- 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 , 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
- Clear formulation of finite-dose allosteric binder selection.
- Connects affinity, selectivity, and concentration in a principled way.
- Shows simple but important counterexamples where affinity-only ranking fails.
Critical weaknesses
- 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.
- 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
- 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
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 . Without actual measurements of conformational populations, these results do not establish whether there is a real improvement. Therefore I keep my original score.
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
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.
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
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.
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
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.