Externally Validated Probabilistic Modeling of a Predefined Entecavir Resistance Pathway in HBV Using Independent Public Repositories

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The study is well structured, and the methodology is well organized. The supplementary material is comprehensive and provides all necessary supporting details. Overall, the manuscript is suitable for publication in its current form.

Author Response

Comment 1: The study is well structured, and the methodology is well organized. The supplementary material is comprehensive and provides all necessary supporting details. Overall, the manuscript is suitable for publication in its current form.

Response 1: We sincerely thank the reviewer for the positive evaluation of our manuscript and for recognizing the quality of the study design, methodology, and supplementary material.

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript presents a transparent and externally validated machine learning framework for identifying entecavir resistance pathways in HBV. The study is well designed, methodologically sound, and clearly written. The use of interpretable modeling and independent external validation is a strength. Overall, the manuscript is suitable for publication after minor revision.

1. The outcome definition is based on predefined resistance mutations, which likely explains the perfect internal performance. The authors should better clarify the added value of this probabilistic model compared with existing rule-based approaches.
2. Since the model targets a genotypic proxy rather than clinical outcomes, a brief discussion on its practical or clinical applicability would strengthen the manuscript.
3. The feature set is predefined based on known RT positions. Please briefly comment on whether alternative feature selections were considered or how this choice may affect generalizability.
4. More details on the external validation dataset (e.g., potential differences in data source, population, or sequence characteristics) would improve interpretability.

Author Response

Comment 1: The outcome definition is based on predefined resistance mutations, which likely explains the perfect internal performance. The authors should better clarify the added value of this probabilistic model compared with existing rule-based approaches.
Response 1: Thank you for pointing this out. We agree with this important comment. We revised the manuscript to clarify that the primary objective of the framework was not to replace existing rule-based resistance interpretation systems or to discover novel resistance mechanisms, but rather to provide a transparent probabilistic implementation enabling calibration assessment, threshold-based evaluation, and formal external validation across independent repositories. Specifically:

1. In the Introduction (page 2, paragraph 1), we revised the discussion of conventional rule-based systems and clarified the methodological rationale for probabilistic modeling by adding the following sentense:
   While such approaches have been instrumental in clinical practice, they generally provide deterministic binary classifications and may be difficult to evaluate quantitatively across heterogeneous sequence datasets [8,9]. In contrast, probabilistic prediction frameworks can provide calibrated risk estimates, support threshold-based interpretation, and enable formal assessment of discrimination, calibration, and transportability across independent datasets [11-14].
2. In the Results section (Internal validation performance) (page 6-7, paragraph 3.2), we clarified that the near-perfect internal performance reflects reconstruction of a predefined biological pathway rather than prediction of an independent clinical endpoint:
   Given the deterministic nature of the outcome definition and feature representation, these results reflect the model’s ability to reproduce a predefined genotypic resistance pathway rather than predict an independent clinical outcome. The primary purpose of the framework was therefore not to outperform existing rule-based interpretation systems in reconstructing known resistance definitions, but to provide a transparent probabilistic implementation that enables calibration assessment, threshold-based evaluation, and formal external validation across independent datasets.
3. In the Discussion (page 9, paragraph 4), we added a dedicated paragraph clarifying the added value of the proposed probabilistic framework compared with conventional rule-based interpretation systems:
   Importantly, the present framework should not be viewed as a replacement for established rule-based resistance interpretation systems, which remain appropriate for well-characterized HBV resistance pathways [4,8,9]. Rather, its added value lies in providing a standardized probabilistic framework that can be externally validated and quantitatively evaluated using discrimination, calibration, and threshold-dependent performance metrics [11-14]. Such an approach may be particularly relevant in settings involving heterogeneous sequence repositories, evolving resistance patterns, or future integration with phenotypic and longitudinal clinical data.
4. In the Limitations section (page 10, paragraph 4), we additionally clarified that the framework represents probabilistic formalization of an established biological pathway rather than de novo resistance discovery:
   In addition, because the outcome definition was partly derived from established resistance-associated substitutions included within the predefined feature panel, the framework should be interpreted primarily as a probabilistic formalization of a known biological resistance pathway rather than de novo discovery of novel resistance mechanisms.
5. Finally, in the Supplementary Material (Supplementary Section 4.5, page 8), we added clarification regarding the intended role of the framework:
   Accordingly, the primary objective of the framework was not discovery of novel resistance rules, but development of a calibrated and externally validated probabilistic implementation for standardized evaluation across independent sequence repositories.
   
   

Comment 2: Since the model targets a genotypic proxy rather than clinical outcomes, a brief discussion on its practical or clinical applicability would strengthen the manuscript.
Response 2: Thank you for this valuable suggestion. To clarify the practical applicability of the proposed approach, we expanded the Discussion section (page 9, paragraph 4) by adding a paragraph describing potential applications of probabilistic HBV sequence interpretation in research and surveillance settings. We clarified that the model is not intended for standalone therapeutic decision-making, but may support large-scale screening of public HBV repositories, identification of sequence patterns requiring expert evaluation, and future integration with phenotypic or longitudinal clinical data. The following text was added:
Although the present model was not designed for direct therapeutic decision-making, the proposed approach may still have practical utility in research and surveillance contexts. Standardized probabilistic interpretation of HBV sequence data could support large-scale screening of publicly available repositories and assist in identifying sequence patterns that warrant further expert evaluation. In addition, quantitative modeling approaches may facilitate more consistent comparison of resistance-associated profiles across heterogeneous datasets and provide a structured basis for future integration with phenotypic resistance data, treatment history, or longitudinal clinical observations [11-14].

Comment 3: The feature set is predefined based on known RT positions. Please briefly comment on whether alternative feature selections were considered or how this choice may affect generalizability.
Response 3: Thank you for this important comment. We clarified the rationale for restricting the feature set to predefined RT positions in the Methods section 2.2 (page 4), emphasizing that this choice was intended to preserve biological interpretability and consistency with established HBV resistance literature. The following text was added in the Methods section:
RT positions were intentionally restricted to previously reported resistance-associated regions to preserve biological interpretability and consistency with established HBV resistance literature [4,8,9]. Alternative data-driven feature selection strategies were not explored because the aim of the study was validation of a predefined resistance pathway rather than de novo mutation discovery.

We also added a statement in the Discussion section 4 (page 10) acknowledging that this approach may limit detection of previously unrecognized resistance-associated sequence patterns.  The following text was added in the Limitations section:
Because feature selection was limited to predefined RT positions, the model may not capture previously unrecognized sequence patterns associated with antiviral resistance.

Comment 4: More details on the external validation dataset (e.g., potential differences in data source, population, or sequence characteristics) would improve interpretability. 
Response 4: Thank you for this important suggestion. We expanded the description of the external validation dataset in the Methods section 2.4 (page 5) to better clarify differences between the NCBI-derived development cohort and the independently curated HBVdb external dataset, as well as the preprocessing procedures applied prior to external evaluation. The following text was added in the External Validation subsection:
External validation was conducted using an independently curated HBVdb-derived cohort distinct from the NCBI-derived development dataset. External sequences underwent separate preprocessing, including six-frame translation, quality-control filtering, motif screening, and deduplication using sequence hash matching to minimize overlap with the development cohort. Preparation of the external dataset was performed independently of model development. The trained model, feature encoder, calibration parameters, and decision thresholds were frozen prior to external evaluation and applied unchanged to the validation dataset. This design was intended to assess model transportability across heterogeneous public sequence repositories processed through partially independent curation workflows.

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript presents a machine learning–based framework for identifying entecavir resistance pathways in hepatitis B virus (HBV) polymerase sequences using publicly available datasets. By integrating curated GenBank sequences with motif-anchored reverse transcriptase numbering and calibrated logistic regression modeling, the authors aim to establish a transparent and reproducible system for probabilistic resistance interpretation. The study further validates the model using an independent HBVdb cohort and demonstrates stable performance despite substantial differences in resistance prevalence between datasets. Overall, the work addresses a clinically relevant problem and highlights the potential utility of machine learning in antiviral resistance prediction. However, the following issues are required for explaining:

1. In addition to the HBVdb dataset, the authors are encouraged to validate the model using additional independent external cohorts to further demonstrate generalizability and robustness across different populations and sequencing platforms.
2. As this is essentially a diagnostic prediction study, the manuscript should report additional clinically relevant performance metrics, including positive predictive value (PPV) and negative predictive value (NPV), especially given the low prevalence of entecavir resistance in the external validation cohort.
3. The authors should clearly describe the strategy and rationale used for selecting the probability cutoff threshold. It is important to explain whether the cutoff was optimized using Youden index, calibration analysis, clinical sensitivity/specificity balance, or other predefined rules.
4. Chronic HBV infection is highly heterogeneous, including differences in viral genotype, disease phase, viral load, cirrhosis status, and prior antiviral exposure. The authors should investigate whether the predictive performance of the model varies across different HBV infection subgroups and perform subgroup analyses where possible.
5. The discussion section should more comprehensively address how this model could be translated into clinical practice, including its potential integration into sequencing workflows, antiviral treatment selection, resistance monitoring, and future clinical decision-support systems.

Author Response

Comment 1: In addition to the HBVdb dataset, the authors are encouraged to validate the model using additional independent external cohorts to further demonstrate generalizability and robustness across different populations and sequencing platforms.
Response 1: Thank you for this important suggestion. We agree that additional external validation across independent cohorts, geographic populations and sequencing platforms would further strengthen assessment of model generalizability and robustness. In the present study, external validation was performed using an independently curated HBVdb-derived cohort processed through a separate preprocessing workflow and demonstrating a substantially different resistance pathway prevalence compared with the NCBI-derived development dataset. The intention of this validation strategy was to evaluate methodological reproducibility and transportability across independently processed public sequence repositories. We additionally explored the feasibility of incorporating further publicly available HBV resistance cohorts. However, many currently accessible HBV sequence resources are substantially derived from overlapping GenBank submissions, which may introduce uncertain dataset overlap and limit true independence between development and validation cohorts despite originating from distinct repositories or publications. To avoid introducing potential non-independence or poorly documented overlap, we retained HBVdb as the external validation cohort and clarified this limitation in the revised manuscript. Accordingly, we added the following sentence in the Limitations paragraph of Discussion section (page 10):
Additional validation across geographically diverse populations, alternative sequencing platforms, and clinically annotated cohorts would further strengthen assessment of model transportability and robustness.

We believe this clarification better contextualizes the scope of the current validation framework and future directions for broader external validation.

Comment 2: As this is essentially a diagnostic prediction study, the manuscript should report additional clinically relevant performance metrics, including positive predictive value (PPV) and negative predictive value (NPV), especially given the low prevalence of entecavir resistance in the external validation cohort.
Response 2: Thank you for this important suggestion. We agree that predictive value metrics are particularly relevant in the context of the marked prevalence shift observed in the external validation cohort. In response, we expanded the reported threshold-dependent performance metrics to explicitly include positive predictive value (PPV) and negative predictive value (NPV). These metrics are now incorporated in the Methods section 2.3 (page 4), Supplementary Table S3, and corresponding supplementary methods descriptions.

We additionally added the following clarification in the Results section 3.3 (page 7):
Despite the low prevalence of the predefined resistance pathway in the external cohort, negative predictive value remained high across evaluated thresholds, while positive predictive value varied according to the selected operating threshold (Supplementary Table S3).

We believe these additions improve the clinical interpretability of threshold-based performance under conditions of substantial outcome imbalance.

Comment 3: The authors should clearly describe the strategy and rationale used for selecting the probability cutoff threshold. It is important to explain whether the cutoff was optimized using Youden index, calibration analysis, clinical sensitivity/specificity balance, or other predefined rules.
Response 3: Thank you for this important suggestion. We agree that clearer explanation of the probability threshold selection strategy improves interpretability of the proposed framework. In response, we revised the Methods section 2.3 (pages 4-5) to more explicitly describe the rationale and procedures used for threshold selection. Specifically, we clarified that decision thresholds were selected using prespecified discrimination-oriented criteria, including maximization of the Youden index and F1 score. The following text was added:
Threshold-based performance was summarized using sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), precision, accuracy, F1 score, and Matthews correlation coefficient, with decision thresholds selected using prespecified discrimination-oriented criteria, including maximization of the Youden index and F1 score. Threshold performance was subsequently evaluated using confusion matrices, calibration analyses, and threshold-dependent classification metrics across internal and external validation datasets.

We additionally clarified in the Results section (page 6) that:
The identical threshold identified by both the Youden index and F1 optimization procedures supported selection of a stable operating threshold balancing sensitivity, specificity, and precision within the internally validated dataset.

Furthermore, the Supplementary Methods were revised to clarify that threshold optimization procedures were predefined prior to external validation and subsequently applied unchanged to the HBVdb external cohort. We believe these revisions improve transparency regarding operating-threshold selection and interpretation.

Comment 4: Chronic HBV infection is highly heterogeneous, including differences in viral genotype, disease phase, viral load, cirrhosis status, and prior antiviral exposure. The authors should investigate whether the predictive performance of the model varies across different HBV infection subgroups and perform subgroup analyses where possible.
Response 4: Thank you for this important observation. We agree that chronic HBV infection is biologically and clinically heterogeneous, including variation in viral genotype, disease phase, antiviral exposure history, viral load dynamics, and underlying liver disease status. However, the publicly available sequence repositories used in the present study contained limited standardized clinical metadata, which restricted robust subgroup analyses according to disease phase, viral load, cirrhosis status, treatment history, or other clinically relevant host-level characteristics. To avoid introducing potentially biased or poorly characterized subgroup analyses based on incomplete metadata, we did not perform additional stratified analyses in the current study. To clarify this limitation, we added the following statements in the Limitations of Discussion section (page 11):
Publicly available HBV sequence repositories contain limited standardized clinical metadata, restricting robust subgroup analyses according to disease phase, viral load, cirrhosis status, treatment history, or other clinically relevant host-level characteristics. 

and:
Additional validation across geographically diverse populations, alternative sequencing platforms, and clinically annotated cohorts would further strengthen assessment of model transportability and robustness.

We believe these revisions better contextualize the scope and limitations of the current repository-based framework.

Comment 5: The discussion section should more comprehensively address how this model could be translated into clinical practice, including its potential integration into sequencing workflows, antiviral treatment selection, resistance monitoring, and future clinical decision-support systems.
Response 5: Thank you for this important suggestion. We agree that further clarification regarding potential translational applications improves the clinical context of the proposed framework. The revised Discussion section already emphasizes that the present model is not intended to replace established rule-based interpretation systems or independently guide therapeutic decisions, but rather to provide a standardized probabilistic framework for reproducible sequence-based resistance interpretation across heterogeneous datasets. To further clarify potential future applications, we additionally incorporated the following statement in the Discussion section (page 10):
In future clinical or laboratory settings, similar frameworks could potentially be incorporated into sequencing-analysis pipelines or decision-support systems to support standardized resistance interpretation alongside established expert-guideline approaches.

We believe this addition better contextualizes how probabilistic resistance interpretation frameworks may eventually complement sequencing workflows, resistance surveillance, and future clinical decision-support applications.

Reviewer 4 Report

Comments and Suggestions for Authors

Dear Authors,

I have several major concerns regarding the conceptual framing and interpretation of the study.

The manuscript is clearly structured and transparently reported.
However, the main modeling task appears circular.
The outcome is defined as a genotypic proxy based on known lamivudine- and entecavir-associated RT mutations, while the predictors are derived from the same predefined RT positions. As a result, the model mainly reconstructs the rule used to define the endpoint, rather than predicting an independent biological or clinical outcome.

This substantially limits the interpretation of the reported near-perfect performance. The high AUC, calibration, and threshold-based metrics are expected under this deterministic setup and should not be presented as evidence of added predictive value or clinical utility.

External validation on HBVdb is useful as a test of reproducibility across databases, but it does not resolve the above issue. The model is still evaluated against the same genotypic rule, not against phenotypic entecavir resistance, virological breakthrough, treatment failure, or longitudinal clinical outcomes.

I recommend that the authors substantially reframe the manuscript. The work should be presented as a reproducible computational implementation of an established genotypic resistance pathway, rather than as a clinically validated machine-learning model. Claims of novelty, prediction, and clinical utility should be reduced accordingly.

A direct comparison with a simple rule-based classifier using the same mutation definition should also be added. This would clarify whether the machine-learning approach provides any performance or interpretability advantage over the underlying rule.

Author Response

Comment 1: The outcome is defined as a genotypic proxy based on known lamivudine- and entecavir-associated RT mutations, while the predictors are derived from the same predefined RT positions. As a result, the model mainly reconstructs the rule used to define the endpoint, rather than predicting an independent biological or clinical outcome.
Response 1: Thank you for this important conceptual observation. We agree that the outcome definition was based on established resistance-associated substitutions and that the predefined RT feature panel was intentionally selected from prior HBV resistance literature. Accordingly, the framework should not be interpreted as de novo prediction of an independent clinical phenotype. To clarify the scope of the study, we revised the following sections:
1. Introduction (page 2) by adding:
The objective of the present study was therefore not de novo discovery of resistance determinants, but development of a transparent probabilistic framework for standardized interpretation and external validation of predefined resistance-associated sequence patterns.
2. We also revised the Results section (page 7) to clarify interpretation of the internal validation findings:
Given the deterministic relationship between the predefined RT features and the outcome definition, these findings primarily reflect probabilistic reconstruction of a well-characterized genotypic resistance pathway rather than prediction of an independent clinical phenotype.
3. In addition, the Discussion section (page 9) was revised to further emphasize the methodological and externally validated nature of the framework. The following text was added:
Rather, it probabilistically formalizes and externally validates an already recognized resistance pathway within a transparent and reproducible analytical framework.

Comment 2: This substantially limits the interpretation of the reported near-perfect performance. The high AUC, calibration, and threshold-based metrics are expected under this deterministic setup and should not be presented as evidence of added predictive value or clinical utility.
Response 2: Thank you for this important clarification. We agree that the near-perfect internal validation performance should be interpreted within the context of the predefined biological and methodological structure of the framework. As noted by the reviewer, the reported discrimination and calibration metrics reflect probabilistic reconstruction of a predefined genotypic resistance pathway rather than prediction of an independent clinical phenotype. To address this point, we revised the Results and Discussion section (page 10) to further clarify the interpretation and scope of the reported performance metrics. The following sentence was added in the Discussion section:
Accordingly, the reported discrimination, calibration, and threshold-based metrics should be interpreted as measures of reproducibility and probabilistic consistency within a predefined genotypic framework rather than direct evidence of independent clinical predictive utility.

We believe this revision better contextualizes the interpretation of the reported model performance.

Comment 3: External validation on HBVdb is useful as a test of reproducibility across databases, but it does not resolve the above issue. The model is still evaluated against the same genotypic rule, not against phenotypic entecavir resistance, virological breakthrough, treatment failure, or longitudinal clinical outcomes.
Response 3: Thank you for this important observation. We agree that external validation against an independent HBVdb-derived cohort does not substitute for validation against phenotypic resistance measurements, virological breakthrough, treatment failure, or longitudinal clinical outcomes. The intention of the external validation procedure was to evaluate reproducibility and transportability of the probabilistic framework across independently processed public sequence repositories rather than to establish direct clinical predictive validity. To clarify this distinction, we revised the Limitations paragraph in Discussion section (page 11) by adding the following sentence:
Consequently, external validation in HBVdb should be interpreted primarily as assessment of methodological reproducibility and transportability across independently processed sequence repositories rather than validation against clinical treatment outcomes.

We believe this clarification better defines the intended scope and interpretation of the external validation analyses.

Comment 4: I recommend that the authors substantially reframe the manuscript. The work should be presented as a reproducible computational implementation of an established genotypic resistance pathway, rather than as a clinically validated machine-learning model. Claims of novelty, prediction, and clinical utility should be reduced accordingly.
Response 4: Thank you for this important recommendation. In response, we revised multiple sections of the manuscript to more clearly define the methodological scope and intended interpretation of the framework. Specifically, we reduced language implying independent clinical prediction or discovery of novel resistance mechanisms and instead emphasized the role of the framework as a transparent, probabilistic, and externally validated computational implementation of a predefined genotypic resistance pathway.

Comment 5: A direct comparison with a simple rule-based classifier using the same mutation definition should also be added. This would clarify whether the machine-learning approach provides any performance or interpretability advantage over the underlying rule.
Response 5: Thank you for this important suggestion. To further contextualize the relationship between the probabilistic framework and the predefined resistance mutation criteria, we performed an additional comparator analysis using a deterministic rule-based classifier directly implementing the same mutation definition used for endpoint construction. As expected, the rule-based classifier demonstrated near-equivalent classification performance and complete concordance during internal validation, reflecting the deterministic structure of the predefined genotypic endpoint. These findings support the interpretation that the primary contribution of the present framework lies not in replacing established biological rules, but in providing a transparent probabilistic implementation enabling calibration assessment, threshold-based evaluation, and external validation across independent datasets. We added the following sentence in the Results section 3.2 (page 6):
Direct comparison with a deterministic rule-based classifier using the same predefined resistance mutation criteria demonstrated near-equivalent classification performance and complete concordance in internal validation, consistent with the deterministic construction of the endpoint definition (Supplementary Table S3).
Moreover, we also added the following statement in the Discussion section (page 9):
Given that the outcome definition was directly derived from predefined resistance-associated substitutions, a deterministic rule-based classifier based on the same mutation criteria would be expected to produce near-equivalent classification results by construction.

Detailed comparator metrics are now provided in Supplementary Table S3.

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

The revised manuscript has made a great improvement. I have no more comments and recommends.

Author Response

Comment 1: The revised manuscript has made a great improvement. I have no more comments and recommends.

Author response: We sincerely thank the reviewer for the careful evaluation of our manuscript and for the very positive assessment of the revised version. 

Reviewer 4 Report

Comments and Suggestions for Authors

Authors,

Please further soften the title, abstract, and conclusions, and avoid language suggesting clinical utility or net clinical benefit, since the model primarily provides a probabilistic reconstruction of a predefined genotypic resistance rule rather than independent clinical prediction.

Author Response

Comment 1: Please further soften the title, abstract, and conclusions, and avoid language suggesting clinical utility or net clinical benefit, since the model primarily provides a probabilistic reconstruction of a predefined genotypic resistance rule rather than independent clinical prediction.
Author response: We sincerely thank the reviewer for this important and constructive comment. In response, we carefully revised the manuscript to further soften the framing of the study and to avoid language implying independent clinical prediction, clinical utility, or net clinical benefit. Throughout the revised manuscript, the work is now more explicitly presented as a transparent probabilistic reconstruction of a predefined genotypic resistance framework rather than an independent clinical prediction model. The following revisions were made:
1. The title was revised from: “Externally Validated Machine Learning Identification of Entecavir Resistance Pathways in HBV Using Independent Public Repositories” to: “Externally Validated Probabilistic Modeling of a Predefined Entecavir Resistance Pathway in HBV Using Independent Public Repositories”.
2. Abstract revisions:
- In the Background section, the sentence: “We aimed to develop and externally validate a transparent sequence-based machine learning framework for identifying the entecavir resistance pathway in HBV polymerase.” was revised to: “We aimed to develop and externally validate a transparent probabilistic framework for reconstructing a predefined entecavir resistance pathway from HBV polymerase sequences.”
- In the Results section, the phrase: “confirmed robust generalization” was revised to: “demonstrated reproducible performance across repositories”.
- In the Conclusions section, the concluding statement was revised to further emphasize the methodological and probabilistic nature of the framework and to avoid implications of independent clinical prediction.

3. Methods section (Section 2.4): The sentence: “Decision curve analysis was additionally performed to evaluate potential clinical utility...” was revised to: “Decision curve analysis was performed as an exploratory assessment of threshold-dependent classification behavior within the predefined genotypic framework.”

4. Decision curve analysis terminology: Throughout the manuscript, the phrase “net clinical benefit” was replaced with the more methodologically appropriate term “decision-analytic behavior”.

5. Results section revisions:
- In Section 3.2: The sentence: “Decision curve analysis indicated a consistent net benefit...” was revised to: “Decision curve analysis demonstrated stable threshold-dependent classification behavior across a range of probability thresholds.”
- In Section 3.3: The phrase: “supporting the model’s transportability” was revised to: “supporting reproducibility of probabilistic classification behavior across repositories with differing pathway prevalence.”

6. Discussion section:
Several statements in the Discussion were revised to further clarify the intended interpretation and scope of the model:
- The sentence: “A clinically relevant aspect of this work is its potential to complement existing genotypic resistance interpretation” was revised to: “A methodological strength of this work is the standardized probabilistic representation of established genotypic resistance criteria.”
- The sentence beginning with: “In future clinical or laboratory settings...” was removed.
- The term: “decision-support systems” was replaced with: “sequence-analysis workflows”.
- The sentence: “probabilistic outputs may be useful for prioritizing sequences for expert review, harmonizing interpretation across large public datasets, or flagging patterns that require cautious interpretation.” was revised to: “probabilistic outputs primarily provide a quantitative representation of predefined genotypic criteria and allow formal evaluation of calibration and threshold-dependent behavior across datasets.”

7. Conclusions section:
The Conclusions section was comprehensively revised to more clearly emphasize that the framework represents a probabilistic formalization of predefined genotypic resistance criteria rather than an independent predictor of clinical outcomes. The revised conclusion now states:
“In summary, this study presents a transparent and externally validated probabilistic framework for reconstructing a predefined entecavir resistance pathway from HBV polymerase sequences derived from publicly available databases. The observed performance primarily reflects the biologically structured nature of the predefined genotypic resistance definition and the consistency of the analytical pipeline across independent repositories. Rather than serving as an independent predictor of clinical resistance outcomes, the framework provides a reproducible probabilistic formalization of established resistance-associated sequence patterns. These findings support the feasibility of applying transparent and externally validated analytical approaches to standardized sequence-based resistance interpretation.”

We believe these revisions substantially improve the clarity and positioning of the manuscript and fully address the reviewer’s concern.