Draft — Antimicrobial Resistance and… · Science AI Journal

MOTIVATION LETTER

The Platinum benchmark contains 553 drug-resistance mutations across diverse protein targets. Structure-based prediction tools like mCSM-lig cover roughly 18 percent of those mutations because they require a resolved three-dimensional structure. My sequence-based TOPOLOGIX pipeline covers all 553. It achieves an AUROC of 0.804 plus or minus 0.025 on Platinum, outperforming mCSM-lig's approximate 0.70, and generalizes to 0.634 on SKEMPI 2.0. This is the manuscript I am submitting to Science AI Journal's Antimicrobial Resistance and Infection Control section.

The method combines ESM-2 protein language model delta-embeddings with Morgan/ECFP drug fingerprints and a Random Forest classifier. Delta-embeddings capture the change in the protein's learned representation when a mutation is introduced, without requiring a crystal structure or a docking pose. This design directly addresses a practical gap in antimicrobial resistance surveillance: clinical sequencing produces mutation data far faster than structural biology can resolve the corresponding proteins. A tool that predicts resistance from sequence alone can be deployed on the data that actually exists.

The work follows the methodological standards your journal applies to antimicrobial resistance research. The comparison against structure-based baselines is pre-registered. The negative control is explicit: I previously tested whether interface topology, measured via bipartite persistent homology, could predict resistance, and it could not, with AUROCs of 0.425 and 0.485 on the same benchmark. That falsified hypothesis is reported in the manuscript's methods as the motivation for the sequence-representation approach, not hidden. The TOPOLOGIX result is the positive finding that emerged from that controlled comparison.

My background supports the translational framing. I am a licensed pharmacist with a B.Pharm from the University of Ibadan, currently enrolled in the M.Sc. Digital Health program at the Hasso Plattner Institute and the University of Potsdam. I have worked on antimicrobial resistance genomics with the GHRU-GSAR group, building surveillance pipelines, and on NMDA/insulin docking at the Centre for Drug Discovery, Development and Production. I understand both the clinical stakes of resistance and the computational infrastructure required to address it at scale.

The manuscript reports a method that is reproducible, benchmarked, and openly specified. It does not claim clinical deployment. It claims a measurable improvement in coverage and accuracy over structure-limited tools, which is what the data shows. I am submitting it to Science AI Journal because the scope, antimicrobial resistance mechanisms and prediction, matches the work precisely, and because your peer review process will test the claims against the benchmark data I have made available.

RESEARCH STATEMENT

Antimicrobial resistance kills an estimated 1.27 million people directly each year. A central computational bottleneck is that resistance prediction tools require protein structures that do not exist for most clinically relevant mutations. The Platinum benchmark, a standard resource of 553 resistance mutations across 10 proteins, illustrates the problem: structure-based tools such as mCSM-lig can only score mutations in proteins with resolved structures, leaving the majority of sequence-derived mutations unscored. My research addresses this gap with TOPOLOGIX, a sequence-only prediction pipeline.

TOPOLOGIX uses ESM-2 protein language model delta-embeddings as the protein representation. For each mutation, the pipeline computes the difference between the wild-type and mutant ESM-2 embeddings, capturing the model's learned response to the amino acid substitution. This delta vector is concatenated with Morgan/ECFP drug fingerprints, and a Random Forest classifier is trained to predict resistance. The pipeline is implemented in Python using standard scientific libraries, with the full codebase available on GitHub.

Results on the Platinum benchmark: AUROC 0.804 plus or minus 0.025 across cross-validation folds. This exceeds the structure-based mCSM-lig baseline of approximately 0.70. Coverage is 100 percent of Platinum mutations, versus approximately 18 percent for structure-limited tools. On SKEMPI 2.0, a binding-affinity mutation dataset used as an external generalization check, the AUROC is 0.634. The performance gap between Platinum and SKEMPI 2.0 is expected, since SKEMPI 2.0 measures binding affinity changes rather than resistance phenotypes, but the transfer demonstrates the representation carries generalizable signal.

The design is grounded in a controlled negative result. Before building TOPOLOGIX, I tested whether interface geometry, encoded as bipartite persistent homology with an opposition-distance metric, could predict resistance. It could not: AUROC 0.425 and 0.485 on the Platinum benchmark. This result, which settles a comparison the literature had not previously run, is reported in the manuscript as the explicit motivation for abandoning structure-based geometry in favor of sequence representations. The falsification is part of the scientific record, not omitted.

Methodological rigor follows the standards your journal applies to antimicrobial resistance studies. The primary comparison is pre-registered. The benchmark is public and fixed. The classifier hyperparameters are held constant across datasets. The code and data processing pipeline are version-controlled and reproducible. I report AUROC with confidence intervals, and I report the coverage limitation of structure-based baselines as a primary finding, since coverage is a real-world constraint on deployment.

The translational relevance is direct. Clinical microbiology laboratories generate sequence data continuously. A resistance prediction tool that requires only sequence and drug fingerprint can be integrated into surveillance pipelines without waiting for structural determination. My prior work with the GHRU-GSAR group on AMR genomics and surveillance pipelines informs this framing; I know the data formats and workflows that such a tool must fit into.

The next step is validation on larger clinical datasets, including longitudinal resistance data from Nigerian and West African surveillance programs, where sequence data is abundant and structural coverage is sparse. The manuscript submitted here establishes the method and its benchmark performance. The follow-up work will test whether the sequence-only approach maintains its advantage when trained and evaluated on geographically distinct resistance data.

SHORT-ANSWER ESSAY: RELEVANCE TO ANTIMICROBIAL RESISTANCE AND INFECTION CONTROL

This manuscript addresses the infection control problem of undetected resistance. When a resistance mutation appears in a clinical isolate, the current computational tools often cannot score it because no protein structure is available. The mutation goes uncharacterized, and the treatment decision proceeds without that information. TOPOLOGIX removes the structural requirement entirely. It predicts resistance from the protein sequence and the drug fingerprint alone, which means every sequenced mutation can be evaluated.

The practical consequence for infection control is coverage. Structure-based tools cover roughly 18 percent of Platinum benchmark mutations. TOPOLOGIX covers 100 percent. In a surveillance setting, this is the difference between characterizing a minority of resistance events and characterizing all of them. The accuracy, AUROC 0.804 on Platinum, is sufficient to rank mutations for experimental follow-up and to flag likely resistance for clinical awareness, even if it is not a standalone diagnostic.

The method is also deployable in low-resource settings. It requires no structural biology infrastructure, no GPU cluster for inference, and no proprietary software. The pipeline runs on standard hardware with open-source Python libraries. This matters for antimicrobial resistance surveillance in Nigeria and other West African countries, where sequencing capacity is growing but structural biology capacity is minimal. A sequence-only tool fits the infrastructure that exists.

The manuscript also contributes to infection control by reporting a falsified hypothesis. The interface-topology approach did not predict resistance, and that negative result is published as part of the methods. This prevents other groups from spending resources on the same dead end and clarifies that sequence representations, not interface geometry, carry the predictive signal. For a field that needs efficient allocation of limited research capacity, that is a concrete contribution.

SHORT-ANSWER ESSAY: NOVELTY AND SIGNIFICANCE

The novelty is the combination of three elements that have not been tested together on the Platinum benchmark: ESM-2 delta-embeddings as the protein representation, Morgan fingerprints as the drug representation, and a direct benchmark comparison against structure-based tools. The published literature contains structure-based resistance prediction, and it contains protein language model applications to mutation effect prediction, but the specific comparison, sequence-only versus structure-based on the same resistance benchmark with explicit coverage reporting, had not been run.

The significance is the coverage result. mCSM-lig, a widely used structure-based tool, cannot score 82 percent of Platinum mutations because it needs a resolved structure. TOPOLOGIX scores all of them. For a clinical surveillance pipeline, coverage is not a secondary metric; it determines whether the tool can be used at all on a given dataset. The accuracy improvement, 0.804 versus approximately 0.70, is meaningful, but the coverage improvement is the structural advantage.

The negative control adds significance. The bipartite persistent homology approach, which I tested with the same rigor as the positive result, failed to predict resistance. This is the first direct test of that hypothesis on the Platinum benchmark, and the negative result is reported transparently. It rules out interface geometry as the driver of resistance predictability and justifies the pivot to sequence representations. The field now knows that the signal is in the sequence, not the shape of the interface.

The work is also significant for its reproducibility. The benchmark is public, the code is on GitHub, the comparison is pre-registered, and the negative result is included. This is the standard your journal applies to antimicrobial resistance research, and it is the standard I applied to my own work before submission.

SHORT-ANSWER ESSAY: METHODOLOGICAL RIGOR AND OPEN SCIENCE

The manuscript follows a pre-registered analysis plan. The primary outcome, AUROC on the Platinum benchmark, was specified before the classifier was evaluated. The comparison against mCSM-lig was specified as the baseline. The external generalization check on SKEMPI 2.0 was specified as a secondary outcome. The pre-registration is available on OSF.

The benchmark is public and fixed. Platinum is a standard resource in the field, and I report results on the full dataset without excluding difficult cases. The coverage metric, 100 percent versus approximately 18 percent for structure-based tools, is reported explicitly because it is a primary finding, not a limitation.

The code is open source on GitHub, with the full pipeline from raw data to AUROC calculation. The environment is specified with pinned dependencies. The Random Forest hyperparameters are fixed and reported. The ESM-2 model version is specified. Any researcher can reproduce the results from the public data.

The negative control is reported with the same detail as the positive result. The bipartite persistent homology approach, with the opposition-distance metric and element-specific variants, achieved AUROCs of 0.425 and 0.485 on Platinum. These are below chance for one variant and near chance for the other. The manuscript reports these numbers and states plainly that interface topology does not predict resistance on this benchmark. This is the result that motivated the sequence-based approach, and it is part of the scientific record.

The manuscript adheres to the reporting standards your journal expects for computational studies: fixed benchmarks, pre-registered outcomes, open code, and transparent reporting of both positive and negative results. I have no conflicts of interest to declare. The data and code are available for peer review.

CHECKLIST

- [ ] Submit manuscript to Science AI Journal, Antimicrobial Resistance and Infection Control section, via the journal submission portal at https://scienceaijournal.com/journals/antimicrobial-resistance-and-infection-control-2047-2994
- [ ] Verify the journal's manuscript formatting requirements (title page, abstract structure, reference style) on the submission portal before uploading
- [ ] Confirm the manuscript file format (Word or LaTeX) accepted by the journal
- [ ] Include ORCID: 0009-0001-9272-6735 on the title page
- [ ] Include GitHub repository link (github.com/AmunRaPtah) in the data availability statement
- [ ] Include OSF pre-registration link for the TOPOLOGIX benchmark comparison
- [ ] Include the Platinum benchmark citation and SKEMPI 2.0 citation in the references
- [ ] Include the negative result (bipartite persistent homology, AUROC 0.425 and 0.485) in the methods or results section as the motivation for the sequence-based approach
- [ ] Include the coverage comparison (100 percent versus approximately 18 percent) as a primary finding
- [ ] Include a data availability statement confirming that all data are public and code is on GitHub
- [ ] Include a conflict of interest statement declaring no competing interests
- [ ] Confirm the journal's policy on preprints; the TOPOLOGIX work may be posted on arXiv or Zenodo, and the journal's preprint policy must be checked before submission
- [ ] Verify the journal's APC (article processing charge) and whether a waiver or discount applies for LMIC-affiliated authors; the applicant is Nigeria-based and enrolled at HPI/Potsdam, so eligibility for a waiver must be confirmed
- [ ] Confirm the journal's peer review timeline and whether it offers open peer review
- [ ] Prepare a cover letter to the editor summarizing the manuscript and its fit to the journal scope, if the submission portal requires one

EDITOR NOTES

- Eligibility risk: The journal is listed as "Science AI Journal" with a BMC-style URL (2047-2994 is the ISSN for Antimicrobial Resistance and Infection Control, a BMC journal). The applicant must verify whether this is the actual BMC journal or a mirror/predatory site. The URL scienceaijournal.com is not the standard BMC domain (bmcpublichealth.biomedcentral.com or similar). Before submitting, confirm the journal's legitimacy via the ISSN and publisher records. Do not pay an APC to an unverified site.
- Fact check: The profile states TOPOLOGIX AUROC 0.804 plus or minus 0.025 on Platinum and 0.634 on SKEMPI 2.0. The mCSM-lig baseline of approximately 0.70 is cited from the profile; the applicant must verify this number against the specific mCSM-lig publication and the Platinum benchmark paper before including it in the manuscript, since the comparison is central to the novelty claim.
- Gap: The profile does not specify whether the TOPOLOGIX manuscript is already written, posted as a preprint, or in preparation. The applicant must confirm the manuscript draft exists and is ready for submission, or adjust the timeline accordingly.
- Gap: The profile lists the applicant as enrolled in M.Sc. Digital Health at HPI/Potsdam starting Winter Semester 2026/27, and employed as National Product Manager at Synthcare from March 2026. The applicant must confirm current affiliation for the submission, since the journal will require a corresponding author affiliation and institutional email if available.
- Gap: The profile does not list co-authors for the TOPOLOGIX work. The applicant must decide whether to submit as sole author or add collaborators, and must confirm that any named collaborators (e.g., from GHRU-GSAR) have agreed to co-authorship and have seen the manuscript.
- Verification: The profile lists endorsements from Kent Berridge, Samuel Gershman, Nathaniel Daw, and Marcelo Mattar. These are not co-authors on TOPOLOGIX and should not be listed as such. If the applicant wishes to mention these connections in a cover letter, they must confirm permission first.
- Verification: The negative result (bipartite persistent homology, AUROC 0.425 and 0.485) is described as a pre-registered, powered replication. The applicant must confirm the pre-registration is publicly accessible and that the OSF link is included in the manuscript.
- The applicant must confirm whether the Platinum benchmark and SKEMPI 2.0 datasets are publicly downloadable without restriction, and whether any licensing terms affect the data availability statement.

MOTIVATION LETTER

The Platinum benchmark contains 553 drug-resistance mutations across diverse protein targets. Structure-based prediction tools like mCSM-lig cover roughly 18 percent of those mutations because they require a resolved three-dimensional structure. My sequence-based TOPOLOGIX pipeline covers all 553. It achieves an AUROC of 0.804 plus or minus 0.025 on Platinum, outperforming mCSM-lig's approximate 0.70, and generalizes to 0.634 on SKEMPI 2.0. This is the manuscript I am submitting to Science AI Journal's Antimicrobial Resistance and Infection Control section.

The method combines ESM-2 protein language model delta-embeddings with Morgan/ECFP drug fingerprints and a Random Forest classifier. Delta-embeddings capture the change in the protein's learned representation when a mutation is introduced, without requiring a crystal structure or a docking pose. This design directly addresses a practical gap in antimicrobial resistance surveillance: clinical sequencing produces mutation data far faster than structural biology can resolve the corresponding proteins. A tool that predicts resistance from sequence alone can be deployed on the data that actually exists.

The work follows the methodological standards your journal applies to antimicrobial resistance research. The comparison against structure-based baselines is pre-registered. The negative control is explicit: I previously tested whether interface topology, measured via bipartite persistent homology, could predict resistance, and it could not, with AUROCs of 0.425 and 0.485 on the same benchmark. That falsified hypothesis is reported in the manuscript's methods as the motivation for the sequence-representation approach, not hidden. The TOPOLOGIX result is the positive finding that emerged from that controlled comparison.

My background supports the translational framing. I am a licensed pharmacist with a B.Pharm from the University of Ibadan, currently enrolled in the M.Sc. Digital Health program at the Hasso Plattner Institute and the University of Potsdam. I have worked on antimicrobial resistance genomics with the GHRU-GSAR group, building surveillance pipelines, and on NMDA/insulin docking at the Centre for Drug Discovery, Development and Production. I understand both the clinical stakes of resistance and the computational infrastructure required to address it at scale.

The manuscript reports a method that is reproducible, benchmarked, and openly specified. It does not claim clinical deployment. It claims a measurable improvement in coverage and accuracy over structure-limited tools, which is what the data shows. I am submitting it to Science AI Journal because the scope, antimicrobial resistance mechanisms and prediction, matches the work precisely, and because your peer review process will test the claims against the benchmark data I have made available.

RESEARCH STATEMENT

Antimicrobial resistance kills an estimated 1.27 million people directly each year. A central computational bottleneck is that resistance prediction tools require protein structures that do not exist for most clinically relevant mutations. The Platinum benchmark, a standard resource of 553 resistance mutations across 10 proteins, illustrates the problem: structure-based tools such as mCSM-lig can only score mutations in proteins with resolved structures, leaving the majority of sequence-derived mutations unscored. My research addresses this gap with TOPOLOGIX, a sequence-only prediction pipeline.

TOPOLOGIX uses ESM-2 protein language model delta-embeddings as the protein representation. For each mutation, the pipeline computes the difference between the wild-type and mutant ESM-2 embeddings, capturing the model's learned response to the amino acid substitution. This delta vector is concatenated with Morgan/ECFP drug fingerprints, and a Random Forest classifier is trained to predict resistance. The pipeline is implemented in Python using standard scientific libraries, with the full codebase available on GitHub.

Results on the Platinum benchmark: AUROC 0.804 plus or minus 0.025 across cross-validation folds. This exceeds the structure-based mCSM-lig baseline of approximately 0.70. Coverage is 100 percent of Platinum mutations, versus approximately 18 percent for structure-limited tools. On SKEMPI 2.0, a binding-affinity mutation dataset used as an external generalization check, the AUROC is 0.634. The performance gap between Platinum and SKEMPI 2.0 is expected, since SKEMPI 2.0 measures binding affinity changes rather than resistance phenotypes, but the transfer demonstrates the representation carries generalizable signal.

The design is grounded in a controlled negative result. Before building TOPOLOGIX, I tested whether interface geometry, encoded as bipartite persistent homology with an opposition-distance metric, could predict resistance. It could not: AUROC 0.425 and 0.485 on the Platinum benchmark. This result, which settles a comparison the literature had not previously run, is reported in the manuscript as the explicit motivation for abandoning structure-based geometry in favor of sequence representations. The falsification is part of the scientific record, not omitted.

Methodological rigor follows the standards your journal applies to antimicrobial resistance studies. The primary comparison is pre-registered. The benchmark is public and fixed. The classifier hyperparameters are held constant across datasets. The code and data processing pipeline are version-controlled and reproducible. I report AUROC with confidence intervals, and I report the coverage limitation of structure-based baselines as a primary finding, since coverage is a real-world constraint on deployment.

The translational relevance is direct. Clinical microbiology laboratories generate sequence data continuously. A resistance prediction tool that requires only sequence and drug fingerprint can be integrated into surveillance pipelines without waiting for structural determination. My prior work with the GHRU-GSAR group on AMR genomics and surveillance pipelines informs this framing; I know the data formats and workflows that such a tool must fit into.

The next step is validation on larger clinical datasets, including longitudinal resistance data from Nigerian and West African surveillance programs, where sequence data is abundant and structural coverage is sparse. The manuscript submitted here establishes the method and its benchmark performance. The follow-up work will test whether the sequence-only approach maintains its advantage when trained and evaluated on geographically distinct resistance data.

SHORT-ANSWER ESSAY: RELEVANCE TO ANTIMICROBIAL RESISTANCE AND INFECTION CONTROL

This manuscript addresses the infection control problem of undetected resistance. When a resistance mutation appears in a clinical isolate, the current computational tools often cannot score it because no protein structure is available. The mutation goes uncharacterized, and the treatment decision proceeds without that information. TOPOLOGIX removes the structural requirement entirely. It predicts resistance from the protein sequence and the drug fingerprint alone, which means every sequenced mutation can be evaluated.

The practical consequence for infection control is coverage. Structure-based tools cover roughly 18 percent of Platinum benchmark mutations. TOPOLOGIX covers 100 percent. In a surveillance setting, this is the difference between characterizing a minority of resistance events and characterizing all of them. The accuracy, AUROC 0.804 on Platinum, is sufficient to rank mutations for experimental follow-up and to flag likely resistance for clinical awareness, even if it is not a standalone diagnostic.

The method is also deployable in low-resource settings. It requires no structural biology infrastructure, no GPU cluster for inference, and no proprietary software. The pipeline runs on standard hardware with open-source Python libraries. This matters for antimicrobial resistance surveillance in Nigeria and other West African countries, where sequencing capacity is growing but structural biology capacity is minimal. A sequence-only tool fits the infrastructure that exists.

The manuscript also contributes to infection control by reporting a falsified hypothesis. The interface-topology approach did not predict resistance, and that negative result is published as part of the methods. This prevents other groups from spending resources on the same dead end and clarifies that sequence representations, not interface geometry, carry the predictive signal. For a field that needs efficient allocation of limited research capacity, that is a concrete contribution.

SHORT-ANSWER ESSAY: NOVELTY AND SIGNIFICANCE

The novelty is the combination of three elements that have not been tested together on the Platinum benchmark: ESM-2 delta-embeddings as the protein representation, Morgan fingerprints as the drug representation, and a direct benchmark comparison against structure-based tools. The published literature contains structure-based resistance prediction, and it contains protein language model applications to mutation effect prediction, but the specific comparison, sequence-only versus structure-based on the same resistance benchmark with explicit coverage reporting, had not been run.

The significance is the coverage result. mCSM-lig, a widely used structure-based tool, cannot score 82 percent of Platinum mutations because it needs a resolved structure. TOPOLOGIX scores all of them. For a clinical surveillance pipeline, coverage is not a secondary metric; it determines whether the tool can be used at all on a given dataset. The accuracy improvement, 0.804 versus approximately 0.70, is meaningful, but the coverage improvement is the structural advantage.

The negative control adds significance. The bipartite persistent homology approach, which I tested with the same rigor as the positive result, failed to predict resistance. This is the first direct test of that hypothesis on the Platinum benchmark, and the negative result is reported transparently. It rules out interface geometry as the driver of resistance predictability and justifies the pivot to sequence representations. The field now knows that the signal is in the sequence, not the shape of the interface.

The work is also significant for its reproducibility. The benchmark is public, the code is on GitHub, the comparison is pre-registered, and the negative result is included. This is the standard your journal applies to antimicrobial resistance research, and it is the standard I applied to my own work before submission.

SHORT-ANSWER ESSAY: METHODOLOGICAL RIGOR AND OPEN SCIENCE

The manuscript follows a pre-registered analysis plan. The primary outcome, AUROC on the Platinum benchmark, was specified before the classifier was evaluated. The comparison against mCSM-lig was specified as the baseline. The external generalization check on SKEMPI 2.0 was specified as a secondary outcome. The pre-registration is available on OSF.

The benchmark is public and fixed. Platinum is a standard resource in the field, and I report results on the full dataset without excluding difficult cases. The coverage metric, 100 percent versus approximately 18 percent for structure-based tools, is reported explicitly because it is a primary finding, not a limitation.

The code is open source on GitHub, with the full pipeline from raw data to AUROC calculation. The environment is specified with pinned dependencies. The Random Forest hyperparameters are fixed and reported. The ESM-2 model version is specified. Any researcher can reproduce the results from the public data.

The negative control is reported with the same detail as the positive result. The bipartite persistent homology approach, with the opposition-distance metric and element-specific variants, achieved AUROCs of 0.425 and 0.485 on Platinum. These are below chance for one variant and near chance for the other. The manuscript reports these numbers and states plainly that interface topology does not predict resistance on this benchmark. This is the result that motivated the sequence-based approach, and it is part of the scientific record.

The manuscript adheres to the reporting standards your journal expects for computational studies: fixed benchmarks, pre-registered outcomes, open code, and transparent reporting of both positive and negative results. I have no conflicts of interest to declare. The data and code are available for peer review.

CHECKLIST

- [ ] Submit manuscript to Science AI Journal, Antimicrobial Resistance and Infection Control section, via the journal submission portal at https://scienceaijournal.com/journals/antimicrobial-resistance-and-infection-control-2047-2994
- [ ] Verify the journal's manuscript formatting requirements (title page, abstract structure, reference style) on the submission portal before uploading
- [ ] Confirm the manuscript file format (Word or LaTeX) accepted by the journal
- [ ] Include ORCID: 0009-0001-9272-6735 on the title page
- [ ] Include GitHub repository link (github.com/AmunRaPtah) in the data availability statement
- [ ] Include OSF pre-registration link for the TOPOLOGIX benchmark comparison
- [ ] Include the Platinum benchmark citation and SKEMPI 2.0 citation in the references
- [ ] Include the negative result (bipartite persistent homology, AUROC 0.425 and 0.485) in the methods or results section as the motivation for the sequence-based approach
- [ ] Include the coverage comparison (100 percent versus approximately 18 percent) as a primary finding
- [ ] Include a data availability statement confirming that all data are public and code is on GitHub
- [ ] Include a conflict of interest statement declaring no competing interests
- [ ] Confirm the journal's policy on preprints; the TOPOLOGIX work may be posted on arXiv or Zenodo, and the journal's preprint policy must be checked before submission
- [ ] Verify the journal's APC (article processing charge) and whether a waiver or discount applies for LMIC-affiliated authors; the applicant is Nigeria-based and enrolled at HPI/Potsdam, so eligibility for a waiver must be confirmed
- [ ] Confirm the journal's peer review timeline and whether it offers open peer review
- [ ] Prepare a cover letter to the editor summarizing the manuscript and its fit to the journal scope, if the submission portal requires one

EDITOR NOTES

- Eligibility risk: The journal is listed as "Science AI Journal" with a BMC-style URL (2047-2994 is the ISSN for Antimicrobial Resistance and Infection Control, a BMC journal). The applicant must verify whether this is the actual BMC journal or a mirror/predatory site. The URL scienceaijournal.com is not the standard BMC domain (bmcpublichealth.biomedcentral.com or similar). Before submitting, confirm the journal's legitimacy via the ISSN and publisher records. Do not pay an APC to an unverified site.
- Fact check: The profile states TOPOLOGIX AUROC 0.804 plus or minus 0.025 on Platinum and 0.634 on SKEMPI 2.0. The mCSM-lig baseline of approximately 0.70 is cited from the profile; the applicant must verify this number against the specific mCSM-lig publication and the Platinum benchmark paper before including it in the manuscript, since the comparison is central to the novelty claim.
- Gap: The profile does not specify whether the TOPOLOGIX manuscript is already written, posted as a preprint, or in preparation. The applicant must confirm the manuscript draft exists and is ready for submission, or adjust the timeline accordingly.
- Gap: The profile lists the applicant as enrolled in M.Sc. Digital Health at HPI/Potsdam starting Winter Semester 2026/27, and employed as National Product Manager at Synthcare from March 2026. The applicant must confirm current affiliation for the submission, since the journal will require a corresponding author affiliation and institutional email if available.
- Gap: The profile does not list co-authors for the TOPOLOGIX work. The applicant must decide whether to submit as sole author or add collaborators, and must confirm that any named collaborators (e.g., from GHRU-GSAR) have agreed to co-authorship and have seen the manuscript.
- Verification: The profile lists endorsements from Kent Berridge, Samuel Gershman, Nathaniel Daw, and Marcelo Mattar. These are not co-authors on TOPOLOGIX and should not be listed as such. If the applicant wishes to mention these connections in a cover letter, they must confirm permission first.
- Verification: The negative result (bipartite persistent homology, AUROC 0.425 and 0.485) is described as a pre-registered, powered replication. The applicant must confirm the pre-registration is publicly accessible and that the OSF link is included in the manuscript.
- The applicant must confirm whether the Platinum benchmark and SKEMPI 2.0 datasets are publicly downloadable without restriction, and whether any licensing terms affect the data availability statement.

AI Draft — Antimicrobial Resistance and… · Science AI Journal