acceptodds
Under review as a conference paper at ICLR 2027

Evaluating Physical Intuition in Large Language Models for Structure-Guided Mutation Design

Abstract

Structure-guided protein design requires interpreting molecular geometry and chemistry to anticipate mutation effects. Existing LLM evaluations emphasize sequence-based prediction or tool-integrated workflows, leaving unclear whether models themselves can reason directly from three-dimensional structures. Here, we systematically evaluate this capacity, termed physical intuition, through a partner-contrastive benchmark that holds the protein and substitution fixed across different binding partners. Models inspect local coordinates with or without Python, without dedicated mutation-effect predictors. Although overall accuracy remains modest, most evaluated LLMs recover more complete partner contrasts when affinity improvements are larger, and successful judgments also occur without Python. Structural interventions and evidence audits support selective dependence on mutation-relevant geometry and chemistry, with auditable observations largely grounded in the supplied coordinates. Case studies further show that models can identify relevant interactions and propose mutant-state hypotheses, but struggle to anticipate structural rearrangements and weigh their energetic consequences. Together, these experiments provide, to our knowledge, the first systematic evidence that general-purpose LLMs can use three-dimensional structural evidence to reason about partner-dependent mutation effects, without relying on specialized predictors to supply those judgments.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Large Language Models, Structure-Guided Protein Design, Physical Intuition, Partner-Dependent Mutation Effects, Protein–Protein Binding Affinity, Structural Reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_WiVBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether general-purpose large language models (LLMs) can reason about protein mutation effects directly from three-dimensional structural information without specialised predictors. The authors introduce a partner-contrastive benchmark and evaluate multiple LLMs against physical prediction methods. Through structural interventions, evidence audits, and case studies, the study finds that LLMs can utilise relevant structural information and identify plausible molecular interactions, particularly for larger affinity improvements. However, their overall prediction accuracy remains modest, with significant challenges in predicting structural rearrangements and energetic consequences.

Critical strengths

The paper addresses an interesting and underexplored question about LLMs' ability to reason directly from three-dimensional protein structures. The proposed partner-contrastive benchmark provides a useful evaluation framework, while the comparisons with physical predictors and structural ablation experiments offer valuable insights into model behaviour. The mechanistic case studies further strengthen the analysis by highlighting both the capabilities and limitations of LLM-based structural reasoning.

Critical weaknesses

The overall prediction accuracy remains relatively low, raising concerns about the reliability and practical applicability of the evaluated LLMs. Although the structural ablation experiments demonstrate sensitivity to geometric information, they do not conclusively establish genuine physical reasoning capabilities. Additionally, the relatively small benchmark and uneven coverage of protein systems limit the generalisability of the findings. Stronger experimental validation and broader evaluations would help substantiate the paper's central claims.

Questions for authors
  1. Given the relatively low prediction accuracy, how do the authors assess the practical usefulness of LLMs for structure-guided mutation design?

  2. How can the authors further distinguish genuine physical reasoning from pattern recognition or reliance on previously learned structural knowledge?

  3. How well do the findings generalise to unseen protein systems, given the limited benchmark size and uneven distribution of protein families?

  4. Could additional experiments or stronger baselines help establish whether the observed structural sensitivity translates into reliable mutation-effect predictions?

5 comments
AuthorsOpenReview ↗

Response to Weaknesses: Part 1-1

Concern: The overall prediction accuracy remains relatively low, raising concerns about the reliability and practical applicability of the evaluated LLMs. Although the structural ablation experiments demonstrate sensitivity to geometric information, they do not conclusively establish genuine physical reasoning capabilities. Additionally, the relatively small benchmark and uneven coverage of protein systems limit the generalisability of the findings. Stronger experimental validation and broader evaluations would help substantiate the paper's central claims.

A: Thank you for raising these concerns. We have expanded the evaluation to address three questions: whether direct LLM structural judgments can contribute to mutation selection, whether their proposed interaction changes receive independent physical support, and whether the observed performance extends beyond the original benchmark.

1. Practical utility

For structure-guided mutation design, a useful capability is to connect the local molecular environment to a proposed substitution. Our results provide systematic evidence that general-purpose LLMs can use three-dimensional structural information to anticipate mutation-induced interaction changes and make partner-dependent affinity judgments, without specialized predictors supplying those judgments. Three observations support this potential:

First, prediction performance exceeds model-matched random expectations and is comparable to the tested physical predictors. Complete-group accuracy requires both partner-specific predictions to be correct. DeepSeek Flash, GLM-5.3 Flash, and Qwen3.8 Flash achieve 14.29%, 19.05%, and 9.52%, respectively, compared with frequency-matched independent expectations of 8.62%, 13.08%, and 6.63%. They also exceed conditional random expectations that preserve which pairs receive identical or different labels: 7.14%, 13.10%, and 4.76%, respectively (Appendix H).

On Dataset 1, GLM-5.3 Flash achieves 19.05% complete-group accuracy, compared with 16.67% for FoldX and 23.81% for Rosetta ddG. On the added A300 evaluation (300 unpaired interface-mutation cases), its task accuracy is 78.00%, compared with 81.33% for FoldX5 and 80.00% for five-model Rosetta flex ddG. The methods also exhibit different prediction patterns and partially different successful cases, suggesting potential complementarity.

Second, correct predictions are accompanied by better-supported mutation hypotheses. MD support for interaction events accompanying correct versus incorrect predictions is 51.22% versus 34.92% for GLM, 43.36% versus 36.36% for DeepSeek, and 42.86% versus 32.05% for Qwen. The same relationship holds within the larger-improvement subset for all three models. Thus, prediction correctness is associated with independently supported mutation-induced interaction changes (Part 2 below).

Third, larger affinity improvements are associated with stronger prediction and hypothesis-support results. For groups whose enhancing member improves affinity by at least twofold, complete-group accuracy is 33.33% for GLM-5.3 Flash and 23.81% for DeepSeek Flash, compared with 4.76% for each in the smaller-improvement subset; Qwen remains at 9.52% in both. Overall MD hypothesis support increases in all three models: 57.81% versus 33.33% for GLM, 46.00% versus 34.96% for DeepSeek, and 49.40% versus 29.91% for Qwen. These patterns are consistent with larger experimental effects being accompanied by more readily identifiable structural consequences.

Taken together, these observations support structure-informed predictive signal beyond random guessing: successful predictions are associated with physically supported hypotheses, and both prediction performance and hypothesis support vary with effect magnitude. Although accuracy remains modest, the models already perform comparably to the tested physical predictors while exhibiting partly different successes. This combination motivates their use in candidate generation and their potential contribution to improved mutation-prediction and design methods.

AuthorsOpenReview ↗

Response to Weaknesses: Part 1-2

To assess utility more directly, we evaluated candidate generation in six BPTI partner-selective design tasks involving binding to ChT and MT. Across sites T11, K15, R17, and V34, each task requires selecting one of 19 non-WT amino acids to strengthen binding to one partner while weakening binding to the other. Recommendations were evaluated against published DMS-derived binding-effect reference data withheld during inference. Success requires both reference effects to reach at least 0.20 kcal/mol in the requested direction; a reference-best recommendation matches the candidate ranked highest by the predefined selectivity score.

MethodSuccessful recommendationsReference-best recommendations
GLM-5.3 Flash, three rounds5/18 (27.78%)2/18 (11.11%)
FoldX52/6 (33.33%)1/6 (16.67%)
Rosetta flex ddG2/6 (33.33%)0/6 (0.00%)

Across three rounds of six tasks each (18 attempts in total), GLM produced five successful recommendations,including two reference-best recommendations, and succeeds at least once on three of the six tasks. FoldX5 and Rosetta flex ddG each succeed on two of six tasks, with each method providing one final recommendation per task after aggregating five internal calculations. The design results demonstrate that the model can propose substitutions that satisfy a partner-selective objective against withheld experimental reference data. A practical use is therefore to generate candidates for subsequent screening together with explicit hypotheses about which contacts or interactions the substitutions would change. Those hypotheses provide concrete targets for structural inspection, focused calculations, or experimental testing. Part 2 below examines their physical support and its association with prediction correctness.

A further opportunity is to use computational or experimental feedback to refine these proposals. Models could suggest alternative interaction hypotheses, identify calculations that distinguish them, and revise candidate choices in light of the results. Existing physical tools encode accumulated biophysical knowledge and empirical observations through human-designed energy functions, structural representations, and modelling procedures. These formulations remain open to improvement. Beyond proposing individual mutations, LLMs could help discover structural descriptors, context-dependent scoring rules, or calculation strategies that capture effects overlooked by existing formulations. Iteratively testing and refining these ideas against computational and experimental evidence could contribute to a next generation of more accurate tools for mutation prediction and protein design.

AuthorsOpenReview ↗

Response to Weaknesses (continued): Part 2

2. Physical support

We expanded the evaluation to examine two questions: whether predictions depend on relevant structural information, and whether the resulting mutation-induced interaction hypotheses are independently supported.

The structural interventions address the first question. Mutation-site side-chain masking changes approximately 42% of group predictions, compared with approximately 10% after distant masking. Removing the partner reduces group accuracy even though enhancement predictions increase in all three models. Among pairs receiving different predicted labels, correct orientation falls from 96.43% to 59.38% for DeepSeek, from 77.78% to 56.76% for GLM, and from 92.59% to 32.14% for Qwen. These results indicate that relevant structural inputs affect the assignment of partner-dependent mutation effects.

To address the second question, we added paired WT/mutant MD comparisons covering 82 mutation–partner cases. The models provide structured responses separating WT structural observations, proposed mutation-induced interaction changes, the final prediction, and its explanation. We extract the declared interaction changes and compare them with changes in contact frequencies measured from the WT and mutant trajectories. These trajectory measurements are extracted independently of model outputs and experimental affinity labels.

The key result is that interaction hypotheses accompanying correct affinity predictions receive higher MD support than those accompanying incorrect predictions in all three models. This relationship is observed both across Dataset 1 and within the subset of 21 partner groups whose enhancing member improves affinity by at least twofold.

ModelAll groups: correct answersAll groups: incorrect answersLarger-improvement groups: correct answersLarger-improvement groups: incorrect answers
GLM-5.3 Flash42/82 (51.22%)22/63 (34.92%)28/46 (60.87%)9/18 (50.00%)
Qwen3.8 Flash48/112 (42.86%)25/78 (32.05%)32/59 (54.24%)9/24 (37.50%)
DeepSeek Flash49/113 (43.36%)40/110 (36.36%)31/61 (50.82%)15/39 (38.46%)

Entries are MD-supported/scorable interaction events, grouped by the correctness of the answer containing the hypothesis.

Restricting the comparison to the larger-improvement subset therefore preserves the relationship between prediction correctness and hypothesis support. Overall hypothesis support is also higher in the larger- than smaller-improvement groups: 57.81% versus 33.33% for GLM, 49.40% versus 29.91% for Qwen, and 46.00% versus 34.96% for DeepSeek.

We further evaluated BBKI F64R hypotheses against existing experimental Arg64 structures withheld during inference. Of 18 extracted interaction events, nine are fully supported, three partially supported, and six unsupported. Supported hypotheses include an Arg64–Asp189 salt bridge and an Arg64–Gly219 backbone hydrogen bond absent from the supplied Phe64 structure. This provides additional evidence that some proposed mutation-induced contacts agree with experimentally observed mutant structures.

These results link prediction correctness to independently supported interaction changes.

3. Broader evaluation

To broaden coverage beyond the original paired benchmark, we added A300, comprising 300 unpaired interface-mutation cases from SKEMPI 2.0. We excluded overlapping records and structural mutations from the original benchmark and capped selection at two cases per mutated-protein name group. Of these cases, 290 are assigned to 139 protein-family labels, while ten remain unassigned. The overlap exclusions concern records and structural mutations; protein families may recur across evaluations.

Results of the added A300 evaluation (300 individual mutation cases):

MethodTask accuracy
GLM-5.3 Flash78.00%
FoldX581.33%
PyRosetta flex ddG80.00%
Always predicting non-enhancement69.33%

GLM-5.3 Flash also achieves 67.16% balanced accuracy and an MCC of 0.4393. Its enhancement precision is 78.26%, whereas recall is 39.13%: the model predicts enhancement in 46 of 300 cases, compared with 92 experimentally enhancing cases. These results extend the evaluation to a broader collection while also showing that the conservative prediction pattern persists. Appendix I provides the construction procedure, family distribution, and full results.

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

This paper constructs a benchmark for predicting partner-dependent mutation effects on binding affinity, using datasets curated from SKEMPI 2.0 and published literature. It compares general-purpose LLMs with the Biomni agentic workflow and physical predictors, including FoldX and Rosetta ddG. Some LLMs achieve group accuracies comparable to the physical predictors, although overall performance remains low. The paper also examines Python assistance and structural interventions, including masking, partner removal, cropping, and rotation. These experiments provide useful insights into information dependence and robustness, but the analysis of how models use structural evidence to reach their final judgments remains limited.

Critical strengths

The benchmark addresses a biologically meaningful problem: the same mutation can affect binding differently across partners. This design tests whether models account for the binding context rather than rely on general preferences for particular amino-acid substitutions. The mutation-group formulation and constrained inputs provide a useful framework for this benchmark. The extensive comparisons and ablations go beyond reporting prediction accuracy. Experiments involving Python assistance and structural perturbations provide useful observations about the information used by the models and the stability of their predictions.

Critical weaknesses

The presentation is difficult to follow. The prose often feels over-edited, favoring abstract evaluation terminology over direct descriptions of the structural biology procedures. Terms such as “audits” and “strata,” although valid, are not always clearly connected to the concrete operations they denote. This terminology-heavy style, together with methodological details scattered across the main text and appendices, makes relatively straightforward experiments unnecessarily difficult to reconstruct. For example, the main method comparisons use three-run majority voting, whereas some intervention conditions use five runs (lines 168–171 and 357–366). A consolidated description of each condition’s input transformation, repetition count, aggregation rule, and evaluation metric would substantially improve readability and reproducibility The low predictive performance is difficult to interpret without stronger baselines and implementation details. On Dataset 1, independent random guessing has an expected group accuracy of 25%, since both binary predictions must be correct. All reported full-dataset point estimates fall below this expectation, although some larger-improvement-subset results exceed it. An explicit random baseline would therefore help contextualize the results. Meanwhile, the configurations of Biomni and the physical predictors are insufficiently described, making it difficult to assess whether their low performance reflects the task’s difficulty or configuration-dependent limitations (lines 175–177; Table 1). The perturbation analysis does not fully distinguish intervention effects from run-to-run variability. The paper reports imperfect consistency even when the coordinates remain unchanged (lines 321–323), alongside prediction changes after mutation-site and distant side-chain masking (lines 297–302). These metrics use different units and cannot be compared directly. Matched repetition controls would help establish how much additional variation is attributable to each intervention. A more systematic analysis of model's chain of thoughts would also help explain how structural evidence influences the final judgment, rather than only showing that predictions change.

Questions for authors

How were Biomni, FoldX, and Rosetta ddG configured and evaluated? Please clarify their input preparation, execution settings, and the procedures used to convert outputs into directional predictions, so that readers can assess whether the comparisons are solid (lines 175–177; Table 1). Does mutation-site masking improve or reduce prediction accuracy? Figure 4a reports substantial changes in predictions, but these could include corrections of previously incorrect answers. Please report accuracy after masking and distinguish changes from correct to incorrect and vice versa (lines 297–302). Figure 3a shows that several models make fewer but more precise enhancement predictions, suggesting a conservative tendency to predict affinity enhancement. What structural evidence leads these models to predict enhancement, and how does their use of evidence differ between true-positive and false-negative cases (lines 259–264)? Among the processed structures whose identities DeepSeek recovered, is prediction accuracy or prediction bias different from that of the remaining cases? Have similar analyses been conducted for other models? This would help assess whether recognition or memorization contributes to the reported performance (Appendix A.6, lines 837–855).

14 comments
AuthorsOpenReview ↗

Response to Weaknesses: Part 1

Q: The presentation is difficult to follow. The prose often feels over-edited, favoring abstract evaluation terminology over direct descriptions of the structural biology procedures. Terms such as “audits” and “strata,” although valid, are not always clearly connected to the concrete operations they denote. This terminology-heavy style, together with methodological details scattered across the main text and appendices, makes relatively straightforward experiments unnecessarily difficult to reconstruct. For example, the main method comparisons use three-run majority voting, whereas some intervention conditions use five runs (lines 168–171 and 357–366). A consolidated description of each condition’s input transformation, repetition count, aggregation rule, and evaluation metric would substantially improve readability and reproducibility

A: Thank you for this feedback. We have revised the presentation to connect the evaluation terminology to the actual structural procedures and to make the experimental conditions easier to reconstruct.

First, we describe the concrete operations underlying the terminology. The starting-structure “audit” checks model-stated contacts and distances against the supplied coordinates using Mol* analysis. The effect-magnitude “strata” are two subsets of Dataset 1, each containing 21 groups, defined by whether the enhancing member shows an affinity improvement of at least twofold or less than twofold. These procedures and selection criteria are now stated explicitly.

Second, Appendix A now provides a consolidated table listing each condition's input transformation, repetition count, corresponding control, aggregation rule, and evaluation metric. This brings the settings for the main comparisons and structural interventions together in one place.

AuthorsOpenReview ↗

Response to Weaknesses: Part 2-1

Q: The low predictive performance is difficult to interpret without stronger baselines and implementation details. On Dataset 1, independent random guessing has an expected group accuracy of 25%, since both binary predictions must be correct. All reported full-dataset point estimates fall below this expectation, although some larger-improvement-subset results exceed it. An explicit random baseline would therefore help contextualize the results.

A: We have added model-matched random references, analyses of within-pair prediction patterns and effect magnitude, and a supplementary evaluation on 300 mutation cases (A300).

1. Clarifying the random reference for complete-group accuracy.

The 25% reference assumes independent predictions with equal probabilities of enhancement and weakening. However, the models predominantly predict non-enhancement. We therefore added model-specific random references that retain the independence assumption while matching each model’s observed enhancement-prediction frequency, p^\hat p. Because each group contains one enhancing and one weakening mutation, the expected complete-group accuracy is $\hat p(1-\hat p)$.

Experimental binding-mutation datasets and deep mutational scanning studies show that affinity-enhancing mutations are less common than non-enhancing mutations, rather than following a uniform 50:50 distribution [1, 2]. Dataset 1 is balanced within pairs; its model-matched references use each model's observed prediction frequencies.

ModelEnhancement predictionsObserved complete-group accuracyFrequency-matched random expectation
DeepSeek Flash8/84 (9.52%)6/42 (14.29%)8.62%
GLM-5.3 Flash13/84 (15.48%)8/42 (19.05%)13.08%
Qwen3.8 Flash6/84 (7.14%)4/42 (9.52%)6.63%

Each expectation is calculated as p^(1−p^)\hat p(1-\hat p), using the model’s observed enhancement frequency p^=K/84\hat p=K/84. The counts use final task-level consensus labels.

As an additional check, randomly reallocating exactly the same number of enhancement predictions among the 84 tasks gives expected group accuracies of 8.72%, 13.24%, and 6.71%, respectively. All three observed accuracies exceed these response-frequency-matched expectations. Appendix H gives the calculation.

2. Accounting for within-pair prediction patterns.

To account for within-pair prediction patterns, we add a conditional reference that preserves which groups receive identical or different labels and randomizes the orientation of the differently labeled groups.

Groups receiving identical labels cannot be fully correct because their experimental effects are opposite. A differently labeled group has a 50% chance of the correct orientation under this randomization. Consequently, if $D$ of the 42 groups receive different labels, the expected complete-group accuracy is D2×42\frac{D}{2\times42}.

For GLM, 11 groups receive different labels: eight are correctly oriented and three are reversed. Randomizing those orientations gives 5.5 correct groups on average, or 13.10% accuracy, compared with the observed 19.05%. The corresponding expectations for DeepSeek and Qwen are 7.14% and 4.76%, compared with observed accuracies of 14.29% and 9.52%.

Appendix H now provides these definitions, formulas, prediction counts, and results. The calculations give analytic expected rates. Together, they show how the results relate both to uniform guessing and to random references preserving the models’ response frequencies or observed pairwise differentiation.

AuthorsOpenReview ↗

Response to Weaknesses: Part 2-2

3. A300 provides additional context for the conservative prediction tendency.

We have also added A300, comprising 300 unpaired interface-mutation cases from SKEMPI 2.0, excluding overlapping records and structural mutations from the original benchmark and limiting selection to two cases per mutated-protein name group. Of these cases, 290 are assigned to 139 protein-family labels.

Although 92/300 cases (30.67%) are experimentally enhancing, GLM predicts enhancement in only 46/300 (15.33%), assigning 84.67% to non-enhancement. This provides an empirical indication of its conservative prediction tendency and a rough proxy for its prediction prior. Using this response frequency in p^(1−p^)\hat p(1-\hat p) gives 12.98% for hypothetical opposite-effect pairs, further illustrating why 25% does not reflect the model’s observed response tendency. The Dataset 1 references above are calculated from Dataset 1’s own predictions.

Results of the added A300 evaluation (300 individual mutation cases):

MethodTask accuracy
GLM-5.3 Flash78.00%
FoldX581.33%
Five-model Raw flex ddG80.00%
Always predicting non-enhancement69.33%

GLM-5.3 Flash also achieves 67.16% balanced accuracy and an MCC of 0.4393. Appendix I reports the dataset construction, evaluation procedures, and full results.

References

[1] Rodrigues, C. H. M., Myung, Y., Pires, D. E. V., and Ascher, D. B. (2019). mCSM-PPI2: predicting the effects of mutations on protein-protein interactions. Nucleic Acids Research, 47(W1), W338-W344. doi:10.1093/nar/gkz383.

[2] Heyne, M., Shirian, J., Cohen, I., Peleg, Y., Radisky, E. S., Papo, N., and Shifman, J. L. (2021). Climbing Up and Down Binding Landscapes through Deep Mutational Scanning of Three Homologous Protein-Protein Complexes. Journal of the American Chemical Society, 143(41), 17261-17275. doi:10.1021/jacs.1c08707.

Reviewer_2EwmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper evaluates whether general-purpose LLMs can infer partner-dependent mutation effects from local protein-complex structures without dedicated mutation-effect predictors. The benchmark contains 60 mutation groups and 126 mutation–partner observations, with the focal protein and substitution held fixed across partners. Models receive independently presented, identity-obscured structural inputs, with or without Python access. Evaluation includes complete-group accuracy, effect-magnitude stratification, structural masking and cropping, rigid-body rotations, structural-claim audits, and mechanistic case studies. The results indicate that predictions depend on structural context and that some coordinate-grounded explanations are plausible, while overall predictive accuracy remains modest.

Critical strengths
  1. A relevant question addressed through a well-motivated contrastive setup. Holding the protein and substitution fixed while changing the binding partner directly tests context dependence and limits the usefulness of generic substitution-level heuristics. Separating evidence-gathering tools from dedicated effect predictors also helps identify the LLM’s contribution to the final judgment.

  2. Multiple complementary analyses of structural evidence use. Masking, partner removal, coordinate removal, rotations, and claim audits provide a broader characterization than prediction accuracy alone. The paper appropriately distinguishes coordinate-supported observations from experimentally validated mechanisms, and includes cases where correct predictions are accompanied by incorrect explanations.

Critical weaknesses
  1. The predictive results do not clearly establish performance beyond chance or response biases. On Dataset 1, LLM complete-group accuracies range from 9.52% to 19.05%, compared with an expected 25% under independent uniform binary guessing. Although this reference does not account for model-specific prediction biases or correlations between outputs, the absence of appropriate null comparisons leaves the significance of successful partner contrasts unclear. The effect-magnitude comparison is also based on only 21 groups per stratum, with multiple groups sharing protein systems. Consequently, the observed successes and stratified differences provide limited evidence of reliable partner-dependent prediction. Similar performance to physical predictors with low absolute accuracy does not strengthen this inference.

  2. Structural information dependence is not sufficiently distinguished from valid physical inference. Removing coordinates or partner information can shift models toward negative predictions, which directly reduces complete-group accuracy on opposite-effect pairs. The resulting performance changes therefore do not isolate correct interpretation of mutation-relevant geometry. The claim audit establishes agreement with supplied coordinates for 46.90% of extracted statements, but does not assess the remaining statements or validate energetic attribution and mutant-state mechanisms. Although the paper acknowledges these limitations, the evidence supports structural sensitivity and the grounding of selected observations more clearly than the central interpretation of physical intuition. Limited protein-system diversity and residual recognition or memorization further constrain that interpretation.

Questions for authors
  1. Performance relative to chance. How should the reported complete-group accuracies be interpreted relative to chance, given that independent uniform binary guessing yields an expected accuracy of 25% on Dataset 1? To what extent could each model’s tendency to predict enhancement, together with the dependence between its predictions for different partners within the same group, explain the observed accuracies?

  2. Interpretation of structural interventions. After structural information is removed, do models lose the ability to distinguish partner-dependent mutation effects, or mainly become more likely to predict weakening for both? What supports the interpretation that the original successful predictions came from correctly understanding the structure?

  3. Intervention specification. In mutation-site masking, which information is hidden, and is the mutation specification still provided?

  4. Recognition and generalization. What constitutes successful protein-identity recovery in the reported 16/84 result: identifying a protein family or a specific protein, and at the chain or task level? Is successful identity recovery associated with correct mutation-effect prediction? What do the recognition results establish about the possible contribution of prior familiarity with the evaluated systems?

  5. Scope and reliability of the claim audit. How were structural claims extracted, and how was the reliability of claims automatically classified as supported assessed? Does the proportion of claims supported by the supplied coordinates differ between correct and incorrect mutation-effect predictions?

8 comments
AuthorsOpenReview ↗

Response to Concern 1: Model-matched random references

Concern 1. The predictive results do not clearly establish performance beyond chance or response biases. On Dataset 1, LLM complete-group accuracies range from 9.52% to 19.05%, compared with an expected 25% under independent uniform binary guessing. Although this reference does not account for model-specific prediction biases or correlations between outputs, the absence of appropriate null comparisons leaves the significance of successful partner contrasts unclear. The effect-magnitude comparison is also based on only 21 groups per stratum, with multiple groups sharing protein systems. Consequently, the observed successes and stratified differences provide limited evidence of reliable partner-dependent prediction. Similar performance to physical predictors with low absolute accuracy does not strengthen this inference.

We thank the reviewer for this comment. We have added model-matched random references, analyses of within-pair prediction patterns and effect magnitude, and a supplementary evaluation on 300 mutation cases (A300).

1. Clarifying the random reference for complete-group accuracy

The 25% reference assumes independent predictions with equal probabilities of enhancement and weakening. However, the models predominantly predict non-enhancement. We therefore added model-specific random references that retain the independence assumption while matching each model’s observed enhancement-prediction frequency, $\hat p$. Because each group contains one enhancing and one weakening mutation, the expected complete-group accuracy is p^(1−p^)\hat p(1-\hat p).

Experimental binding-mutation datasets and deep mutational scanning studies show that affinity-enhancing mutations are less common than non-enhancing mutations, rather than following a uniform 50:50 distribution [1, 2]. Dataset 1 is balanced within pairs; its model-matched references use each model's observed prediction frequencies.

ModelEnhancement predictionsObserved complete-group accuracyFrequency-matched independent expectation
DeepSeek Flash8/84 (9.52%)6/42 (14.29%)8.62%
GLM-5.3 Flash13/84 (15.48%)8/42 (19.05%)13.08%
Qwen3.8 Flash6/84 (7.14%)4/42 (9.52%)6.63%

As an alternative reference, we considered randomly reassigning a fixed number, KK, of enhancement predictions among the 84 observations. This preserves the exact enhancement count and gives an expected complete-group accuracy of K(84−K)84×83\frac{K(84-K)}{84\times83}: 8.72%, 13.24%, and 6.71% for the three models, respectively. All three observed accuracies exceed both frequency-matched expectations.

AuthorsOpenReview ↗

Response to Concern 1 (continued): Within-pair patterns and A300 evaluation

2. Accounting for within-pair prediction patterns

To account for within-pair prediction patterns, we added a conditional reference that preserves whether each group receives two enhancement predictions, two weakening predictions, or different labels, and randomizes only the orientation of groups receiving different labels.

Because the experimental effects within each group are opposite, groups assigned the same label cannot be fully correct. For groups assigned different labels, a random orientation has a 50% probability of being correct. Thus, if DD of the 42 groups receive different labels, the expected complete-group accuracy is D2×42\frac{D}{2\times42}.

ModelGroups receiving different labelsCorrectly oriented groupsObserved complete-group accuracyConditional random expectation
DeepSeek Flash6614.29%7.14%
GLM-5.3 Flash11819.05%13.10%
Qwen3.8 Flash449.52%4.76%

These analyses, now reported in Appendix H, distinguish three aspects of performance: how often a model predicts enhancement, how often it assigns opposite effects to the two partners, and how often it assigns those effects in the correct orientation.

3. Extending the evaluation with A300

We have also added A300, comprising 300 unpaired interface-mutation cases from SKEMPI 2.0, excluding overlapping records and structural mutations from the original benchmark and limiting selection to two cases per mutated-protein name group. Of these cases, 290 are assigned to 139 protein-family labels.

Although 92/300 cases (30.67%) are experimentally enhancing, GLM predicts enhancement in only 46/300 (15.33%), assigning 84.67% to non-enhancement. This provides an empirical indication of its conservative prediction tendency and a rough proxy for its prediction prior. Using this response frequency in p^(1−p^)\hat p(1-\hat p) gives 12.98% for hypothetical opposite-effect pairs, further illustrating why 25% does not reflect the model’s observed response tendency. The Dataset 1 references above are calculated from Dataset 1’s own predictions.

Results of the added A300 evaluation (300 individual mutation cases):

MethodTask accuracy
GLM-5.3 Flash78.00%
FoldX581.33%
PyRosetta flex ddG80.00%
Always predicting non-enhancement69.33%

GLM-5.3 Flash also achieves 67.16% balanced accuracy and an MCC of 0.4393. Appendix I reports the dataset construction, evaluation procedures, and full results.

References

[1] Rodrigues, C. H. M., Myung, Y., Pires, D. E. V., and Ascher, D. B. (2019). mCSM-PPI2: predicting the effects of mutations on protein-protein interactions. Nucleic Acids Research, 47(W1), W338-W344. doi:10.1093/nar/gkz383.

[2] Heyne, M., Shirian, J., Cohen, I., Peleg, Y., Radisky, E. S., Papo, N., and Shifman, J. L. (2021). Climbing Up and Down Binding Landscapes through Deep Mutational Scanning of Three Homologous Protein-Protein Complexes. Journal of the American Chemical Society, 143(41), 17261-17275. doi:10.1021/jacs.1c08707.

AuthorsOpenReview ↗

Response to Concern 2: Point1

Concern 2. Structural information dependence is not sufficiently distinguished from valid physical inference. Removing coordinates or partner information can shift models toward negative predictions, which directly reduces complete-group accuracy on opposite-effect pairs. The resulting performance changes therefore do not isolate correct interpretation of mutation-relevant geometry. The claim audit establishes agreement with supplied coordinates for 46.90% of extracted statements, but does not assess the remaining statements or validate energetic attribution and mutant-state mechanisms. Although the paper acknowledges these limitations, the evidence supports structural sensitivity and the grounding of selected observations more clearly than the central interpretation of physical intuition. Limited protein-system diversity and residual recognition or memorization further constrain that interpretation.

We thank the reviewer for highlighting the distinction between sensitivity to structural inputs and physically supported inference. We have expanded the evaluation to assess both whether predictions depend on the supplied structure and whether the proposed mutation-induced interaction changes are supported by paired WT/mutant simulations and experimental mutant structures. The revised manuscript distinguishes three questions: whether a model accurately describes the supplied structure, whether it anticipates interaction changes upon mutation, and whether it correctly predicts the resulting affinity effect.

1. Partner removal reduces accuracy without shifting predictions toward non-enhancement.

We now report five runs for both intact and partner-removed inputs, including enhancement-prediction counts and the full distribution of paired outcomes (Appendix N). Each condition comprises 420 task predictions per model: 84 observations evaluated over five runs.

ModelEnhancement predictions, intact → removedComplete-group accuracy, intact → removedCorrect orientation among pairs assigned different labels, intact → removed
DeepSeek Flash40 → 4812.86% → 9.05%96.43% → 59.38%
GLM-5.3 Flash67 → 7316.67% → 10.00%77.78% → 56.76%
Qwen3.8 Flash31 → 3411.90% → 4.29%92.59% → 32.14%

Enhancement predictions increase for all three models after partner removal, while complete-group accuracy declines. Across the 210 group predictions per condition, the number of groups with both effect directions reversed rises from 1 to 13 for DeepSeek Flash, from 10 to 16 for GLM-5.3 Flash, and from 2 to 19 for Qwen3.8 Flash. The accuracy loss therefore cannot be explained by a general shift toward non-enhancement; partner removal also impairs the models’ ability to assign the correct direction of the partner contrast.

To examine the role of mutation-site geometry more specifically, we added a side-chain-masking analysis. Masking the mutation-site side chain while retaining the backbone, residue identity, and mutation specification changes approximately 42% of group predictions, compared with approximately 10% when a distant side chain with cross-chain contacts is masked. This difference supports preferential sensitivity to mutation-site geometry. The mutant-state evaluations below separately assess whether the proposed interaction changes have physical support.

Reviewer_w8FLOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper evaluates whether general-purpose LLMs can judge partner-dependent mutation effects directly from local coordinates, without dedicated predictors. The benchmark holds the focal protein and substitution fixed while varying the binding partner: 60 groups and 126 observations, split into 42 SKEMPI-derived groups and 18 literature-curated groups. Complexes are presented independently as identity-obscured 20 Å crops, and a group counts as correct only if every partner direction is right. Four LLMs are run with and without Python against FoldX, Rosetta Flex ddG, and Biomni, supplemented by masking, cropping, rotation, and claim-audit experiments.

Critical strengths

The partner-contrastive design is strong. Each group pairs one enhancing and one weakening observation of the same substitution, and both have to be correct, so any partner-blind scorer is pushed to 0% by construction. That is a much sharper probe than aggregate accuracy. Section 4 is also the right shape of experiment for a benchmark that asks what a judgment depends on, and the paper is careful about its own limits, repeatedly saying what each result does not establish.

Critical weaknesses
  1. The claim audit supports a narrower conclusion than the paper uses it for. The headline number is 1,574/(1,574+2), after 1,784 claims (53.10%) are set aside as out of scope. On top of that, Appendix G.2 shows 94.79% of the audited claims are hydrogen bonds or salt bridges, against only 2.86% for hydrophobic and van der Waals contacts, demonstrating the imbalance audit distribution. The audit is also not split by condition, and under PDB + Python the model computes distances with Biopython from the same file Mol* then re-reads, so part of the agreement is just two deterministic computations matching. Please report the audited fraction separately for the PDB-only and PDB + Python conditions, and audit a sample of the out-of-scope claims.

  2. I cannot interpret the Biomni comparison as stated. Biomni is a harness, so its result depends on the backing model, the tool whitelist and the orchestration, and none of these is reported anywhere. Besides, Biomni ships a large tool library, so whether it was permitted to call any mutation-effect or energetic predictor also bears on Section 2.2’s exclusion of such tools.

  3. The model panel makes the results hard to place for no frontier model is included. Every entry in the 42-group column of Table 1 sits at or below the 25% you would expect from coin-flipping on opposite-effect pairs. Therefore, I cannot tell whether this task is beyond current LLMs or just beyond the models evaluated in this paper. Running two or three frontier models, such as Kimi K3, Opus 5.5 or GPT-6, under the Appendix F prompt would be informative whichever way it comes out.

Questions for authors

Please see the above weaknesses.

6 comments
AuthorsOpenReview ↗

Response to Concern 1-1

Q1. The claim audit supports a narrower conclusion than the paper uses it for. The headline number is 1,574/(1,574+2), after 1,784 claims (53.10%) are set aside as out of scope. On top of that, Appendix G.2 shows 94.79% of the audited claims are hydrogen bonds or salt bridges, against only 2.86% for hydrophobic and van der Waals contacts, demonstrating the imbalance audit distribution. The audit is also not split by condition, and under PDB + Python the model computes distances with Biopython from the same file Mol* then re-reads, so part of the agreement is just two deterministic computations matching. Please report the audited fraction separately for the PDB-only and PDB + Python conditions, and audit a sample of the out-of-scope claims.

A: We have revised the evaluation to report audit results separately by tool condition, distinguish audit coverage from support rate, and complement automated contact checks with case studies of broader structural claims and MD-based evaluation of mutation hypotheses.

1. Separate audit results for PDB-only and PDB + Python

The original 3,360 claims were generated under the PDB + Python condition. We additionally audited 756 predictions (252 per model) under PDB-only, where Python execution was unavailable. We define audit coverage as the proportion of extracted claims that are assessable under the audit rules, and support rate as the proportion of assessable claims supported by the coordinate checks.

ConditionExtracted claimsAssessable claimsSupportedConflictsOutside scopeAudit coverageSupport rate
PDB + Python3,3601,5761,57421,78446.90%99.87%
PDB-only1,3176786661263951.48%98.23%

The PDB-only results are further broken down by model:

ModelExtracted claimsSupportedConflictsOutside scope
GLM-5.3 Flash6363186312
DeepSeek Flash3751926177
Qwen3.8 Flash3061560150

Agreement with the supplied coordinates remains high among assessable claims even when models cannot execute Python to calculate distances.

In the PDB + Python condition, models receive anonymized PDB coordinates and may write code to inspect them. Mol interaction records are used only in the subsequent audit and are not supplied to the models.*

2. Claim extraction and interaction-type coverage

Models return structured JSON answers containing a structural_evidence array. We examine every entry and extract independently checkable relationships from the type, observation, and measurement fields. Each claim is matched against Mol* records for the same task and input PDB using the mutation site, interaction type, partner chain and residue, and any specified backbone or side-chain restrictions. Where distances are comparable, we apply a tolerance of 0.15 Å.

The reported fractions of hydrogen-bond/salt-bridge claims (94.79%) and hydrophobic/van der Waals claims (2.86%) describe the audited claim distribution. This distribution reflects both what the models report and what the matching rules can assess. Hydrogen bonds and salt bridges map readily to discrete interaction records, whereas statements about packing, burial, cavities, and flexibility often cannot be scored automatically under the current matching rules and are therefore excluded from the automated audit.

AuthorsOpenReview ↗

Response to Concern 1-2

3. Case studies assess broader structural explanations

We complement the automated audit with manual examination of selected explanations, using structural measurements, experimental mutant structures, and the associated experimental reports:

  • ShPI-1 K13L: We examine the proposed roles of buried charge and hydrophobic pocket preference. The analysis identifies polar compensation around the buried Lys, qualifying the model’s interpretation of the local environment.
  • BBKI F64R: We assess packing changes, direct and water-mediated contacts, and alternative Arg conformations. Some predicted contacts are supported, whereas others specify incorrect acceptor atoms or interaction types.
  • EVA-A L39P/Y44A: We examine pocket rearrangement and accommodation of the partner Tyr. For L39P, the model considers favorable Tyr repacking but discounts it in the final judgment; the experimental mutant structure shows deeper Tyr accommodation and favorable aromatic interactions.
  • BLIP E73W: We compare the proposed aromatic-packing explanation with the mutant structure, which reveals a prominent Trp73–Tyr105 contact and an altered local interaction pattern.

These case studies examine claims about burial, packing, water-mediated interactions, and conformational changes using structural measurements and experimental comparisons.

4. MD comparisons assess mutation hypotheses in relation to prediction accuracy

We also used paired WT/mutant MD simulations to test whether the interaction changes proposed by the models occur after mutation, including changes in hydrogen bonds, salt bridges, nonpolar packing, and aromatic interactions.

ModelMD support for events accompanying correct predictionsMD support for events accompanying incorrect predictions
GLM-5.3 Flash42/82 (51.22%)22/63 (34.92%)
Qwen3.8 Flash48/112 (42.86%)25/78 (32.05%)
DeepSeek Flash49/113 (43.36%)40/110 (36.36%)

For all three models, MD support is higher for interaction events accompanying correct affinity predictions. The same pattern holds within the 21 groups whose enhancing member improves affinity by at least twofold: 60.87% versus 50.00% for GLM-5.3 Flash, 54.24% versus 37.50% for Qwen3.8 Flash, and 50.82% versus 38.46% for DeepSeek Flash. Across the evaluated cases and within the larger-improvement subset, correct affinity predictions are accompanied by higher MD support for the proposed interaction changes.

The revised evaluation combines coordinate checks, structural case studies, and paired WT/mutant MD comparisons to assess structural observations and proposed mutation-induced changes.

AuthorsOpenReview ↗

Response to Concern 2

Q2. I cannot interpret the Biomni comparison as stated. Biomni is a harness, so its result depends on the backing model, the tool whitelist and the orchestration, and none of these is reported anywhere. Besides, Biomni ships a large tool library, so whether it was permitted to call any mutation-effect or energetic predictor also bears on Section 2.2’s exclusion of such tools.

A: We evaluated Biomni through the publicly accessible free web interface on the official Biomni website with the "Max - Most capable model, highest quota" option selected. The backing model and orchestration were managed by the platform. The interface does not expose the underlying model identifier. It received the original, uncropped and non-anonymized PDB structures together with the mutation specification.

All tools available in that interface were permitted, including energetic and mutation-related prediction tools. Recorded executions included OpenMM, PRODIGY, and ProteinMPNN, alongside structural inspection, rotamer scanning, local energy minimization, and MM/GBSA. Biomni selected these operations through its workflow; its final affinity-direction judgment was scored as the task prediction.

The exclusion of dedicated energetic and mutation-effect predictors in Section 2.2 applies to the direct LLM evaluation conditions, not to Biomni. The direct LLMs received anonymized local structures and, where enabled, Python for structural inspection. Biomni serves as an external comparator with broader structural inputs and unrestricted tool access. We have expanded Appendix A to describe this distinction and the configuration underlying its reported results.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Overall Response and Summary of Revisions (1/2)

We thank all four reviewers for their constructive feedback. We have made four main revisions:

  1. Independent evaluation of mutation hypotheses. We added paired WT/mutant MD comparisons covering 82 mutation–partner cases, examining how support for proposed interaction changes relates to prediction correctness, affinity-improvement magnitude, and interaction type, with a separate PyRosetta-based paired side-chain repacking comparator. We also expanded the BBKI analysis to score 18 interaction predictions against experimental mutant structures withheld during inference.

  2. Broader prediction and exploratory design evaluations. We added A300, comprising 300 additional unpaired interface-mutation cases (139 protein-family), with documented selection criteria, overlap exclusions, protein-family annotations, and comparisons with FoldX5 and Rosetta flex ddG. We further added six BPTI partner-selective design tasks, evaluating proposed substitutions against withheld experimental reference data and physical-method recommendations.

  3. Random references and expanded intervention analyses. We added random expectations matched to prediction frequencies, fixed enhancement counts, and within-pair label patterns. We expanded the partner-removal analysis with five-run results, response counts, and paired-outcome distributions, and report task/group accuracies and correctness transitions after mutation-site masking. Additional analyses connect masking with changes in both individual answers and MD hypothesis support.

  4. Revised presentation and expanded methodological reporting. We revised the manuscript to distinguish observations of the supplied structure, hypotheses about mutation-induced changes, and predictions of affinity effects. Figures 2–4 and the accompanying tables now present prediction performance, hypothesis support, and intervention outcomes together, with coverage distinguished from support among assessable outputs. A consolidated experiment table records input transformations, repetitions, controls, aggregation rules, and metrics. We also expanded the configurations of FoldX5, Rosetta flex ddG, and Biomni, and clarified the structured outputs, hypothesis-extraction procedures, MD protocols, and scoring criteria.

We have also carefully addressed every concern and question in detailed, point-by-point responses to each reviewer. Below, we discuss the shared concerns about prediction performance and physical inference, followed by their implications for rational design.

1. Interpreting prediction performance against appropriate random references.

The 25% complete-group expectation assumes independent, uniform binary guessing. To account for the models’ tendency to predict non-enhancement and their within-pair prediction patterns, we added random references that preserve each model’s observed enhancement-prediction frequency, together with a conditional reference that preserves which pairs receive identical or different labels and randomizes only the orientation of differently labeled pairs.

All three evaluated Flash models exceed these model-matched expectations. For example, GLM-5.3 Flash achieves 19.05% complete-group accuracy, compared with 13.08% under frequency-matched independent guessing and 13.10% under conditional orientation randomization (Appendix H). These analytic references distinguish enhancement-prediction frequency from successful assignment of opposite effects to the appropriate partners.

The additional A300 evaluation broadens this assessment to 300 individual mutation cases. GLM achieves 78.00% accuracy, compared with 69.33% for always predicting non-enhancement, with 67.16% balanced accuracy and an MCC of 0.4393. Together, these analyses characterize predictive signal alongside the conservative response tendency.

AuthorsOpenReview ↗

Overall Response and Summary of Revisions (2/2)

2. Connecting successful predictions with independently supported mutation hypotheses.

The most substantial addition is an evaluation of the interaction changes that models propose for the unobserved mutant state. This addresses a distinction emphasized by the reviewers: accurately describing supplied coordinates, anticipating mutation-induced changes, and predicting their net affinity consequence are separate achievements.

We added paired WT/mutant MD comparisons covering 82 mutation–partner cases. Interaction fingerprints were extracted independently of the model hypotheses and experimental affinity labels, and the resulting changes were compared with the models’ explicitly stated hypotheses. Across all three models, hypotheses accompanying correct affinity predictions receive greater MD support than those accompanying incorrect predictions: 51.22% versus 34.92% for GLM, 43.36% versus 36.36% for DeepSeek, and 42.86% versus 32.05% for Qwen. This direction is retained when each scorable answer receives equal weight and when the comparison is restricted to groups with at least twofold affinity improvement.

The larger-improvement groups also show higher overall hypothesis support in all three models, alongside higher complete-group accuracy for GLM and DeepSeek; Qwen’s group accuracy is unchanged. Thus, prediction success, effect magnitude, and support for the proposed physical changes exhibit related patterns across several comparisons.

Structural interventions provide complementary evidence. Mutation-site masking changes substantially more predictions than distant-side-chain masking. Removing the binding partner reduces complete-group accuracy and impairs the orientation of partner contrasts even though enhancement predictions increase in all three models. This loss therefore cannot be explained simply by a shift toward negative answers. Selected experimental mutant structures provide a further check: in the BBKI F64R case, models anticipate specific contacts, including an Arg64–Asp189 salt bridge and an Arg64–Gly219 backbone hydrogen bond, that are absent from the supplied starting structure and observed in the withheld mutant structure.

Taken together, these observations support structure-grounded physical inference by linking prediction correctness to independently supported interaction hypotheses and demonstrating selective dependence on mutation-relevant structural information.

3. Practical value and potential for improved mutation-design tools.

The central contribution is evidence that general-purpose LLMs can themselves connect molecular structural information to partner-dependent mutation judgments, without dedicated energetic or mutation-effect predictors supplying those judgments. Computation may assist coordinate inspection, while the model selects and interprets the evidence and proposes the resulting mutation effect.

Direct LLM judgments achieve performance comparable to the tested physical predictors. On A300 dataset (300 cases,139 protein-family), GLM’s 78.00% accuracy compares with 81.33% for FoldX5 and 80.00% for Rosetta flex ddG. The paired benchmark additionally reveals different prediction profiles and partially different successful cases, suggesting potential complementarity. The exploratory BPTI partner-selective mutation design evaluation provides a more direct illustration: GLM generates five successful partner-selective recommendations across three rounds of six tasks each (18 attempts in total), succeeding at least once on three of six tasks against withheld experimental reference data. These results support a role in proposing candidates and testable structural hypotheses for subsequent screening.

The longer-term implication concerns how structural knowledge can be used and improved. FoldX and Rosetta encode biophysical knowledge through energy functions and structural modelling procedures. LLMs could contribute by proposing alternative mutation hypotheses, identifying calculations that distinguish them, and revising judgments using computational or experimental feedback. Successful strategies might ultimately inform inspectable structural descriptors, context-dependent scoring rules, or more effective calculation-selection procedures. Developing and testing such strategies could contribute to a next generation of more accurate mutation-prediction and protein-design tools.

The revised results connect partner-dependent affinity predictions with physically supported mutation hypotheses, show performance comparable to the tested physical predictors with partly different successful cases, and demonstrate candidate generation against withheld experimental reference data. Together, these findings motivate further development of LLM-based structural inference for mutation prediction and design.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.