Multi-Criteria Decision Analysis in Drug Discovery

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Your article presents a novel method to integrate Multicriteria Decision Analysis (MCDA) with the AIDDTM generative chemistry framework. Discovery of drugs could advance chemical space exploration. Although the VIKOR approach for compound ranking improves escape the constraints of conventional Pareto optimization, its originality is greatly overdone. The study has to be able to clearly differentiate its framework from already applied approaches such as TOPSIS and SMAA-TOPSIS. The paper also addresses computing requirements and scale issues of applying MCDA in generative chemistry. It is oversimplified to assume that the distribution of weights and the usage of VIKOR always yield the best results, as it ignores potential negative consequences, such as erroneous exploration resulting from an unsuitable weight distribution.

You conclude that this strategy "significantly increases efficiency and reduces overall costs." However, the lack of a comprehensive cost-benefit analysis or computational benchmarks casts doubt on this assertion. In addition to the need for careful mathematical investigation, the overemphasis on VIKOR overshadows the importance of integrating MCDA with generative chemistry. Also, the validity is weakened by the lack of a comparative study using other MCDA methods, like TOPSIS or AHP. Some of the innovative assertions regarding large language models and breakthroughs in machine learning need further evidence since there are few references for them. Though additional cases with conflicting goals would have added more complexity, the Pregnane X Receptor (PXR) example shows sufficiently the actual use of the strategy. Although the survey presents interesting ideas, it would benefit from a more balanced perspective that emphasizes general effects instead of disproportionately focusing on particular methods like VIKOR, and examines pragmatic limits by including comparison studies.

In order to justify the choice of VIKOR as the recommended approach, I advise a more extensive comparison study of MCDA approaches so enhancing the comprehensiveness, practicality, and emphasis of this research. Although a complete justification of VIKOR's better applicability in this context can greatly improve the case, a detailed description of VIKOR stresses its efficacy, and adding alternative methodologies such TOPSIS or AHP will help to support this. The study would benefit from addressing pragmatic concerns including processing demands, scalability, and the probable risks of providing erroneous weights to goals, therefore misguiding the optimization process. Including actual benchmarks or a cost-benefit analysis would help to validate claims of efficiency and cost control. We might focus more on the mathematical details to pay more attention to the most important effects of mixing MCDA with generative chemistry, especially when we think about problems and uses in drug development that happen in the real world. More citations would help to improve the scientific legitimacy of assertions about the transformational power of generative models and machine learning. The case studies would be more useful if they were more in-depth and looked at a wider range of scenarios, such as those with competing goals or unexpected failures in optimization. Instead of emphasizing the relevance of a single instrument, the publication should highlight the framework as a complementary approach that offers strategies, thereby reflecting the complexity of real drug development. By addressing these elements, the study could be comprehensive and successful, as it would strike a fair balance between theoretical breakthroughs and practical applicability.

Author Response

Reviewers’ remark:

Your article presents a novel method to integrate Multicriteria Decision Analysis (MCDA) with the AIDDTM generative chemistry framework. Discovery of drugs could advance chemical space exploration. Although the VIKOR approach for compound ranking improves escape the constraints of conventional Pareto optimization, its originality is greatly overdone. The study has to be able to clearly differentiate its framework from already applied approaches such as TOPSIS and SMAA-TOPSIS. The paper also addresses computing requirements and scale issues of applying MCDA in generative chemistry. It is oversimplified to assume that the distribution of weights and the usage of VIKOR always yield the best results, as it ignores potential negative consequences, such as erroneous exploration resulting from an unsuitable weight distribution.

Authors' response:

Thank you for this thorough insight into our work. We are aware that the VIKOR method itself has been invented years ago, but the application of this MCDA approach in the drug design, especially in the context of \textit{de novo} design it seems to be novel. At least we have not found any literature item that would thoroughly expand on this topic. To address the issue of potential erroneous/unwanted exploration of chemical space, we introduced one more, probably the most sophisticated use case involving the oral bioavailability optimization of Cyprofloxacin (Section 3.3). We carefully inspected the benefits coming from the MCDA pruning, and compared this to the tradition Pareto-based approach.

Reviewers’ remark:

You conclude that this strategy "significantly increases efficiency and reduces overall costs." However, the lack of a comprehensive cost-benefit analysis or computational benchmarks casts doubt on this assertion. In addition to the need for careful mathematical investigation, the overemphasis on VIKOR overshadows the importance of integrating MCDA with generative chemistry. Also, the validity is weakened by the lack of a comparative study using other MCDA methods, like TOPSIS or AHP. Some of the innovative assertions regarding large language models and breakthroughs in machine learning need further evidence since there are few references for them. Though additional cases with conflicting goals would have added more complexity, the Pregnane X Receptor (PXR) example shows sufficiently the actual use of the strategy. Although the survey presents interesting ideas, it would benefit from a more balanced perspective that emphasizes general effects instead of disproportionately focusing on particular methods like VIKOR, and examines pragmatic limits by including comparison studies.

Authors' response:

Thank you for this comment. To address these remarks we implemented and integrated with AIDD suggested by the reviewer additional MCDA approach: TOPSIS. The comparative study was illustrated with the molecular optimization of the ciprofloxacin. The optimization process used a multi-criteria decision analysis (MCDA) framework, with key objectives being 3D structural similarity, fraction bioavailable (\%Fb), and synthetic difficulty. Among the MCDA pruning methods tested, VIKOR and TOPSIS provided superior results compared to the Pareto approach, with VIKOR offering the best balance between structural similarity and bioavailability. The optimized compound retained the quinolone scaffold while significantly enhancing pharmacokinetic properties, as demonstrated by plasma concentration curves. The ability to weight objectives in MCDA pruning was critical there. We also discussed an issue of potential danger related to the improper use of weighting schemes in the MCDA-supported molecular optimization.

Reviewers’ remark:

In order to justify the choice of VIKOR as the recommended approach, I advise a more extensive comparison study of MCDA approaches so enhancing the comprehensiveness, practicality, and emphasis of this research. Although a complete justification of VIKOR's better applicability in this context can greatly improve the case, a detailed description of VIKOR stresses its efficacy, and adding alternative methodologies such TOPSIS or AHP will help to support this. The study would benefit from addressing pragmatic concerns including processing demands, scalability, and the probable risks of providing erroneous weights to goals, therefore misguiding the optimization process. Including actual benchmarks or a cost-benefit analysis would help to validate claims of efficiency and cost control. We might focus more on the mathematical details to pay more attention to the most important effects of mixing MCDA with generative chemistry, especially when we think about problems and uses in drug development that happen in the real world.

Authors' response:

Thank you for this valuable comment, and a motivation to pursue additional efforts resulting in the development and integration of additional MCDA method, i.e. TOPSIS. To address this remark we designed additional use case involving the molecular optimization of cyprofloxacin, where the key objectives were oral bioavailability and 3D similarity to the reference crystal structure. A comparative study evaluated the performance of VIKOR and TOPSIS pruning methods, both of which outperformed the Pareto approach by better balancing the trade-offs between objectives. VIKOR demonstrated a superior ability to balance structural similarity and bioavailability, identifying 3 compounds within the top 90th percentiles of both objectives, while TOPSIS identified 7 such compounds but with slightly less consistency in trade-offs. The study highlighted the importance of regret and utility measures in VIKOR, which provided a nuanced optimization approach compared to the Euclidean distance-based metrics of TOPSIS. We believe this addition enhances the overall contribution of the paper and offers a clearer insight into the practical applications of our approach. We sincerely appreciate your valuable feedback.

Reviewers’ remark:

More citations would help to improve the scientific legitimacy of assertions about the transformational power of generative models and machine learning. The case studies would be more useful if they were more in-depth and looked at a wider range of scenarios, such as those with competing goals or unexpected failures in optimization. Instead of emphasizing the relevance of a single instrument, the publication should highlight the framework as a complementary approach that offers strategies, thereby reflecting the complexity of real drug development. By addressing these elements, the study could be comprehensive and successful, as it would strike a fair balance between theoretical breakthroughs and practical applicability.

Authors' response:

Thank you for the comment. We added a substantial number of additional literature items, also these directing the readers to recent review papers discussing thoroughly the history and recent advances of various generative chemistry approaches. We hope that it makes the study more consistent and provides broad overview of techniques applied in de-novo drug design.

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript presents the application of a Multicriteria Decision Aiding (MCDA) method within the framework of AI-driven Drug Design (AIDDTM) technology. The study addresses the challenges of drug development by evaluating multiple criteria to identify optimal drug candidates with favorable pharmacokinetic and pharmacodynamic profiles, or to eliminate compounds that are unlikely to meet therapeutic safety standards. Specifically, the research investigates the effectiveness of the VIKOR approach in steering chemical space exploration towards compounds with the desired properties.

Moreover, I would expect the authors to study and probe more test-cases for a full article. Practically the authors work on only two case-studies (benzimidazole sulfonomide & Sphingosine 1-Phosphate Receptor) out of one pdb reference (5a86).

Author Response

Reviewers’ remark:

The manuscript presents the application of a Multicriteria Decision Aiding (MCDA) method within the framework of AI-driven Drug Design (AIDDTM) technology. The study addresses the challenges of drug development by evaluating multiple criteria to identify optimal drug candidates with favorable pharmacokinetic and pharmacodynamic profiles, or to eliminate compounds that are unlikely to meet therapeutic safety standards. Specifically, the research investigates the effectiveness of the VIKOR approach in steering chemical space exploration towards compounds with the desired properties.

Moreover, I would expect the authors to study and probe more test-cases for a full article. Practically the authors work on only two case-studies (benzimidazole  sulfonomide \& Sphingosine 1-Phosphate Receptor) out of one pdf reference (5a86).

Authors' response:

Thank you for your valuable feedback and for suggesting the inclusion of additional use case to illustrate the application of our optimization approach. In response, we have added a detailed use case centered on ciprofloxacin, a well-known broad-spectrum fluoroquinolone antibiotic (Section 3.3).

This use case demonstrates the application of our multi-criteria decision analysis (MCDA) framework to optimize ciprofloxacin's pharmacokinetic properties, particularly its oral bioavailability. By focusing on objectives such as 3D similarity, fraction bioavailable (\%Fb), and synthetic difficulty, we highlight the practical utility of our methodology in a real-world drug design scenario. Structural modifications to ciprofloxacin led to significant improvements in hydrophilicity and permeability, resulting in an increase in oral bioavailability by over 30\%.

As part of this use case, we also conducted a comparative study of the VIKOR and TOPSIS pruning methods. Both methods significantly outperformed the Pareto approach, with VIKOR demonstrating nuanced trade-off optimization through its regret and utility measures, identifying 3 compounds in the top 90th percentiles for both 3D similarity and \%Fb. TOPSIS, using Euclidean distance-based metrics, identified 7 such compounds with a slightly different optimization profile. These results showcase the strengths of both methods in effectively navigating the chemical space.

This addition to the manuscript not only aligns with your request but also enriches the discussion by providing a concrete example of the methodology's capabilities. We hope this strengthens the overall contribution of the paper and provides a clearer perspective on the practical implications of our approach. Thank you again for your valuable input.

Reviewer 3 Report

Comments and Suggestions for Authors

This paper describes the integration of the VIKOR multicriteria decision analysis (MCDA) method into the AI-powered Drug Design (AIDDTM) technology to enhance drug discovery by allowing the navigation of chemical spaces and enabling user-defined prioritization of molecular properties.

Some comments and suggestions:

1. Did the authors perform a comparative analysis on this novel method against traditional approaches  in terms of success rates or efficiency metrics? This would help strengthen the validity of the claims on the effectiveness of this method.

2. It would be good if the authors could expand the potential biases introduced by subjective weighting of objectives in MCDA in the discussion section.

3. Could the authors provide more detailed information on the specific methods used for QSAR modeling, data curation, and also the rationale for the selected test datasets. This would provide greater clarity in the results presented.

4. Chemical structures presented in Figure 5 and Figure 9 should probably be redrawn as the labels are too small to be seen clearly, especially the side chains. 

 

Comments on the Quality of English Language

Generally, the article is written well. Some suggestions for improvements:

1. Some sentences should be simplified to improve readability. For example, avoiding excessive subclauses.

2. The flow between sections could be improved by using transition sentences that summarize key findings or implications.

Author Response

Reviewers’ remark:

This paper describes the integration of the VIKOR multicriteria decision analysis (MCDA) method into the AI-powered Drug Design (AIDDTM) technology to enhance drug discovery by allowing the navigation of chemical spaces and enabling user-defined prioritization of molecular properties.

Some comments and suggestions:

1. Did the authors perform a comparative analysis on this novel method against traditional approaches in terms of success rates or efficiency metrics? This would help strengthen the validity of the claims on the effectiveness of this method.

Authors' response:

Thank you for your thoughtful comment. In response to your query, we would like to emphasize that in each use case presented in the manuscript - namely crystal structure rediscovery, ligand optimization for low PXR activity, and a new study oriented on oral bioavailability optimization of ciprofloxacin - we performed a comparative analysis of the MCDA-based pruning schemes against the default Pareto-based optimization approach, which we consider as the traditional method.

Notably, the last use case which was added  in the frame of current revision was specifically oriented toward a comparative study. It was designed to analyze the performance of MCDA-based pruning schemes, particularly VIKOR and TOPSIS, in relation to the traditional Pareto-based approach. This case highlighted the superior ability of the MCDA methods to balance trade-offs between key objectives, such as 3D similarity and fraction bioavailable (%Fb), thereby validating the effectiveness and efficiency of the proposed methods.

Reviewers’ remark:

2. It would be good if the authors could expand the potential biases introduced by subjective weighting of objectives in MCDA in the discussion section.

Authors' response:

Thank you for this comment. Indeed, the process of assigning weights to objective functions during molecular optimization plays a crucial role in shaping the resulting compound population. While this approach is intended to reflect user priorities, improper application of MCDA pruning can cause the algorithm to explore regions of the chemical space that do not align with the intended objectives. Conversely, when applied thoughtfully, MCDA pruning - particularly using the VIKOR method - proves to be a highly effective optimization strategy, directing the exploration of chemical space toward areas that closely align with the desired outcomes. To address this, we have extended Section 3.3 with a paragraph reflecting this perspective.

Reviewers’ remark:

3. Could the authors provide more detailed information on the specific methods used for QSAR modeling, data curation, and also the rationale for the selected test datasets. This would provide greater clarity in the results presented.

Authors' response:

Thank you for this comment. We extended the description at the beginning of Section 3.2 (in yellow) uncovering the details related to data preparation phase.

Reviewers’ remark:

4. Chemical structures presented in Figure 5 and Figure 9 should probably be redrawn as the labels are too small to be seen clearly, especially the side chains.

Authors' response:

Thank you for this comment. All structures have been redrawn such that the labels are large enough.

Reviewers’ remark:

Generally, the article is written well. Some suggestions for improvements:\newline

1. Some sentences should be simplified to improve readability. For example, avoiding excessive subclauses.
2. The flow between sections could be improved by using transition sentences that summarize key findings or implications.

Authors' response:

Thank you for these comments. The entire manuscript has been reviewed by English native speakers with technical expertise.

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

the authors have adequately amended the manuscript, so I recommend publication.