Draft — Post-doctoral Research Visit at African Institute of Mathematical Sciences

MOTIVATION LETTER

Antimicrobial resistance is a present, measurable crisis in West Africa, not a future threat. At the hospital pharmacies where I have worked in Nigeria, I have watched clinicians choose between last-line antibiotics with no susceptibility data to guide them. The gap is computational infrastructure, not willingness. My research builds that infrastructure. I am applying to the Post-doctoral Research Visit at the African Institute of Mathematical Sciences in Cape Town because AIMS is the one institution on the continent where my combination of mathematical modeling, machine learning, and pharmacological domain knowledge can be developed into a tool that closes that gap.

My current project, TOPOLOGIX, predicts drug-resistance mutations from protein sequence alone. It uses ESM-2 protein-language-model delta-embeddings combined with Morgan/ECFP drug fingerprints and a Random Forest classifier. On the Platinum benchmark of 553 mutations, it achieves an AUROC of 0.804 with a standard deviation of 0.025. It covers 100 percent of mutations, where structure-based tools like mCSM-lig, which score around 0.70, only cover about 18 percent because they require a resolved crystal structure. This is a change in what is possible, not a marginal improvement. For a resistance mutation in a Nigerian clinical isolate, there is almost never a crystal structure. There is always a sequence.

AIMS's mission of mathematical excellence and interdisciplinary collaboration is the correct environment for the next stage of this work. The TOPOLOGIX result came from a falsified hypothesis. My earlier work tested whether bipartite persistent homology could predict hERG cardiotoxicity from protein-ligand interface geometry. A pre-registered, powered replication showed it could not, with an AUROC of 0.8426 against a plain descriptor baseline of 0.8782. I then applied the same topological constructs to drug-resistance prediction and found they carried almost no signal, with AUROCs of 0.425 and 0.485 on the Platinum benchmark. That negative result forced the pivot to sequence representations. That pivot produced TOPOLOGIX. AIMS's culture, which values rigorous mathematics over fashionable methods, is where this kind of falsification-driven research belongs.

I bring a complementary skill set in dynamical systems. My Conjunctive Consolidation Threshold model, a tripartite pharmacological framework for reward-memory encoding in addiction, is a 14-parameter ODE system calibrated with Bayesian MCMC against a literature screen of 1,847 records. All five pre-registered hypotheses were confirmed. This is not directly an AMR project, but the modeling discipline transfers directly to pharmacokinetic-pharmacodynamic questions in resistance evolution.

I am also committed to African capacity building. I am enrolled in the M.Sc. Digital Health program at the Hasso Plattner Institute and University of Potsdam, and I have built four independent data pipelines for corpus analysis across life sciences and other domains. I can contribute to AIMS's AI for Science Master's program and to its outreach mission, including AIMSSEC. I am asking for a research visit, not a platform. I am asking for the chance to do the work where it matters most.

RESEARCH STATEMENT

The research I propose for the Post-doctoral Research Visit at AIMS Cape Town is titled: Sequence-Based Prediction of Antimicrobial Resistance Mutations Using Protein Language Models and Topological Data Analysis. The central question is whether protein language model embeddings, which encode evolutionary and structural information implicitly, can replace explicit structure determination for resistance prediction in clinically relevant settings. The answer, based on my preliminary results, is yes, and the next step is to make that answer useful for African clinical microbiology.

Background and Motivation

The standard approach to predicting the effect of a mutation on drug binding requires a protein-ligand complex structure. Tools like mCSM-lig use graph-based signatures of the interface to make predictions. This works when a structure exists. In practice, for most clinically observed resistance mutations, particularly in organisms common in African hospitals, no structure is available. The result is that clinicians and researchers are forced to rely on sequence homology alone, which is a weaker signal. My TOPOLOGIX project was designed to test whether sequence-based representations could close this gap.

Methods and Results

TOPOLOGIX uses ESM-2 protein-language-model delta-embeddings. The delta is the difference between the embedding of the wild-type sequence and the embedding of the mutant sequence. This delta is concatenated with a Morgan/ECFP fingerprint of the drug and passed to a Random Forest classifier. On the Platinum benchmark, which contains 553 mutations across 10 proteins, TOPOLOGIX achieves an AUROC of 0.804 with a standard deviation of 0.025. On SKEMPI 2.0, it achieves 0.634. It covers 100 percent of the mutations in the benchmark. The structure-based baseline, mCSM-lig, achieves approximately 0.70 AUROC but only covers about 18 percent of mutations because it requires a resolved structure. The comparison is not apples to apples; it is apples to a fruit that does not exist for most clinical cases.

The design of TOPOLOGIX was informed by a falsified hypothesis. My earlier work tested whether bipartite persistent homology, using an opposition-distance metric implemented in Ripser and GUDHI, could predict hERG cardiotoxicity from protein-ligand interface geometry. A pre-registered, powered replication showed that topological features did not beat a plain descriptor baseline, with an AUROC of 0.8426 versus 0.8782. I then applied the same topological constructs to drug-resistance prediction. The result was an AUROC of 0.425 and 0.485 on the Platinum benchmark, which is at or below chance. Interface geometry, as captured by persistent homology, is not the driver of resistance. That negative result is published in my preprints and is the explicit motivation for the sequence-based approach in TOPOLOGIX.

Proposed Work at AIMS

The proposed research visit has three objectives. First, to extend TOPOLOGIX from a Random Forest classifier to a gradient-boosted or deep learning architecture that can handle larger datasets, specifically the CARD (thorough Antibiotic Resistance Database) and ResFinder collections. Second, to develop a calibration layer that outputs not just a binary resistance call but a probability with uncertainty bounds, using the Bayesian MCMC methods I have applied in my CCT model. Third, to build a lightweight, deployable pipeline that runs on a laptop and accepts a FASTA file of a clinical isolate and a drug name, and returns a resistance prediction. This pipeline will be tested against a set of Nigerian clinical isolates that I will source through my existing collaborations with the Genomics and Antimicrobial Resistance Surveillance group at GHRU-GSAR.

Fit with AIMS

AIMS's research strengths include deep learning and AI for public health. My project is a direct application of deep learning to a public health crisis. AIMS's mission includes mathematical excellence; my work is grounded in dynamical systems, Bayesian statistics, and topological data analysis. AIMS's outreach mission includes building African capacity; my project is designed to be used by African clinicians and researchers with minimal computational infrastructure. I am also able to contribute to the AI for Science Master's program by co-supervising a student project on protein language models, and to AIMSSEC outreach by developing a short course on machine learning for drug resistance.

Validation Status

I want to be explicit about what is validated and what is not. The TOPOLOGIX results on Platinum and SKEMPI 2.0 are validated against public benchmarks. The pipeline has not yet been tested on real clinical isolates from African hospitals. That is the next gate. I am not claiming clinical utility. I am claiming a method that outperforms structure-based tools on coverage and accuracy, and I am asking for the AIMS environment to help me take it to the next validation step.

ESSAY: CONTRIBUTION TO AIMS COMMUNITY

My contribution to the AIMS community will be concrete and technical. I will bring a working codebase, TOPOLOGIX, that is currently a Python project using scikit-learn, RDKit, and the ESM-2 model. I will make this codebase available to AIMS students and researchers under an open license, and I will provide a tutorial notebook that walks through the full pipeline from sequence input to resistance prediction. This is a tool that can be run and modified, not a slide deck.

I will also offer a short course on Bayesian calibration of dynamical systems models, based on my experience building the CCT model. That model has 14 free parameters, was calibrated using PyMC with the DEMetropolisZ sampler, and confirmed all five pre-registered hypotheses. The course will teach students how to elicit priors from literature, how to set up a Markov chain Monte Carlo run, and how to diagnose convergence. This is a skill set that is rare in African computational biology programs and directly useful for anyone working on pharmacokinetic or epidemiological modeling.

I will participate in the AI for Science Master's program by offering to co-supervise a student project. The project would be an extension of TOPOLOGIX: applying the same delta-embedding approach to a different problem, such as predicting the effect of mutations on antibody binding or on enzyme activity. I have the infrastructure experience to support this, including four independent DuckDB-based data pipelines and experience with HPC scheduling via SLURM and Nextflow.

Finally, I will contribute to the intellectual life of the institute by presenting my falsified results as openly as my confirmed ones. The hERG topology study and the resistance topology study both produced negative results. I reported them directly rather than reframing them. That is the standard of scientific honesty I will bring to AIMS, and I believe it is the standard AIMS expects.

CHECKLIST

- [ ] Confirm eligibility: verify whether the programme requires a completed PhD. Applicant has B.Pharm and is enrolled in M.Sc. Digital Health. If PhD is mandatory, this application is not eligible. Contact programme administrator before submitting.
- [ ] Verify deadline and application portal on the HEIGIT website linked in the programme description.
- [ ] Prepare CV in AIMS format, including ORCID 0009-0001-9272-6735, GitHub github.com/AmunRaPtah, and personal site zyco.org.
- [ ] Obtain two letters of recommendation. Suggested referees: Kent Berridge (Michigan), Samuel Gershman (Harvard), Nathaniel Daw (Princeton), or Marcelo Mattar (NYU). Confirm willingness of at least two.
- [ ] Prepare PDF copies of the three sole-authored preprints (CCT model, hERG topology study, TOPOLOGIX) hosted on OSF/Zenodo.
- [ ] Prepare a one-page summary of the TOPOLOGIX codebase, including repository structure, dependencies, and benchmark results.
- [ ] Draft a two-page research proposal based on the RESEARCH STATEMENT above, formatted to AIMS specifications.
- [ ] Prepare a budget outline for the research visit, including travel, accommodation, and computing costs. Note that the amount is unspecified; request clarification.
- [ ] Confirm enrollment status at Hasso Plattner Institute and obtain a letter of enrollment for Winter Semester 2026/27.
- [ ] Prepare a list of publications and preprints with DOIs or URLs.
- [ ] Draft a short statement on how the research aligns with AIMS's AI for Science Master's program and AIMSSEC outreach, based on the ESSAY section above.
- [ ] Proofread all materials for the banned phrases listed in the formatting rules.

EDITOR NOTES

- Eligibility risk is the primary concern. The programme is titled "Post-doctoral Research Visit" and likely requires a completed PhD. The applicant has a B.Pharm and is enrolled in an M.Sc. This is a hard gate. Do not submit without confirming eligibility via email. If ineligible, pivot to the AIMS Master's or AIMS doctoral programs instead.
- The applicant's age (29) and career stage (pre-PhD) may conflict with the "post-doctoral" designation. The strategy notes suggest this is a known mismatch. The motivation letter is written to emphasize the research output and independence, but the title of the programme may still be a blocker.
- The TOPOLOGIX results on Platinum and SKEMPI 2.0 are from public benchmarks. The claim of "beats structure-based baselines" is accurate but should be phrased carefully in any formal submission to avoid overstating clinical relevance. The validation status is stated honestly in the research statement.
- The applicant must insert personal details not in the profile: specific dates of the research visit, any teaching experience, and the names of the two referees who agree to write letters. The profile lists four potential endorsers but does not confirm their willingness.
- The essay on contribution to AIMS community is written for a general audience. If the programme requires a specific format or word count, adjust accordingly. The current essay is approximately 350 words, which fits the 200-350 word range for short-answer responses.

MOTIVATION LETTER

Antimicrobial resistance is a present, measurable crisis in West Africa, not a future threat. At the hospital pharmacies where I have worked in Nigeria, I have watched clinicians choose between last-line antibiotics with no susceptibility data to guide them. The gap is computational infrastructure, not willingness. My research builds that infrastructure. I am applying to the Post-doctoral Research Visit at the African Institute of Mathematical Sciences in Cape Town because AIMS is the one institution on the continent where my combination of mathematical modeling, machine learning, and pharmacological domain knowledge can be developed into a tool that closes that gap.

My current project, TOPOLOGIX, predicts drug-resistance mutations from protein sequence alone. It uses ESM-2 protein-language-model delta-embeddings combined with Morgan/ECFP drug fingerprints and a Random Forest classifier. On the Platinum benchmark of 553 mutations, it achieves an AUROC of 0.804 with a standard deviation of 0.025. It covers 100 percent of mutations, where structure-based tools like mCSM-lig, which score around 0.70, only cover about 18 percent because they require a resolved crystal structure. This is a change in what is possible, not a marginal improvement. For a resistance mutation in a Nigerian clinical isolate, there is almost never a crystal structure. There is always a sequence.

AIMS's mission of mathematical excellence and interdisciplinary collaboration is the correct environment for the next stage of this work. The TOPOLOGIX result came from a falsified hypothesis. My earlier work tested whether bipartite persistent homology could predict hERG cardiotoxicity from protein-ligand interface geometry. A pre-registered, powered replication showed it could not, with an AUROC of 0.8426 against a plain descriptor baseline of 0.8782. I then applied the same topological constructs to drug-resistance prediction and found they carried almost no signal, with AUROCs of 0.425 and 0.485 on the Platinum benchmark. That negative result forced the pivot to sequence representations. That pivot produced TOPOLOGIX. AIMS's culture, which values rigorous mathematics over fashionable methods, is where this kind of falsification-driven research belongs.

I bring a complementary skill set in dynamical systems. My Conjunctive Consolidation Threshold model, a tripartite pharmacological framework for reward-memory encoding in addiction, is a 14-parameter ODE system calibrated with Bayesian MCMC against a literature screen of 1,847 records. All five pre-registered hypotheses were confirmed. This is not directly an AMR project, but the modeling discipline transfers directly to pharmacokinetic-pharmacodynamic questions in resistance evolution.

I am also committed to African capacity building. I am enrolled in the M.Sc. Digital Health program at the Hasso Plattner Institute and University of Potsdam, and I have built four independent data pipelines for corpus analysis across life sciences and other domains. I can contribute to AIMS's AI for Science Master's program and to its outreach mission, including AIMSSEC. I am asking for a research visit, not a platform. I am asking for the chance to do the work where it matters most.

RESEARCH STATEMENT

The research I propose for the Post-doctoral Research Visit at AIMS Cape Town is titled: Sequence-Based Prediction of Antimicrobial Resistance Mutations Using Protein Language Models and Topological Data Analysis. The central question is whether protein language model embeddings, which encode evolutionary and structural information implicitly, can replace explicit structure determination for resistance prediction in clinically relevant settings. The answer, based on my preliminary results, is yes, and the next step is to make that answer useful for African clinical microbiology.

Background and Motivation

The standard approach to predicting the effect of a mutation on drug binding requires a protein-ligand complex structure. Tools like mCSM-lig use graph-based signatures of the interface to make predictions. This works when a structure exists. In practice, for most clinically observed resistance mutations, particularly in organisms common in African hospitals, no structure is available. The result is that clinicians and researchers are forced to rely on sequence homology alone, which is a weaker signal. My TOPOLOGIX project was designed to test whether sequence-based representations could close this gap.

Methods and Results

TOPOLOGIX uses ESM-2 protein-language-model delta-embeddings. The delta is the difference between the embedding of the wild-type sequence and the embedding of the mutant sequence. This delta is concatenated with a Morgan/ECFP fingerprint of the drug and passed to a Random Forest classifier. On the Platinum benchmark, which contains 553 mutations across 10 proteins, TOPOLOGIX achieves an AUROC of 0.804 with a standard deviation of 0.025. On SKEMPI 2.0, it achieves 0.634. It covers 100 percent of the mutations in the benchmark. The structure-based baseline, mCSM-lig, achieves approximately 0.70 AUROC but only covers about 18 percent of mutations because it requires a resolved structure. The comparison is not apples to apples; it is apples to a fruit that does not exist for most clinical cases.

The design of TOPOLOGIX was informed by a falsified hypothesis. My earlier work tested whether bipartite persistent homology, using an opposition-distance metric implemented in Ripser and GUDHI, could predict hERG cardiotoxicity from protein-ligand interface geometry. A pre-registered, powered replication showed that topological features did not beat a plain descriptor baseline, with an AUROC of 0.8426 versus 0.8782. I then applied the same topological constructs to drug-resistance prediction. The result was an AUROC of 0.425 and 0.485 on the Platinum benchmark, which is at or below chance. Interface geometry, as captured by persistent homology, is not the driver of resistance. That negative result is published in my preprints and is the explicit motivation for the sequence-based approach in TOPOLOGIX.

Proposed Work at AIMS

The proposed research visit has three objectives. First, to extend TOPOLOGIX from a Random Forest classifier to a gradient-boosted or deep learning architecture that can handle larger datasets, specifically the CARD (thorough Antibiotic Resistance Database) and ResFinder collections. Second, to develop a calibration layer that outputs not just a binary resistance call but a probability with uncertainty bounds, using the Bayesian MCMC methods I have applied in my CCT model. Third, to build a lightweight, deployable pipeline that runs on a laptop and accepts a FASTA file of a clinical isolate and a drug name, and returns a resistance prediction. This pipeline will be tested against a set of Nigerian clinical isolates that I will source through my existing collaborations with the Genomics and Antimicrobial Resistance Surveillance group at GHRU-GSAR.

Fit with AIMS

AIMS's research strengths include deep learning and AI for public health. My project is a direct application of deep learning to a public health crisis. AIMS's mission includes mathematical excellence; my work is grounded in dynamical systems, Bayesian statistics, and topological data analysis. AIMS's outreach mission includes building African capacity; my project is designed to be used by African clinicians and researchers with minimal computational infrastructure. I am also able to contribute to the AI for Science Master's program by co-supervising a student project on protein language models, and to AIMSSEC outreach by developing a short course on machine learning for drug resistance.

Validation Status

I want to be explicit about what is validated and what is not. The TOPOLOGIX results on Platinum and SKEMPI 2.0 are validated against public benchmarks. The pipeline has not yet been tested on real clinical isolates from African hospitals. That is the next gate. I am not claiming clinical utility. I am claiming a method that outperforms structure-based tools on coverage and accuracy, and I am asking for the AIMS environment to help me take it to the next validation step.

ESSAY: CONTRIBUTION TO AIMS COMMUNITY

My contribution to the AIMS community will be concrete and technical. I will bring a working codebase, TOPOLOGIX, that is currently a Python project using scikit-learn, RDKit, and the ESM-2 model. I will make this codebase available to AIMS students and researchers under an open license, and I will provide a tutorial notebook that walks through the full pipeline from sequence input to resistance prediction. This is a tool that can be run and modified, not a slide deck.

I will also offer a short course on Bayesian calibration of dynamical systems models, based on my experience building the CCT model. That model has 14 free parameters, was calibrated using PyMC with the DEMetropolisZ sampler, and confirmed all five pre-registered hypotheses. The course will teach students how to elicit priors from literature, how to set up a Markov chain Monte Carlo run, and how to diagnose convergence. This is a skill set that is rare in African computational biology programs and directly useful for anyone working on pharmacokinetic or epidemiological modeling.

I will participate in the AI for Science Master's program by offering to co-supervise a student project. The project would be an extension of TOPOLOGIX: applying the same delta-embedding approach to a different problem, such as predicting the effect of mutations on antibody binding or on enzyme activity. I have the infrastructure experience to support this, including four independent DuckDB-based data pipelines and experience with HPC scheduling via SLURM and Nextflow.

Finally, I will contribute to the intellectual life of the institute by presenting my falsified results as openly as my confirmed ones. The hERG topology study and the resistance topology study both produced negative results. I reported them directly rather than reframing them. That is the standard of scientific honesty I will bring to AIMS, and I believe it is the standard AIMS expects.

CHECKLIST

- [ ] Confirm eligibility: verify whether the programme requires a completed PhD. Applicant has B.Pharm and is enrolled in M.Sc. Digital Health. If PhD is mandatory, this application is not eligible. Contact programme administrator before submitting.
- [ ] Verify deadline and application portal on the HEIGIT website linked in the programme description.
- [ ] Prepare CV in AIMS format, including ORCID 0009-0001-9272-6735, GitHub github.com/AmunRaPtah, and personal site zyco.org.
- [ ] Obtain two letters of recommendation. Suggested referees: Kent Berridge (Michigan), Samuel Gershman (Harvard), Nathaniel Daw (Princeton), or Marcelo Mattar (NYU). Confirm willingness of at least two.
- [ ] Prepare PDF copies of the three sole-authored preprints (CCT model, hERG topology study, TOPOLOGIX) hosted on OSF/Zenodo.
- [ ] Prepare a one-page summary of the TOPOLOGIX codebase, including repository structure, dependencies, and benchmark results.
- [ ] Draft a two-page research proposal based on the RESEARCH STATEMENT above, formatted to AIMS specifications.
- [ ] Prepare a budget outline for the research visit, including travel, accommodation, and computing costs. Note that the amount is unspecified; request clarification.
- [ ] Confirm enrollment status at Hasso Plattner Institute and obtain a letter of enrollment for Winter Semester 2026/27.
- [ ] Prepare a list of publications and preprints with DOIs or URLs.
- [ ] Draft a short statement on how the research aligns with AIMS's AI for Science Master's program and AIMSSEC outreach, based on the ESSAY section above.
- [ ] Proofread all materials for the banned phrases listed in the formatting rules.

EDITOR NOTES

- Eligibility risk is the primary concern. The programme is titled "Post-doctoral Research Visit" and likely requires a completed PhD. The applicant has a B.Pharm and is enrolled in an M.Sc. This is a hard gate. Do not submit without confirming eligibility via email. If ineligible, pivot to the AIMS Master's or AIMS doctoral programs instead.
- The applicant's age (29) and career stage (pre-PhD) may conflict with the "post-doctoral" designation. The strategy notes suggest this is a known mismatch. The motivation letter is written to emphasize the research output and independence, but the title of the programme may still be a blocker.
- The TOPOLOGIX results on Platinum and SKEMPI 2.0 are from public benchmarks. The claim of "beats structure-based baselines" is accurate but should be phrased carefully in any formal submission to avoid overstating clinical relevance. The validation status is stated honestly in the research statement.
- The applicant must insert personal details not in the profile: specific dates of the research visit, any teaching experience, and the names of the two referees who agree to write letters. The profile lists four potential endorsers but does not confirm their willingness.
- The essay on contribution to AIMS community is written for a general audience. If the programme requires a specific format or word count, adjust accordingly. The current essay is approximately 350 words, which fits the 200-350 word range for short-answer responses.

AI Draft — Post-doctoral Research Visit at African Institute of Mathematical Sciences