acceptodds
Under review as a conference paper at ICLR 2027

From Regression to Dose–Response: Predicting GPCR Activity and EC50 via Concentration-Conditional Modelling

Abstract

ML models have revolutionised drug discovery, yet they still struggle with modelling the molecule-induced response of G-protein coupled receptors (GPCRs) — cellular membrane "sensors" that account for nearly % targets of approved drugs. The central challenge in GPCRs lies in predicting activity and potency (i.e., half maximal effective concentration, EC50). To estimate these properties, we propose a unified learning framework inspired by *in vitro* dose-response assays that models concentration-conditional activation. Specifically, for each training protein-molecule pair , we uniformly sample concentrations and generate labels that are deterministic far from the EC50 and soft in the transition region. We then train a concentration-conditional model that learns the probability at any . Finally, querying the model across concentrations enables fitting a Hill curve from which both activity (upper asymptote) and EC50 (inflection point in ) are derived. On the challenging M2OR dataset, our approach achieves an EC50 estimation error of log units, an improvement of log units compared to drug–target affinity baselines, log units compared to censored regression, and log units compared to the Boltz-2 affinity module, while achieving comparable performance in activity classification to the SOTA baselines (MCC in the i.i.d. case).

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein-molecule interaction, Dose–response modelling, EC50 estimation, Activity prediction, GPCRs, Olfactory receptors (ORs), Protein language models, Graph neural networks

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_U3jyOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Instead of directly predicting the EC50 (regression task), this paper casts it to a classification problem by generating the classification label by comparing the sampled concentrations against the experimentally measured EC50. In the inference time, the predicted EC50 can be obtained by fitting a logistic curve on the predicted probability w.r.t many sampled concentrations. The paper evaluated on unseen compounds, proteins, compound cluster, and protein cluster for both the classification task performance and regression task performance (MCC).

Critical strengths

It casted a regression problem to a classification problem such that it can perform both activity classification and potency prediction and obtained better performance in comparison to 4 potency regression baseline and 5 activity classification baselines and ablation studies clarify what does not drive the performance gains.

Critical weaknesses

• The paper evaluates the regression performance (RMSLE) on the intersection of the correctly predicted active pairs of ASMI-DR and the DTA baselines. I assume the reason for this constraint comparison is that ASMI-DR can only estimate EC50 when its predicted response curve crosses certain probabilities, so pairs predicted as inactive have no potency value. However, this excludes false negatives from RMSLE and likely favours easier pairs. Additionally, this is a limitation of the method such that the other DTA baselines do not have. • As the benefit of the approach is that it can perform both classification and regression task, comparing against approaches such as Nesso-1 that can also perform both activity classification and affinity prediction would be helpful. Additionally, the proposed method seems to perform similar or worse than HyperAttnDTI on all the four generalization tasks.

Questions for authors

• Sensitivity analysis between the performance and the sampling procedure. There are a few sampling decisions: number of sampled classification pairs, number and range of the sampled concentrations, sampled decisions when the concentration is close to the ground truth EC50. Understanding the sensitivity of these parameters to the performance is critical to justify the robustness of the model. • Figure 4 is a bit confusing. What do the error bars around the baseline methods represent? Evaluating performance across multiple folds produces a standard deviation, but selecting different thresholds also changes the number of interactions among the correctly predicted active pairs for ASMI-DR and the DTA baselines, which could produce a standard deviation as well. A clarification would be helpful.

5 comments

We thank the reviewer for their constructive feedback and for recognising the advantages of jointly predicting activity and potency. We have expanded the evaluation to address the concerns about the selection of active pairs, added the Nesso-1 comparison, and clarified the sensitivity of our approach to sampling hyperparameters. We address each concern below. The changes in the revised pdf are highlighted in blue.

W1 & Q2: EC50 evaluation, false negatives, and recall–RMSLE trade-off

We agree with the reviewer that evaluating RMSLE on the intersection of correctly predicted active pairs reduces the evaluation set and may select easier interactions. However, this evaluation is intentional: it isolates EC50 estimation accuracy from activity classification by comparing all models on the same successfully identified active pairs. False negatives are not ignored, but accounted for separately through MCC and recall.

We would also like to emphasise that we consider the absence of a reported EC50 estimate for pairs classified as inactive by ASMI-DR to be a desirable property of our formulation rather than its limitation. Experimentally, EC50 is only defined when a measurable response is observed and a dose-response curve can be fitted. Although regression baselines produce a numerical EC50 for every pair, this does not necessarily imply that their estimates for pairs classified as inactive are informative.

In fact, for the regression baselines, the activity decision is based directly on the estimated EC50, with predictions above 10−1 M10^{-1}\,\mathrm{M} classified as inactive. Since all experimentally measured EC50 values in our test sets are at most 10−2 M10^{-2}\,\mathrm{M}, every false-negative prediction necessarily has an EC50 error exceeding one log unit by construction. Evaluating these predictions together with correctly identified active pairs would therefore conflate activity-classification failures with the accuracy of potency estimates.

To investigate whether these numerical predictions nevertheless retain meaningful potency information, we additionally evaluated the regression baselines separately on their true positives (TP) and false negatives (FN):

Table 1: EC50 estimation on experimentally active pairs in the i.i.d. setting, separated by each model’s activity decision. Counts are averaged over five splits. ASMI-DR does not report potency estimates for predicted inactive pairs.

ModelnRMSLErhonRMSLErho
True positivesFalse negatives
DTIAM212.01.5160.40836.23.5210.024
DeepDTAGen187.61.4170.33860.63.8520.159
ProSmith208.81.4020.42539.43.6470.066
ASMI-Reg180.41.4640.38967.84.0350.014
ASMI-DR178.20.7080.65070.0——

While the regression baselines achieve Spearman correlations of 0.34–0.43 on correctly identified active pairs, these correlations decrease to 0.01–0.16 on false negatives, accompanied by RMSLE values of 3.52–4.04. The same pattern is observed across all four OOD scenarios, where Spearman correlations on false negatives range from −0.12-0.12 to 0.160.16. These results suggest that although the regression baselines produce numerical estimates for false-negative pairs, these predictions provide little useful information about their relative potency.

To further address the reviewer’s concern about the reduced evaluation set, we have substantially extended Figure 4 to evaluate the recall-RMSLE trade-off across varying activity thresholds for all models, rather than only ASMI-DR. We vary the decision threshold of ASMI-DR over τ∈[0.02,0.8]\tau \in [0.02, 0.8] and the regression baselines over T∈[10−3,10−0.5] MT \in [10^{-3}, 10^{-0.5}]\,\mathrm{M}. Each model is evaluated on its own true-positive predictions at each threshold.

The revised Figure 4 shows that ASMI-DR achieves lower EC50 error than the regression baselines across their shared recall ranges in all five evaluation scenarios. In contrast to regression models, whose RMSLE generally increases substantially as recall increases, ASMI-DR maintains relatively stable potency estimation accuracy. Thus, the advantage of ASMI-DR is not simply a consequence of evaluating fewer, easier active pairs.

We thank the reviewer for pointing out the ambiguity in the original Figure 4, which helped us to greatly enhance the clarity for the reader. The uncertainty bands now represent one standard deviation of RMSLE across the five data splits, evaluated independently at each threshold. Changing the threshold changes the selected true-positive pairs and consequently the recall shown on the horizontal axis. The revised caption also explicitly defines NactiveN_{\mathrm{active}} as the mean number of experimentally active test pairs per split.

W2: Comparison with Nesso-1 and activity classification performance

We thank the reviewer for suggesting Nesso-1 and have added it to our experiments (Figure 3 and Tables J2–J4).

Similarly to Boltz-2, Nesso-1 predicts protein-ligand binding likelihood and binding affinity, whereas our primary objective is to predict functional receptor activation and its potency. This distinction is particularly important for GPCRs, where ligand binding does not necessarily induce receptor activation. A binder may instead act as an inverse agonist or produce no measurable functional response.

Consistent with this distinction, Nesso-1 achieves very low activity recall (0.033–0.075 across the evaluated settings), identifying only a small fraction of experimentally active pairs. Consequently, its reported potency errors are evaluated on active pairs identified by other models, most of which Nesso-1 itself classifies as inactive. ASMI-DR achieves lower RMSLE in four of five scenarios, with statistically significant improvements in two. The exception is Sequence-Cluster, where the difference is not statistically significant. These results highlight the difficulty of directly transferring binding affinity predictions to functional GPCR activation.

Regarding HyperAttentionDTI, we agree that ASMI-DR does not consistently outperform dedicated activity-classification models. However, our principal contribution is the substantial improvement in EC50 estimation while retaining competitive activity prediction, rather than establishing a new state of the art in activity classification alone, which we now emphasise more in the revised version.

In the i.i.d. setting, ASMI-DR achieves MCC 0.665 compared with 0.656 for HyperAttentionDTI. The differences are also small in three of the four OOD settings: Sequence-Single (0.481 vs. 0.480), Sequence-Cluster (0.218 vs. 0.232), and Molecule-Single (0.593 vs. 0.608). We acknowledge a larger difference in Molecule-Cluster (0.398 vs. 0.482), where HyperAttentionDTI performs better.

Unlike HyperAttentionDTI, however, ASMI-DR simultaneously provides an EC50 estimate for predicted active pairs, substantially outperforming the dedicated regression baselines in EC50 estimation. We therefore view the competitive classification performance, combined with the substantial improvements in EC50 prediction, as the main practical advantage of the proposed framework.

Q1: Sensitivity to concentration sampling and hyperparameters

We thank the reviewer for highlighting the importance of sensitivity to the sampling procedure. The original Appendix F already included several relevant ablations, and we have now made these analyses more explicit in Section 6.3 in the revised text.

We evaluated the sensitivity to the main sampling choices as follows:

  • Transition margins (Table F1): Varying ϵL=ϵU=ϵ\epsilon_L = \epsilon_U = \epsilon over {0.1,0.25,0.5,1.0}\{0.1, 0.25, 0.5, 1.0\} yields stable performance, with i.i.d. RMSLE between 0.708 and 0.755 and MCC between 0.652 and 0.673.
  • Soft-label generation (Table F2): We compare uniformly sampled soft labels with linear interpolation across the transition region. Both retain substantially better EC50 estimation than the DTA regression baselines.
  • Concentration sampling (Table F3): We compare uniform sampling with a strategy that preferentially samples concentrations near the measured EC50. Both yield comparable performance, indicating that the improvement does not depend on concentrating training samples around the transition region.
  • Sample weighting (Table F4): We additionally compare dynamic concentration-dependent class weights with standard class-imbalance weights, finding that the formulation remains effective under both strategies.

Regarding the number of concentrations used at training, this is adapted by the best checkpoint selection. During training, one concentration is sampled per protein-molecule pair in each epoch, with a new concentration drawn in subsequent epochs. Thus, the number of sampled concentrations per pair over the full training is given by the index of the best training epoch.

As for the number of concentrations used at inference, Appendix G evaluates this trade-off explicitly. Reducing the number of queried concentrations from 62 to only 7 changes RMSLE by less than 0.015, substantially reducing inference cost without materially affecting performance.

Finally, the training concentration range is not a free hyperparameter, and it is determined by the concentrations tested in the underlying experimental assays.

Taken together, these experiments demonstrate that the improvements in EC50 estimation are robust to the main choices introduced by the concentration-conditional formulation, supporting our conclusion that the gain primarily originates from the learning formulation rather than a particular sampling configuration.

Conclusion

We hope that these additional results and clarifications address the reviewer’s concerns about the evaluation protocol, the comparison with joint activity-affinity models, and the robustness of the proposed approach.

Reviewer_aP4hOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3(+1)
Presentation
2(+1)
Soundness
3(+1)
Read review
Paper summary

This paper proposes a novel framework for GPCRs activity and potency prediction. In detail, the proposed method is based on the biological prior in vitro dose-response assays that concentration-conditional activation would accurately capture the true activity of the GPCR. Extensive experiments are conducted to prove the effectiveness of the proposed method.

Critical strengths
  1. The proposed method incorporates biomedical prior knowledge into the model through supervision signals. This is a promising approach that enhances the interpretability of the model.
Critical weaknesses
  1. The core of this work is built around EC50EC_{50}. From a methodological perspective, the proposed approach appears to employ a cross-attention network as the backbone of a Bayesian framework, while manually incorporating EC50EC_{50}-related information as an additional conditioning signal. However, it remains unclear whether the reported performance improvements can genuinely be attributed to the incorporation of EC50EC_{50} information. After tokenization, there is no guarantee that the model actually captures the underlying biological meaning of EC50EC_{50}. Instead, the model may simply learn to predict the final outputs through a regression mechanism. This raises concerns about whether the proposed framework truly achieves the mechanism-informed learning claimed in the abstract.
  2. The authors introduce numerous predefined parameters in the methodology, such as ClowC_{low}, ChighC_{high}, and ϵ\epsilon. A major concern with this may substantially compromise the generalization capability of the model.
  3. The novelty of the proposed method is limited. Following the previous weakness 1, although the introduce of EC50EC_{50} is interesting, the proposed method simply incorporates EC50EC_{50} information into the model as an additional input. From a methodological perspective, the framework does not introduce any specific architectural adaptations to account for the underlying biological mechanisms, potentially resulting in a substantial loss of biologically meaningful information.
  4. From the model's performance, the concerns raised from weakness 3 also reflected in the experimental results, as the proposed method didn't achieve substantial improvements compared with other baselines.
  5. The presentation of the paper is limited. The writing lacks logical coherence in several parts of the paper. While I understand that the authors may have used AI tools for language polishing, the logical flow between sentences and paragraphs is often unclear and inconsistent, making the manuscript difficult to follow. Also some symbols in the equations are not explained, i.e., in line 149, what is i and j represented for?
Questions for authors
  1. The authors should provide more discussions regarding the predifined hype-parameters, especially explain whether these design will affect the model's generalization ability or not.
  2. The authors need to highlight their contributions. From the current version, the advantage of the use of EC50EC_{50} is not clear and the way of incooprating EC50EC_{50} to the model is not reasonable.
  3. The representation of the paper need to be improved. The methodology section should follow a clear and coherent logical structure, where each component is progressively introduced with a well-defined motivation, ultimately forming a unified framework. However, the current presentation reads more like a collection of individual ideas presented sequentially, without sufficient consideration of the logical connections between them. Rather than systematically explaining how and why each component is designed and how they work together, the authors simply stack different methodological components without establishing a coherent narrative. The authors are encouraged to substantially reorganize this section to improve its logical flow and clearly justify the motivation behind each design choice.
5 comments

We thank the reviewer for their constructive feedback and for recognising the value of incorporating biological prior knowledge through supervision. We appreciate the concerns regarding the methodological contribution, the role of EC50, the robustness to predefined hyperparameters, and the clarity of the presentation. We have clarified the learning formulation and expanded the experimental analysis to address these points. The changes in the revised pdf are highlighted in blue.

W1, W3 & Q2: Role of EC50 and methodological contribution

We would first like to clarify an important distinction that may not have been sufficiently clear in the original manuscript: EC50 is not an input to our model, nor is it tokenised or embedded by the network. The model receives only the protein sequence ss, molecular graph mm, and a queried concentration cc. Experimentally measured EC50 values are used exclusively during training to construct concentration-dependent supervision. At inference, no EC50 information is provided. Instead, the model is queried at multiple concentrations, and predicted EC50 is recovered from the inflection point of the fitted Hill curve (Section 4, Algorithms 1-2). Ultimately, our goal is to estimate EC50 for previously unseen protein-molecule pairs.

Therefore, our contribution is not to provide EC50 as an additional conditioning signal to a cross-attention network, but rather to reformulate EC50 regression as a concentration-conditioned classification task. For each protein-molecule pair, we sample concentrations and generate labels according to their position relative to the measured EC50. This allows us to exploit the concentration ordering implied by experimental dose-response assays, while naturally incorporating inactive and censored observations without assigning them arbitrary finite potency values.

The biological prior is thus incorporated primarily through the supervision, rather than through an explicit mechanistic model of receptor activation. The architecture serves as a flexible function approximator for the concentration-dependent response. Specifically, the scalar concentration is mapped to a learned embedding and incorporated into cross-attention queries, allowing protein-molecule representations to vary with concentration (Section 5). We do not claim that this mechanism explicitly reproduces receptor conformational changes or intracellular signalling.

We also acknowledge the connection to ordinal regression and regression-as-classification, now discussed explicitly in Section 2. To isolate the contribution of the task formulation, we have introduced ASMI-Ord, a discretised cumulative-link variant of our approach using the same backbone and comparable treatment of inactive and censored observations (Section D.6). ASMI-Ord improves over conventional regression approaches, while ASMI-DR further increases the performance and allows concentration to be treated as a continuous input, which enables predictions at arbitrary concentrations rather than a fixed set of thresholds.

Finally, we distinguish the learned concentration-conditioned score from an exact reconstruction of biological response amplitudes. The model is trained from EC50-derived supervision rather than complete experimental curves, and its output should primarily be interpreted as a surrogate response score (Section 7). Nevertheless, Appendix E.2 compares the predictions with independent normalised experimental dose-response measurements and reports a Brier score of 0.08, providing additional evidence that the learned scores capture meaningful aspects of the experimental dose-response curves.

W4: Performance improvements and source of the gains

We agree that improvements in activity classification alone are not consistently substantial. However, our principal objective is to improve EC50 estimation while retaining competitive activity classification, rather than outperforming dedicated binary activity predictors on every evaluation split.

In the i.i.d. setting, ASMI-DR achieves an RMSLE of 0.683, compared with 1.082 for the strongest DTA regression baseline (ProSmith), an improvement of 0.399 log units. This advantage persists across all four OOD scenarios (Figure 3, Figure 4 and Tables J2-J4). At the same time, ASMI-DR achieves MCC 0.665 in the i.i.d. setting, compared with 0.656 for HyperAttentionDTI. We acknowledge that dedicated classification models perform better in some OOD scenarios, particularly Molecule-Cluster, but they do not provide corresponding EC50 predictions.

More importantly, the same-backbone ablations distinguish the contribution of the learning formulation from that of the architecture. In the i.i.d. setting, ASMI-Reg, trained with direct regression, achieves RMSLE 1.335. Explicitly accounting for censored observations with ASMI-Cen reduces this to 0.888. Replacing regression with the discretised threshold formulation (ASMI-Ord) further reduces RMSLE to 0.791, while continuous concentration conditioning (ASMI-DR) achieves 0.683.

These results show that the improvement cannot be attributed to the ASMI backbone alone, nor solely to the treatment of censored observations. Rather, the concentration-threshold learning formulation is the principal source of the improvement, with continuous concentration conditioning providing an additional benefit.

W2 & Q1: Predefined hyperparameters and generalisation

We thank the reviewer for raising the question of robustness to predefined hyperparameters. We have clarified their roles and expanded the discussion of the sensitivity analyses in Section 6.3 and Appendix F.1.

First, the concentration bounds ClowC_{\mathrm{low}} and ChighC_{\mathrm{high}} are determined by the underlying experimental assays rather than treated as tunable hyperparameters. The margins ϵL\epsilon_L and ϵU\epsilon_U, on the other hand, control the width of the transition region around the measured EC50, where soft labels are sampled.

To evaluate their influence, we vary the symmetric margins ϵL=ϵU=ϵ\epsilon_L=\epsilon_U=\epsilon over {0.1,0.25,0.5,1.0}\{0.1,0.25,0.5,1.0\} (Table F1). The resulting i.i.d. RMSLE remains between 0.708 and 0.755, while MCC varies between 0.652 and 0.673, indicating that performance is relatively insensitive to the precise choice of margins within the evaluated range.

We additionally evaluate alternative soft-label distributions in the transition region (Table F2), concentration sampling strategies that preferentially sample near EC50 (Table F3), and alternative class-weighting schemes (Table F4). The results remain broadly comparable across these choices, with all evaluated variants retaining substantial improvements in EC50 estimation over the DTA regression baselines.

Finally, Appendix G examines the number of concentration queries used at inference. Reducing the number of queried concentrations from 62 to only 7 changes RMSLE by less than 0.015, demonstrating that the inference procedure is not sensitive to a finely discretised concentration grid.

Together with the improvements observed across the four OOD evaluation scenarios, these results support the robustness of the proposed formulation.

W5 & Q3: Presentation and methodological clarity

We thank the reviewer for the detailed suggestions regarding the presentation. We recognise that some wording in the original manuscript could have obscured the distinction between the training supervision and the model inputs, and we have revised Sections 4 and 5 to make this distinction explicit.

In particular, Section 4 now states that measured EC50 values are available only during training, while the model inputs consist exclusively of the protein sequence, molecular graph, and concentration. We have also clarified the mathematical notation: ii indexes protein–molecule pairs, and jj indexes sampled concentrations for each pair. In practice, the training data are generated incrementally by sampling one concentration per pair at each epoch.

We have further clarified the relationship between the different methodological components. The motivation follows from the limitations of direct regression on datasets containing inactive and censored observations. These measurements are converted into concentration-dependent supervision, which is used to train a conditional response model. At inference, the model is queried across concentrations, and the fitted response curve provides both an activity decision and an EC50 estimate.

Section 5 now distinguishes the architecture-agnostic learning formulation from its specific implementation in ASMI-DR, and explicitly explains that the biological dose-response prior enters through supervision rather than through the cross-attention mechanism itself. We have also expanded Sections 6.2-6.3 to connect the experimental results directly to the respective design choices, including the distinction between improvements due to censoring, threshold-based learning, and continuous concentration conditioning.

Conclusion

We hope these revisions and additional experiments clarify the methodological contribution, its relationship to established regression and classification approaches, and the evidence supporting the reported improvements.

Reviewer_d9KpOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a concentration-conditioned framework for predicting GPCR activity and EC₅₀. Instead of directly regressing EC₅₀, it samples concentrations and constructs classification labels from measured EC₅₀ values and activity annotations. At inference, the model is queried across concentrations. A Hill curve is then fitted to obtain an activity decision and an EC₅₀ estimate.

Critical strengths

The framework connects the learning task to the experimental dose–response workflow. It provides a unified way to use measured EC₅₀ values, inactive observations, and lower-bound observations. In particular, it avoids assigning arbitrary finite EC₅₀ values to inactive pairs.

Critical weaknesses

The core supervision is closely related to threshold classification. Away from the transition region, the generated label indicates whether the queried concentration exceeds the measured EC₅₀. Related ideas appear in ordinal regression and regression-as-classification.

The training data do not contain complete experimental response curves. The concentration labels are constructed from EC₅₀ values using predefined rules. Therefore, the results establish improved EC₅₀ prediction more directly than recovery of actual biological responses.

The paper should explicitly state the complete training objective, including the response loss, dynamic sample weights, and the weight of the auxiliary MLM objective. It should also clarify whether concentration monotonicity is enforced or only encouraged by the generated labels, and how failed or unstable Hill fits are handled.

Questions for authors

How does the method compare with a standard threshold-classification or ordinal-regression model using the same backbone and comparable treatment of inactive observations?

Does the predicted concentration-dependent score correspond to exceeding half-maximal response, or to observing any measurable activation? These interpretations are different.

Can the model be evaluated against independent, experimentally measured dose–response curves?

How often are raw predictions non-monotonic, and how often does Hill fitting fail or produce unstable parameters?

3 comments

We thank the reviewer for their thoughtful assessment and for recognising the advantages of our framework in jointly modelling activity and potency while accommodating inactive and censored observations. We particularly appreciate the suggestions concerning the connection to ordinal regression, the interpretation of the predicted scores, and the stability of the inference procedure. We have addressed these points in the revised manuscript, as detailed below. The changes in the revised pdf are highlighted in blue.

W1 & Q1: Connection to ordinal regression and threshold classification

We agree with the reviewer that our formulation is closely related to threshold classification and ordinal regression. We have made this connection explicit in Section 2, where we interpret concentration-conditioned learning as a continuous-threshold formulation of cumulative-link modelling. In the idealised setting with zero transition margins ϵL\epsilon_L and ϵU\epsilon_U, the supervision indicates whether the queried concentration exceeds the measured EC50, and the Bayes-optimal predictor under ordinary binary cross-entropy corresponds to the conditional cumulative distribution function (CDF) P(EC50≤c∣s,m)P(EC_{50} \leq c \mid s,m).

The main difference from standard ordinal regression is that ASMI-DR treats concentration as a continuous input rather than predicting probabilities at a fixed set of discrete thresholds. This allows the model to learn a concentration-dependent function that can be queried at arbitrary concentrations. Moreover, concentration directly conditions the protein-molecule interaction layers, rather than being incorporated only into the output head.

To evaluate the importance of these differences, we have introduced ASMI-Ord, an ordinal-regression variant that shares the ASMI backbone but replaces concentration conditioning with a monotone cumulative-link head predicting KK ordered threshold probabilities. It uses comparable supervision and treatment of inactive and censored observations as ASMI-DR (details in Section D.6 in the revised version).

The results (Figure 3, Figure 4 and Tables J2-J4 in the revision) show that ASMI-Ord improves over censored regression (ASMI-Cen) in all five scenarios, reducing RMSLE by 0.097-0.578 log units. This confirms that the threshold-based task reformulation accounts for a substantial part of the improvement. Nevertheless, ASMI-DR achieves lower RMSLE than ASMI-Ord in four of five scenarios and higher activity MCC in all five. In the i.i.d. setting, ASMI-DR achieves RMSLE 0.683 compared with 0.791 for ASMI-Ord and 0.888 for ASMI-Cen.

These results suggest that the main improvement originates from the concentration-threshold learning formulation, while continuous concentration conditioning provides an additional empirical benefit. Beyond predictive accuracy, the continuous formulation also allows inference at arbitrary concentrations, and the number of queried concentrations can be reduced without retraining (Appendix G).

W2 & Q2–Q3: Interpretation of the predicted scores and experimental validation

We agree with the reviewer that exceeding half-maximal response and observing any measurable activation are distinct concepts and our supervision is derived from the former.

More precisely, the training labels are constructed from the comparison between the queried concentration cc and the experimentally measured EC50. In the idealised zero-margin setting, the corresponding probabilistic target is P(EC50≤c∣s,m)P(EC_{50} \leq c \mid s,m), representing the probability that the queried concentration reaches the half-maximal effective concentration. With transition margins ϵL\epsilon_L and ϵU\epsilon_U and class weighting, the learned output is more appropriately interpreted as a smoothed, concentration-conditioned response score rather than an exactly calibrated probability of experimentally observed activation.

The transition margins ϵL\epsilon_L and ϵU\epsilon_U control the region around EC50 in which soft supervision is applied. They also admit a direct interpretation. For a Hill slope qq, the normalised response at c=EC50+ϵUc=EC_{50}+\epsilon_U is

f(EC50+ϵU)=11+10−qϵU.f(EC_{50}+\epsilon_U) = \frac{1}{1+10^{-q\epsilon_U}}.

For example, assuming q=1q=1, our default symmetric margins ϵL=ϵU=ϵ=0.25\epsilon_L=\epsilon_U=\epsilon=0.25 correspond to a transition interval spanning approximately 36-64% of the normalised response. Importantly, the margins primarily control the width of the learned transition rather than its location. Table F1 shows that EC50 estimation remains stable across ϵ∈{0.1,0.25,0.5,1.0}\epsilon \in \{0.1,0.25,0.5,1.0\}.

Regarding experimental validation, we acknowledge that our method is trained from EC50-derived supervision and is not intended to reconstruct complete experimental dose-response curves. Raw measurements are generally unavailable in M2OR, and their amplitudes are not directly comparable across experimental assays. Nevertheless, the framework could in principle be extended to incorporate measured concentration-dependent responses when sufficiently standardised data are available.

We additionally evaluate the relationship between our predictions and independent wet lab measurements in Appendix E.2 (Figures E2-E3). For held-out protein-molecule pairs with available raw dose-response measurements from Mainland et al. (2014), we compare the predicted concentration-dependent scores with normalised experimental responses.

For this analysis, we set ϵL=ϵU=log⁡10(9)\epsilon_L=\epsilon_U=\log_{10}(9), corresponding to the 10-90% response interval under the Hill slope q=1q=1, which was assumed by Mainland et al. in their Hill curve fits. The resulting predictions broadly follow the observed response progression, achieving a Brier score of 0.08. For reference, fitting a Hill curve directly to the experimental measurements achieves 0.01, although this is a substantially easier setting because the fit has direct access to the evaluated observations.

These results provide evidence that the predicted scores capture aspects of the experimental dose-response curves, while also confirming that they should not be interpreted as exact reconstructions of biological response amplitudes. In particular, the fitted asymptotes and slope should not be identified directly with experimentally measured efficacy, basal activity, or cooperativity.

W3 & Q4: Training objective, monotonicity, and Hill-fitting stability

We thank the reviewer for highlighting the need to clarify the training objective and the stability of the inference procedure.

We have now explicitly stated the complete training objective in Section D, eqn. (6). The response loss is binary cross-entropy with concentration-dependent class weights, while the auxiliary masked-language-modelling loss is the cross-entropy over masked amino-acid positions. The class weights are computed dynamically from the label distribution in each batch, as detailed in Appendix B, and the auxiliary loss coefficient is λMLM=1\lambda_{\mathrm{MLM}}=1 (Section D.1).

Regarding monotonicity, ASMI-DR does not explicitly constrain its raw predictions to be monotone in concentration. Instead, monotonicity is encouraged by the generated supervision, which assigns inactive labels below EC50 and positive labels above it, outside the transition region. At inference, fitting a Hill curve imposes a monotone sigmoidal shape on the extracted response curve.

To assess how closely the raw predictions follow this expected behaviour, we have added a monotonicity and fitting-stability analysis in Section F.4 (Table F8). We quantify deviations using the total sum of downward variations DD. For a monotonically non-decreasing response, D=0D=0.

Across the five evaluation scenarios, the median DD is below 0.015, and below 10−410^{-4} in three scenarios, indicating that typical predicted active curves are nearly monotone. Larger deviations occur in the upper tail, particularly in Sequence-Cluster, where the 95th percentile reaches 0.438.

We also explicitly quantify Hill-fitting failures. Across all scenarios, convergence errors occur for at most 0.19% of evaluated pairs, which are classified as inactive. Among pairs predicted active, nearly flat fitted curves with amplitude M^−b^<0.1\hat M-\hat b<0.1, for which the inflection point may be poorly determined, occur in at most 0.17% of cases and are absent in three scenarios.

Finally, to verify that the observed improvements do not depend on imposing the Hill shape during inference, Appendix F.3 compares Hill fitting with three alternative EC50 estimators: direct threshold crossing, isotonic regression, and monotone I-splines. Their resulting RMSLE values differ by less than approximately 0.1 log unit across the evaluated scenarios. This demonstrates that the improvement is robust to the particular post-hoc EC50 extraction procedure, and the Hill curve readout of EC50 is not crucial for the performance.

Conclusion

We hope that the additional ordinal-regression comparison, the clarified probabilistic interpretation, and the explicit analysis of monotonicity and fitting stability address the reviewer’s concerns and further strengthen the methodological justification of the proposed framework.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Minor correction to ASMI-Reg results

During preparation of the revised manuscript, we identified that ASMI-Reg had been evaluated using an incorrect model checkpoint rather than the checkpoint selected during validation. We have corrected the affected results in Section J and Figure 3 and Figure 4 using the correct checkpoint. This correction does not change any conclusion regarding the improvements in EC50 estimation achieved by ASMI-DR.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.