-direction accuracy improves only slightly over the state-only baseline. Although the authors acknowledge this limitation, further analysis of how it affects binder-specific ranking would make the approach more convincing.
  • The finite-dose optimization achieves exact results when using measured affinities, but its practical effectiveness depends on the accuracy of the predicted affinities. Although the authors provide additional analyses using prediction errors in Appendix E, the reported reachability accuracy drops to 0.633 on separating tasks when replaying actual residuals. A clearer end-to-end evaluation of binder and dose selection using predicted affinities would better demonstrate the practical utility of the approach.
  • The peptide generation results rely on learned evaluators, and the improvements become smaller under the independent evaluator (0.500 to 0.557, compared to 0.500 to 0.613 under the primary evaluator). Although the independent evaluation supports the proposed approach, further analysis of the robustness of these improvements across different targets and evaluators would strengthen the results.
Questions for authors
  • Given the additive architecture of Qbind (Eq. 5), the predicted state-affinity difference becomes independent of binder identity when the biochemical contexts match. The authors also report that the predicted state differences are identical across 214 peptide candidates in Appendix I. What drives the improvements in Qbind-guided generation and binder ranking if the model cannot distinguish binder-specific conformational selectivity in this setting? Does this limitation affect other target pairs as well?
  • In Table 3, predicted affinities are used for binder selection, but measured affinities are used to determine the optimal dose. Although Appendix E provides additional analyses of prediction errors, I am curious how the method performs when both binder selection and dose estimation rely on predicted affinities. Could the authors report the resulting population error, evaluated using measured affinities, particularly for target-population matching?
  • The finite-dose formulation builds on established thermodynamic equilibrium equations. While I understand the motivation for jointly optimizing binder selection and concentration, I am not sure how much methodological novelty this formulation provides beyond applying these equations to binder selection. Could the authors clarify the main methodological contribution and discuss the practical advantages over simpler affinity- or selectivity-based selection methods?
1 comment

We thank the reviewer for the thoughtful questions and for recognizing the value of considering dose together with affinity and selectivity. We took these suggestions very seriously and performed new interaction-model comparisons, an end-to-end binder-and-dose evaluation, and a generation analysis across targets. Below, we'll explain what drives the improvements and how these additions strengthen the contribution.

1. What drives candidate ranking?

“What drives the improvements in Qbind-guided generation and binder ranking ...?”

We traced the matched-context rankings to predicted potency within a receptor-state pair. Across the 214 Appendix I candidates, predicted state-affinity differences varied by at most 8.6 x 10⁻⁸. All 33 matched-context pools tied at reporting precision. Ranking by either state's affinity selected the same candidate in all 72 pools. We will explain this attribution directly so readers can distinguish potency-based prioritization from binder-specific state preference.

To test an alternative, we trained explicit binder-by-state readouts on identical receptor-cold folds, with 25 models per arm.

ReadoutDirection macro ↑Exact-Δ MAE ↓Screening regret ↓
Additive Qbind0.6110.9930.0397
State-only0.610N/AN/A
Rank-16 interaction0.5380.9910.1221
Attention0.4361.0250.2024
No-change predictorN/A0.964N/A

We also tested bilinear, low-rank, contrastive, and ridge variants with and without potency inputs. The new readouts matched or trailed the state-free control on pooled ordering. The 215 available within-transition orderings span 13 groups, all with changing composition, and 147 involve affinity bounds. We retain the additive model for the present screening task and will pursue interaction learning with additional matched-context measurements.

2. Selecting both binder and dose from predictions

“Could the authors report the resulting population error, evaluated using measured affinities ...?”

We completed this evaluation. We fixed both decisions using out-of-fold predictions, then calculated populations from measured affinities under the equilibrium model. On 3,956 separating target-matching tasks, we obtained the following results. “Measured-dose” means optimizing the dose with measured affinities after fixing the nominated binder. Lower is better throughout.

PolicyPredicted-dose errorPredicted-dose regretMeasured-dose errorMeasured-dose regret
Random0.24780.09610.22100.0693
Affinity-only0.24160.08990.19900.0474
Finite-dose Qbind0.24180.09010.18930.0376

At predicted doses, the two selection policies achieved nearly identical errors. With measured-dose optimization, finite-dose selection reduced regret by 20.7% relative to affinity-only selection. We now separate binder-selection quality from concentration-estimation error.

For the submitted desired-state maximization screen, the optimal dose is zero or the ceiling. Predicted affinity directions were correct for all 28 candidates, so prediction-based dose selection preserved the reported five-model screening regret of 0.0397.

3. Generation across targets and evaluators

We recomputed results across 24 targets, two directions, and three seeds. Gains are evaluator-assigned improvements, with reported intervals in brackets.

ComparisonPrimary evaluatorSecond evaluator
Overall gain+0.113 [0.047, 0.186]+0.058 [0.001, 0.119]
Different-composition targets (16)+0.169 [0.078, 0.263]+0.086 [0.002, 0.168]
Guidance added to reranking+0.044 [0.016, 0.076]+0.028 [0.006, 0.053]

We obtained positive gains with both reporting ensembles, which share a model family and were used during development. We describe these findings as computational candidate enrichment.

4. Methodological contribution

“Could the authors clarify the main methodological contribution ...?”

We develop joint binder-and-concentration selection for a requested equilibrium population, including reachable populations and dose-dependent decision regimes. We build on classical linkage (Wyman, 1964 and Monod et al., 1965). Across three measured-affinity panels, affinity-only screening missed the optimum in 70.4% of 513 separating decisions. This establishes why concentration-dependent selection matters.

We are grateful to the reviewer for helping us strengthen the evaluation. Having completed the requested interaction, binder-and-dose, and cross-target analyses, we would be very appreciative if the reviewer would increase the score and support acceptance of AlloShift!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We thank the reviewers for the thoughtful questions about AlloShift. During rebuttal, we ran new experiments to answer three questions: How well can we choose a binder and dose from predicted affinities? How does our predictor distinguish protein states? How much do we gain by guiding peptide generation?

Importance of binder and dose selection

When we design a binder to favor one protein state, we also need to know how much binder it will take to reach the desired population. AlloShift uses the affinity for each state and the maximum concentration we can use to select a binder and dose together. Across 28 binders in three panels with measured affinities, affinity ranking missed the optimal binder in 70.4% of the 513 tasks where the selection rules differed. At a dose ceiling of 0.1 μM, affinity ranking selected the optimal binder. At 100 μM, selectivity ranking did. The concentration we can use changes which binder we should choose.

New experiments for each reviewer

  • Reviewer uyAQ asked how well we can select a binder and dose when we predict both affinities. We tested the full workflow on 3,956 target-matching tasks where the selection rules differed. We predicted both the binder and its dose, then calculated the resulting population using measured affinities. AlloShift and affinity ranking achieved similar mean population errors (0.2418 and 0.2416). We then tested a workflow we could use in the lab. We kept the binder each method nominated and used its measured affinities to choose the dose. AlloShift reduced mean regret by 20.7%, from 0.0474 to 0.0376. In practical terms, we nominated binders that came closer to the desired population once we set the dose from measurements.

  • Reviewer pb63 asked how Qbind uses protein state to rank peptides. We compared Qbind with two models that explicitly learn interactions between a peptide and a protein state. Qbind predicted the direction of state preference with 61.1% macro-averaged accuracy across receptors, compared with 53.8% for the rank-16 interaction model and 43.6% for the attention model. A model using state alone reached 61.0%. We also checked whether changing the state changed the top-ranked peptide. Ranking by predicted affinity for either state selected the same peptide in all 72 matched pools. These experiments show that Qbind learns the direction of preference largely from the receptor state and ranks peptides within each pool primarily by predicted potency.

  • Reviewer aieH asked what AlloShift contributes beyond established thermodynamics and how we evaluated generation. We use thermodynamic linkage to calculate which populations a binder can reach at a feasible dose. Our contribution is to use that calculation both to choose among binders and to guide the design of new peptides. We tested generation on 24 targets in both shift directions across three seeds. Compared directly with reranking generated peptides, guiding generation improved the mean score by 0.044 and 0.028 for the two directions on our development evaluators. We have also begun prospective work on NRAS macrocycles to test the predicted population shifts by NMR.

Case for acceptance

We believe AlloShift deserves acceptance because it answers a practical question in protein design: which binder should we advance, and at what dose can it reach the population we want? With measured affinities, we showed how often affinity ranking misses the best choice. With predicted affinities, we showed that we can nominate better candidates for experimental follow-up when we use measurements to set the final dose. We appreciate the reviewers pushing us to test the method under both conditions, and we respectfully ask the Area Chair and reviewers to support acceptance of AlloShift.

Thank you, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.