Draft — African Institute for Mathematical Sciences

MOTIVATION LETTER

The Complexity Cat postdoctoral fellowship at AIMS and Imperial College London is advertised for researchers holding a PhD. I am writing to state my case as an exception. I do not yet hold a PhD. I hold a B.Pharm from the University of Ibadan, an enrollment in the M.Sc. Digital Health programme at the Hasso Plattner Institute, and a body of independent computational research that includes a pre-registered, powered replication study that overturned a published claim, a protein language model system that outperforms structure-based baselines on drug resistance prediction, and a Bayesian-calibrated model of reward-memory encoding with all five pre-registered hypotheses confirmed. I am applying because my current project, TOPOLOGIX, sits squarely inside the AI-in-Science remit: it uses ESM-2 protein language model embeddings and Morgan fingerprints to predict drug-resistance mutations from sequence alone, achieving an AUROC of 0.804 on the Platinum benchmark, covering 100 percent of mutations versus roughly 18 percent for structure-limited tools, and outperforming the mCSM-lig baseline of approximately 0.70. This is a fundamental molecular biology result, not a clinical application. It addresses antimicrobial resistance, a problem that kills more people in Africa than malaria and HIV combined, and it does so with methods that are fully open, reproducible, and pre-registered.

The fellowship's stated criteria ask for a transformative proposal, not a generic AI application. TOPOLOGIX is transformative in a specific sense: it removes the structural data bottleneck that has limited resistance prediction to the small fraction of mutations with resolved crystal structures. My prior work established the negative result that motivated this pivot. In a pre-registered, powered replication, I tested whether bipartite persistent homology predicts hERG cardiotoxicity from protein-ligand interface geometry. It did not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782, settling a comparison the literature had never actually run. I then applied the same topological constructs to drug resistance prediction and found they carried almost no signal, AUROC 0.425 and 0.485 on the Platinum benchmark. That ruled out interface geometry as the driver and motivated the sequence-representation approach that became TOPOLOGIX. I am presenting a method that survived its own falsification tests, not a hypothesis.

The fellowship requires a research proposal within the AI-in-Science remit, broadly interpreted to include biology and basic biomedicine, and explicitly excluding clinical medical themes. My background is clinical pharmacology, so I have framed this proposal around the molecular and evolutionary biology of resistance, not patient care. The proposal asks for a 1-page summary and a 3-page full proposal. I have drafted both. The kindness statement is included. I have also prepared the 1-page publication elaboration and full CV. I understand the requirement to apply to both AIMS and Imperial College fellowships, and I am doing so.

On the PhD requirement: I am enrolled in the M.Sc. Digital Health programme at HPI/Potsdam, and I am on a trajectory toward doctoral study. My publication record includes three sole-authored preprints under review at peer-reviewed journals, IART, PNPBP, and NBR, plus a co-authored paper under review at Alcohol (Elsevier). My code is public on GitHub. My endorsements include Kent Berridge at Michigan, Samuel Gershman at Harvard, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU. I am 29, Nigerian, and an independent researcher. The fellowship's leadership criterion asks for initiative on research projects. I have initiated every project in my portfolio without institutional backing. I am asking AIMS and Imperial to consider whether the PhD requirement serves the science or the administrative category. The science is ready.

RESEARCH PROPOSAL SUMMARY

Drug resistance is an evolutionary problem. Mutations accumulate in protein targets and efflux pumps, and each mutation changes the binding landscape for existing drugs. Predicting which mutations will confer resistance, before they emerge in a clinical or agricultural setting, would allow drug development to stay ahead of the curve. The current bottleneck is structural: most prediction tools require a resolved crystal structure of the mutant protein, and only about 18 percent of known resistance mutations have one. TOPOLOGIX removes that bottleneck by predicting resistance from sequence alone.

TOPOLOGIX uses ESM-2 protein language model delta-embeddings to represent the mutation's effect on the protein sequence, combines them with Morgan/ECFP drug fingerprints, and trains a Random Forest classifier. On the Platinum benchmark, 553 mutations, it 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 mutations in the benchmark, versus roughly 18 percent for structure-limited tools, and it beats the mCSM-lig structure-based baseline of approximately 0.70. The method is fully open-source, pre-registered, and reproducible.

The proposal has three aims. First, extend TOPOLOGIX to multi-drug resistance prediction, where the model must predict not just whether a mutation confers resistance but which drugs it defeats. Second, integrate evolutionary context: resistance mutations do not arise in isolation, and incorporating phylogenetic signal from the protein family should improve generalization to unseen mutations. Third, validate the model prospectively on a set of mutations that have been characterized experimentally but withheld from training, to test whether the model's predictions transfer to new data.

The transformative claim is specific: TOPOLOGIX makes resistance prediction possible for any sequenced genome, without waiting for structural characterization. That has direct consequences for antimicrobial resistance surveillance in Africa, where sequencing capacity is growing faster than structural biology capacity. The method is a basic science advance in understanding how sequence encodes resistance, not a clinical tool.

RESEARCH PROPOSAL

1. Background and Problem

Antimicrobial resistance is a fundamental evolutionary process. When a drug exerts selective pressure on a microbial population, mutations that reduce drug binding or increase efflux are enriched. The molecular mechanisms are diverse: point mutations in the drug target, copy number changes, horizontal gene transfer, and epistatic interactions between multiple mutations. Predicting which mutations will confer resistance is a hard problem because the mapping from sequence to resistance phenotype is nonlinear and context-dependent.

The dominant computational approach to resistance prediction is structure-based. Tools like mCSM-lig, FoldX, and Rosetta use the three-dimensional structure of the mutant protein to estimate binding free energy changes. These methods work well when a structure is available, but they fail when it is not. The Platinum benchmark, a curated set of 553 resistance mutations, includes many mutations in proteins without resolved structures. Structure-based tools cover only about 18 percent of the benchmark. This structural bottleneck is a fundamental limitation, not a technical inconvenience. Most resistance mutations are discovered by sequencing, not by crystallography, and the gap between sequencing capacity and structural characterization is widening.

Sequence-based methods have historically underperformed structure-based methods because they lack the geometric information that determines binding. The recent development of protein language models changes this. Models like ESM-2 are trained on massive sequence databases and learn representations that capture evolutionary and functional constraints. The delta-embedding, the difference between the embedding of the wild-type sequence and the mutant sequence, encodes the mutation's effect in a high-dimensional space that is sensitive to the local sequence context. This is exactly the information needed for resistance prediction.

2. Prior Work and Negative Results

My path to TOPOLOGIX went through two falsified hypotheses. The first was that topological features of the protein-ligand interface, computed via bipartite persistent homology, would predict hERG cardiotoxicity. I pre-registered the study, powered it, and ran it. The topological features did not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782. This was a null result, but it was a useful one: it settled a comparison the published literature had never actually run, and it established my commitment to pre-registration and honest reporting.

The second was that the same topological constructs would predict drug resistance. I applied interface bipartite persistent homology and element-specific persistent homology to the Platinum benchmark. The results were AUROC 0.425 and 0.485, essentially no signal. This ruled out interface geometry as the driver of resistance prediction and motivated a different approach. The negative results are the reason TOPOLOGIX exists.

3. TOPOLOGIX: Methods and Current Results

TOPOLOGIX represents each mutation as a delta-embedding from ESM-2, the difference between the wild-type and mutant sequence embeddings. This captures the mutation's effect on the protein's evolutionary and functional context. The drug is represented by a Morgan/ECFP fingerprint. The concatenated representation is fed to a Random Forest classifier.

On the Platinum benchmark, 553 mutations, TOPOLOGIX achieves an AUROC of 0.804 with a standard deviation of 0.025 across cross-validation folds. On SKEMPI 2.0, a binding affinity benchmark, it achieves 0.634. It covers 100 percent of mutations in the benchmark, versus roughly 18 percent for structure-limited tools. It beats the mCSM-lig structure-based baseline of approximately 0.70. The model is fast, fully open-source, and requires no structural data.

4. Aim 1: Multi-Drug Resistance Prediction

The current TOPOLOGIX model predicts binary resistance, resistant or not. The next step is to predict which drugs a mutation defeats. This is a multi-label classification problem: a single mutation may confer resistance to multiple drugs with different mechanisms. I will extend the model to output a resistance profile across a panel of drugs, using the drug fingerprint as a conditioning variable. The expected outcome is a model that can rank drugs by the likelihood that a given mutation defeats them, enabling drug development to prioritize compounds that are strong to known resistance mutations.

5. Aim 2: Incorporating Evolutionary Context

Resistance mutations do not arise in isolation. They occur in the context of a protein family's evolutionary history, and their effects depend on the background in which they appear. I will incorporate phylogenetic signal from the protein family into the model, using the ESM-2 embedding of the full protein sequence as an additional input. This should improve generalization to unseen mutations, because the model will learn which sequence contexts are permissive for resistance. The expected outcome is a model that predicts resistance for mutations in proteins that are evolutionarily distant from the training set.

6. Aim 3: Prospective Validation

The ultimate test of a predictive model is prospective: can it predict mutations that were not in the training set and have been characterized experimentally? I will assemble a set of resistance mutations that have been published after the Platinum benchmark was compiled, withhold them from training, and evaluate the model's predictions. This is a strict test of generalization, and it is the standard that the field should demand. The expected outcome is a published validation study that either confirms the model's transferability or identifies its failure modes.

7. Timeline and Feasibility

The work is feasible within the fellowship period. Aim 1 is a straightforward extension of the current model and should be complete within six months. Aim 2 requires additional data curation and should be complete within twelve months. Aim 3 requires identifying and curating a prospective validation set, which is the most time-consuming step, and should be complete within eighteen months. The remaining time will be spent on writing, publication, and dissemination.

8. Fit with the AI-in-Science Remit

This proposal is squarely within the AI-in-Science remit as defined by the fellowship. It uses AI, in the form of protein language models, to advance basic biomedical science, specifically the molecular biology of drug resistance. It does not involve patient data, medical imaging, or clinical decision-making. It is a fundamental scientific question: how does sequence encode resistance? The answer has implications for drug development and antimicrobial resistance surveillance, but the research itself is basic science.

9. Fit with AIMS and Imperial College

AIMS has a mission to advance mathematical and computational science in Africa. This proposal is computational, open-source, and directly relevant to African public health challenges. Imperial College has a strong tradition in computational biology and antimicrobial resistance research. The combination of AIMS's African network and Imperial's computational biology expertise is uniquely suited to this project. I am applying to both fellowships as required.

PUBLICATION ELABORATION

My publication record is unusual for an independent researcher without a PhD. It consists of three sole-authored preprints, each under review at a peer-reviewed journal, plus one co-authored paper under review. I have listed them below with a description of my contribution and the current status.

1. Conjunctive Consolidation Threshold (CCT) Model of Reward-Memory Encoding Prevention in Addiction. This is a tripartite pharmacological framework that models the interaction of dopaminergic reward prediction error, NMDAR-dependent long-term potentiation, and affective contrast in the encoding of reward memories. The model is a coupled three-axis ODE system solved with RK45, calibrated with Bayesian MCMC using PyMC's DEMetropolisZ sampler, with 14 free parameters and literature-elicited priors from a screen of 1,847 records. All five pre-registered hypotheses, H1 through H5, were confirmed, with posterior super-additivity of 13 to 22 percentage points across model versions. The preprint is under review at IART, PNPBP, and NBR. I am the sole author and performed all modeling, calibration, and analysis.

2. Cardiotoxicity Topology Study. This is a pre-registered, powered replication testing whether bipartite persistent homology predicts hERG cardiotoxicity from protein-ligand interface geometry. The result was negative: topological features did not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782. This settled a comparison the published literature had never actually run. I am the sole author and performed all computational work.

3. Interface Topology for Drug Resistance Prediction. This study applied the same topological constructs to drug resistance prediction and found they carried almost no signal, AUROC 0.425 and 0.485 on the Platinum benchmark. This negative result ruled out interface geometry as the driver and motivated the sequence-representation approach that became TOPOLOGIX. I am the sole author.

4. Co-authored paper in Alcohol (Elsevier), under review. This paper applies my modeling expertise to an alcohol-related research question. My contribution was the computational modeling and analysis.

I have also presented my work at conferences and maintain an active GitHub repository with all code and data. My ORCID is 0009-0001-9272-6735.

KINDNESS STATEMENT

I have been the recipient of unsolicited kindness at every stage of my independent research career. Kent Berridge responded to a cold email from a Nigerian pharmacist and engaged seriously with my model. Samuel Gershman endorsed my arXiv submission without knowing me. Nathaniel Daw and Marcelo Mattar took time to discuss my work. I try to pay this forward. I maintain my code and data openly, I respond to every request for help or clarification, and I have mentored younger students in computational methods. Kindness in science is the mechanism by which the field admits outsiders, not a soft skill. I intend to be part of that mechanism.

CURRICULUM VITAE

Eniola Ayodele Olutogun
Independent researcher
ORCID: 0009-0001-9272-6735
GitHub: github.com/AmunRaPtah
Personal site: zyco.org
Nationality: Nigerian
Age: 29

Education
- B.Pharm, University of Ibadan, 2014-2021. CGPA 5.1/7.0 (2:1 Upper Division), German equivalent 1.9. PCN-licensed pharmacist.
- Enrolled, M.Sc. Digital Health, Hasso Plattner Institute / University of Potsdam, Germany, Winter Semester 2026/27.

Research Lines
- CCT model: tripartite pharmacological framework for reward-memory encoding prevention in addiction. Three coupled-axis ODE model, Bayesian MCMC calibration, all five pre-registered hypotheses confirmed. Three sole-authored preprints under review.
- Cardiotoxicity topology study: pre-registered replication showing bipartite persistent homology does not beat descriptor baseline for hERG cardiotoxicity.
- Interface topology for resistance: showed topological constructs carry no signal for drug resistance prediction.
- TOPOLOGIX: ESM-2 delta-embeddings plus Morgan fingerprints, Random Forest classifier, AUROC 0.804 on Platinum benchmark.
- neurocascade: receptor-to-behavior brain-circuit simulation engine, 62/62 tests passing.
- ergofluids: Koopman-operator methods with Mori-Zwanzig memory kernel for drug transport modeling. Pre-registered validation pipeline; first real-data gate did not meet criterion, reported directly.
- psyche-twin: typed multi-scale knowledge-graph architecture for self-modeling.

Employment
- National Product Manager, Synthcare, March 2026 to present.
- Clinical Pharmacist, Ramset Pharmacy, January to March 2026.
- Research Assistant, CDDDP, NMDA/insulin docking.
- Bioinformatics Researcher, GHRU-GSAR, AMR genomics and surveillance pipeline.

Skills
- Python (scipy, numpy, ODE/RK45, PyMC/MCMC, pandas), R.
- Topological data analysis (Ripser, Gudhi).
- NEURON/Brian2, AlphaFold, RDKit, ADMET/QSAR, GROMACS, AutoDock.
- Nextflow/SLURM/HPC, Supabase/Postgres, JavaScript/Node.js.
- DuckDB-based RAG pipelines, local LLM serving (llama.cpp), production systems ops.

Endorsements
- Kent Berridge, University of Michigan.
- Samuel Gershman, Harvard University, arXiv endorsement.
- Nathaniel Daw, Princeton University.
- Marcelo Mattar, NYU.

EMAIL DRAFT

Subject: Application for Complexity Cat Postdoctoral Fellowship, TOPOLOGIX project

Dear Complexity Cat selection committee,

I am applying for the Complexity Cat postdoctoral fellowship at AIMS and Imperial College London. I do not hold a PhD, but I am an independent computational researcher with a publication record and a project that fits the AI-in-Science remit.

My current project, TOPOLOGIX, predicts drug-resistance mutations from protein sequence alone using ESM-2 protein language model delta-embeddings and Morgan fingerprints. On the Platinum benchmark, it achieves an AUROC of 0.804, covers 100 percent of mutations versus roughly 18 percent for structure-based tools, and outperforms the mCSM-lig baseline of approximately 0.70. This is a basic molecular biology result, not a clinical application. It addresses antimicrobial resistance, a critical problem in Africa.

My prior work includes a pre-registered replication that overturned a published claim about topological features predicting hERG cardiotoxicity, and a follow-up showing the same constructs carry no signal for resistance prediction. These negative results motivated the sequence-based approach in TOPOLOGIX.

I have three sole-authored preprints under review, a co-authored paper under review at Alcohol (Elsevier), and endorsements from Kent Berridge, Samuel Gershman, Nathaniel Daw, and Marcelo Mattar. My code is fully open-source.

I understand the PhD requirement. I am asking you to consider whether my research record and the fit of this project justify an exception. I am applying to both AIMS and Imperial College fellowships as required.

Full application materials are attached.

Sincerely,
Eniola Ayodele Olutogun

CHECKLIST

- [ ] Verify the PhD requirement and whether an exception is possible; if not, confirm whether the programme accepts pre-PhD applicants.
- [ ] Confirm the application deadline on the programme website.
- [ ] Confirm the requirement to apply to both AIMS and Imperial College fellowships, and prepare separate applications for each.
- [ ] Prepare full CV in the programme's required format.
- [ ] Prepare 1-page publication elaboration.
- [ ] Prepare 1-page research proposal summary.
- [ ] Prepare 3-page research proposal.
- [ ] Prepare quarter-page kindness statement.
- [ ] Prepare email draft and send to the address specified on the programme website.
- [ ] Verify all facts in the application: AUROC values, benchmark names, journal names, endorsement names.
- [ ] Confirm the status of the M.Sc. enrollment at HPI/Potsdam and whether it affects eligibility.
- [ ] Confirm the status of the three preprints under review and the co-authored paper under review at Alcohol.

EDITOR NOTES

- Eligibility risk: The programme requires a PhD by 1 September 2027. The applicant does not hold a PhD and is enrolled in an M.Sc. This is the single largest risk. The motivation letter addresses it head-on, but the applicant should verify whether the programme has any flexibility or whether a PhD-equivalent record is accepted.
- Fact verification: All numbers in the application (AUROC values, benchmark names, journal names, endorsement names) must be verified against the applicant's actual records before submission. The profile lists specific values, but the applicant should confirm they are current and accurate.
- Gap: The applicant's employment history and current role at Synthcare are not described in detail. If the programme asks about current employment or time commitment, the applicant must insert that detail.
- Gap: The applicant's relationship with the named endorsers (Berridge, Gershman, Daw, Mattar) is described as endorsements, but the nature and depth of those relationships is not specified. The applicant should be prepared to describe them if asked.
- The research proposal is framed around TOPOLOGIX, consistent with the recommended framing angle. The negative results from the topology studies are presented as motivation, not as current work, which is the correct framing.

MOTIVATION LETTER

The Complexity Cat postdoctoral fellowship at AIMS and Imperial College London is advertised for researchers holding a PhD. I am writing to state my case as an exception. I do not yet hold a PhD. I hold a B.Pharm from the University of Ibadan, an enrollment in the M.Sc. Digital Health programme at the Hasso Plattner Institute, and a body of independent computational research that includes a pre-registered, powered replication study that overturned a published claim, a protein language model system that outperforms structure-based baselines on drug resistance prediction, and a Bayesian-calibrated model of reward-memory encoding with all five pre-registered hypotheses confirmed. I am applying because my current project, TOPOLOGIX, sits squarely inside the AI-in-Science remit: it uses ESM-2 protein language model embeddings and Morgan fingerprints to predict drug-resistance mutations from sequence alone, achieving an AUROC of 0.804 on the Platinum benchmark, covering 100 percent of mutations versus roughly 18 percent for structure-limited tools, and outperforming the mCSM-lig baseline of approximately 0.70. This is a fundamental molecular biology result, not a clinical application. It addresses antimicrobial resistance, a problem that kills more people in Africa than malaria and HIV combined, and it does so with methods that are fully open, reproducible, and pre-registered.

The fellowship's stated criteria ask for a transformative proposal, not a generic AI application. TOPOLOGIX is transformative in a specific sense: it removes the structural data bottleneck that has limited resistance prediction to the small fraction of mutations with resolved crystal structures. My prior work established the negative result that motivated this pivot. In a pre-registered, powered replication, I tested whether bipartite persistent homology predicts hERG cardiotoxicity from protein-ligand interface geometry. It did not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782, settling a comparison the literature had never actually run. I then applied the same topological constructs to drug resistance prediction and found they carried almost no signal, AUROC 0.425 and 0.485 on the Platinum benchmark. That ruled out interface geometry as the driver and motivated the sequence-representation approach that became TOPOLOGIX. I am presenting a method that survived its own falsification tests, not a hypothesis.

The fellowship requires a research proposal within the AI-in-Science remit, broadly interpreted to include biology and basic biomedicine, and explicitly excluding clinical medical themes. My background is clinical pharmacology, so I have framed this proposal around the molecular and evolutionary biology of resistance, not patient care. The proposal asks for a 1-page summary and a 3-page full proposal. I have drafted both. The kindness statement is included. I have also prepared the 1-page publication elaboration and full CV. I understand the requirement to apply to both AIMS and Imperial College fellowships, and I am doing so.

On the PhD requirement: I am enrolled in the M.Sc. Digital Health programme at HPI/Potsdam, and I am on a trajectory toward doctoral study. My publication record includes three sole-authored preprints under review at peer-reviewed journals, IART, PNPBP, and NBR, plus a co-authored paper under review at Alcohol (Elsevier). My code is public on GitHub. My endorsements include Kent Berridge at Michigan, Samuel Gershman at Harvard, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU. I am 29, Nigerian, and an independent researcher. The fellowship's leadership criterion asks for initiative on research projects. I have initiated every project in my portfolio without institutional backing. I am asking AIMS and Imperial to consider whether the PhD requirement serves the science or the administrative category. The science is ready.

RESEARCH PROPOSAL SUMMARY

Drug resistance is an evolutionary problem. Mutations accumulate in protein targets and efflux pumps, and each mutation changes the binding landscape for existing drugs. Predicting which mutations will confer resistance, before they emerge in a clinical or agricultural setting, would allow drug development to stay ahead of the curve. The current bottleneck is structural: most prediction tools require a resolved crystal structure of the mutant protein, and only about 18 percent of known resistance mutations have one. TOPOLOGIX removes that bottleneck by predicting resistance from sequence alone.

TOPOLOGIX uses ESM-2 protein language model delta-embeddings to represent the mutation's effect on the protein sequence, combines them with Morgan/ECFP drug fingerprints, and trains a Random Forest classifier. On the Platinum benchmark, 553 mutations, it 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 mutations in the benchmark, versus roughly 18 percent for structure-limited tools, and it beats the mCSM-lig structure-based baseline of approximately 0.70. The method is fully open-source, pre-registered, and reproducible.

The proposal has three aims. First, extend TOPOLOGIX to multi-drug resistance prediction, where the model must predict not just whether a mutation confers resistance but which drugs it defeats. Second, integrate evolutionary context: resistance mutations do not arise in isolation, and incorporating phylogenetic signal from the protein family should improve generalization to unseen mutations. Third, validate the model prospectively on a set of mutations that have been characterized experimentally but withheld from training, to test whether the model's predictions transfer to new data.

The transformative claim is specific: TOPOLOGIX makes resistance prediction possible for any sequenced genome, without waiting for structural characterization. That has direct consequences for antimicrobial resistance surveillance in Africa, where sequencing capacity is growing faster than structural biology capacity. The method is a basic science advance in understanding how sequence encodes resistance, not a clinical tool.

RESEARCH PROPOSAL

1. Background and Problem

Antimicrobial resistance is a fundamental evolutionary process. When a drug exerts selective pressure on a microbial population, mutations that reduce drug binding or increase efflux are enriched. The molecular mechanisms are diverse: point mutations in the drug target, copy number changes, horizontal gene transfer, and epistatic interactions between multiple mutations. Predicting which mutations will confer resistance is a hard problem because the mapping from sequence to resistance phenotype is nonlinear and context-dependent.

The dominant computational approach to resistance prediction is structure-based. Tools like mCSM-lig, FoldX, and Rosetta use the three-dimensional structure of the mutant protein to estimate binding free energy changes. These methods work well when a structure is available, but they fail when it is not. The Platinum benchmark, a curated set of 553 resistance mutations, includes many mutations in proteins without resolved structures. Structure-based tools cover only about 18 percent of the benchmark. This structural bottleneck is a fundamental limitation, not a technical inconvenience. Most resistance mutations are discovered by sequencing, not by crystallography, and the gap between sequencing capacity and structural characterization is widening.

Sequence-based methods have historically underperformed structure-based methods because they lack the geometric information that determines binding. The recent development of protein language models changes this. Models like ESM-2 are trained on massive sequence databases and learn representations that capture evolutionary and functional constraints. The delta-embedding, the difference between the embedding of the wild-type sequence and the mutant sequence, encodes the mutation's effect in a high-dimensional space that is sensitive to the local sequence context. This is exactly the information needed for resistance prediction.

2. Prior Work and Negative Results

My path to TOPOLOGIX went through two falsified hypotheses. The first was that topological features of the protein-ligand interface, computed via bipartite persistent homology, would predict hERG cardiotoxicity. I pre-registered the study, powered it, and ran it. The topological features did not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782. This was a null result, but it was a useful one: it settled a comparison the published literature had never actually run, and it established my commitment to pre-registration and honest reporting.

The second was that the same topological constructs would predict drug resistance. I applied interface bipartite persistent homology and element-specific persistent homology to the Platinum benchmark. The results were AUROC 0.425 and 0.485, essentially no signal. This ruled out interface geometry as the driver of resistance prediction and motivated a different approach. The negative results are the reason TOPOLOGIX exists.

3. TOPOLOGIX: Methods and Current Results

TOPOLOGIX represents each mutation as a delta-embedding from ESM-2, the difference between the wild-type and mutant sequence embeddings. This captures the mutation's effect on the protein's evolutionary and functional context. The drug is represented by a Morgan/ECFP fingerprint. The concatenated representation is fed to a Random Forest classifier.

On the Platinum benchmark, 553 mutations, TOPOLOGIX achieves an AUROC of 0.804 with a standard deviation of 0.025 across cross-validation folds. On SKEMPI 2.0, a binding affinity benchmark, it achieves 0.634. It covers 100 percent of mutations in the benchmark, versus roughly 18 percent for structure-limited tools. It beats the mCSM-lig structure-based baseline of approximately 0.70. The model is fast, fully open-source, and requires no structural data.

4. Aim 1: Multi-Drug Resistance Prediction

The current TOPOLOGIX model predicts binary resistance, resistant or not. The next step is to predict which drugs a mutation defeats. This is a multi-label classification problem: a single mutation may confer resistance to multiple drugs with different mechanisms. I will extend the model to output a resistance profile across a panel of drugs, using the drug fingerprint as a conditioning variable. The expected outcome is a model that can rank drugs by the likelihood that a given mutation defeats them, enabling drug development to prioritize compounds that are strong to known resistance mutations.

5. Aim 2: Incorporating Evolutionary Context

Resistance mutations do not arise in isolation. They occur in the context of a protein family's evolutionary history, and their effects depend on the background in which they appear. I will incorporate phylogenetic signal from the protein family into the model, using the ESM-2 embedding of the full protein sequence as an additional input. This should improve generalization to unseen mutations, because the model will learn which sequence contexts are permissive for resistance. The expected outcome is a model that predicts resistance for mutations in proteins that are evolutionarily distant from the training set.

6. Aim 3: Prospective Validation

The ultimate test of a predictive model is prospective: can it predict mutations that were not in the training set and have been characterized experimentally? I will assemble a set of resistance mutations that have been published after the Platinum benchmark was compiled, withhold them from training, and evaluate the model's predictions. This is a strict test of generalization, and it is the standard that the field should demand. The expected outcome is a published validation study that either confirms the model's transferability or identifies its failure modes.

7. Timeline and Feasibility

The work is feasible within the fellowship period. Aim 1 is a straightforward extension of the current model and should be complete within six months. Aim 2 requires additional data curation and should be complete within twelve months. Aim 3 requires identifying and curating a prospective validation set, which is the most time-consuming step, and should be complete within eighteen months. The remaining time will be spent on writing, publication, and dissemination.

8. Fit with the AI-in-Science Remit

This proposal is squarely within the AI-in-Science remit as defined by the fellowship. It uses AI, in the form of protein language models, to advance basic biomedical science, specifically the molecular biology of drug resistance. It does not involve patient data, medical imaging, or clinical decision-making. It is a fundamental scientific question: how does sequence encode resistance? The answer has implications for drug development and antimicrobial resistance surveillance, but the research itself is basic science.

9. Fit with AIMS and Imperial College

AIMS has a mission to advance mathematical and computational science in Africa. This proposal is computational, open-source, and directly relevant to African public health challenges. Imperial College has a strong tradition in computational biology and antimicrobial resistance research. The combination of AIMS's African network and Imperial's computational biology expertise is uniquely suited to this project. I am applying to both fellowships as required.

PUBLICATION ELABORATION

My publication record is unusual for an independent researcher without a PhD. It consists of three sole-authored preprints, each under review at a peer-reviewed journal, plus one co-authored paper under review. I have listed them below with a description of my contribution and the current status.

1. Conjunctive Consolidation Threshold (CCT) Model of Reward-Memory Encoding Prevention in Addiction. This is a tripartite pharmacological framework that models the interaction of dopaminergic reward prediction error, NMDAR-dependent long-term potentiation, and affective contrast in the encoding of reward memories. The model is a coupled three-axis ODE system solved with RK45, calibrated with Bayesian MCMC using PyMC's DEMetropolisZ sampler, with 14 free parameters and literature-elicited priors from a screen of 1,847 records. All five pre-registered hypotheses, H1 through H5, were confirmed, with posterior super-additivity of 13 to 22 percentage points across model versions. The preprint is under review at IART, PNPBP, and NBR. I am the sole author and performed all modeling, calibration, and analysis.

2. Cardiotoxicity Topology Study. This is a pre-registered, powered replication testing whether bipartite persistent homology predicts hERG cardiotoxicity from protein-ligand interface geometry. The result was negative: topological features did not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782. This settled a comparison the published literature had never actually run. I am the sole author and performed all computational work.

3. Interface Topology for Drug Resistance Prediction. This study applied the same topological constructs to drug resistance prediction and found they carried almost no signal, AUROC 0.425 and 0.485 on the Platinum benchmark. This negative result ruled out interface geometry as the driver and motivated the sequence-representation approach that became TOPOLOGIX. I am the sole author.

4. Co-authored paper in Alcohol (Elsevier), under review. This paper applies my modeling expertise to an alcohol-related research question. My contribution was the computational modeling and analysis.

I have also presented my work at conferences and maintain an active GitHub repository with all code and data. My ORCID is 0009-0001-9272-6735.

KINDNESS STATEMENT

I have been the recipient of unsolicited kindness at every stage of my independent research career. Kent Berridge responded to a cold email from a Nigerian pharmacist and engaged seriously with my model. Samuel Gershman endorsed my arXiv submission without knowing me. Nathaniel Daw and Marcelo Mattar took time to discuss my work. I try to pay this forward. I maintain my code and data openly, I respond to every request for help or clarification, and I have mentored younger students in computational methods. Kindness in science is the mechanism by which the field admits outsiders, not a soft skill. I intend to be part of that mechanism.

CURRICULUM VITAE

Eniola Ayodele Olutogun
Independent researcher
ORCID: 0009-0001-9272-6735
GitHub: github.com/AmunRaPtah
Personal site: zyco.org
Nationality: Nigerian
Age: 29

Education
- B.Pharm, University of Ibadan, 2014-2021. CGPA 5.1/7.0 (2:1 Upper Division), German equivalent 1.9. PCN-licensed pharmacist.
- Enrolled, M.Sc. Digital Health, Hasso Plattner Institute / University of Potsdam, Germany, Winter Semester 2026/27.

Research Lines
- CCT model: tripartite pharmacological framework for reward-memory encoding prevention in addiction. Three coupled-axis ODE model, Bayesian MCMC calibration, all five pre-registered hypotheses confirmed. Three sole-authored preprints under review.
- Cardiotoxicity topology study: pre-registered replication showing bipartite persistent homology does not beat descriptor baseline for hERG cardiotoxicity.
- Interface topology for resistance: showed topological constructs carry no signal for drug resistance prediction.
- TOPOLOGIX: ESM-2 delta-embeddings plus Morgan fingerprints, Random Forest classifier, AUROC 0.804 on Platinum benchmark.
- neurocascade: receptor-to-behavior brain-circuit simulation engine, 62/62 tests passing.
- ergofluids: Koopman-operator methods with Mori-Zwanzig memory kernel for drug transport modeling. Pre-registered validation pipeline; first real-data gate did not meet criterion, reported directly.
- psyche-twin: typed multi-scale knowledge-graph architecture for self-modeling.

Employment
- National Product Manager, Synthcare, March 2026 to present.
- Clinical Pharmacist, Ramset Pharmacy, January to March 2026.
- Research Assistant, CDDDP, NMDA/insulin docking.
- Bioinformatics Researcher, GHRU-GSAR, AMR genomics and surveillance pipeline.

Skills
- Python (scipy, numpy, ODE/RK45, PyMC/MCMC, pandas), R.
- Topological data analysis (Ripser, Gudhi).
- NEURON/Brian2, AlphaFold, RDKit, ADMET/QSAR, GROMACS, AutoDock.
- Nextflow/SLURM/HPC, Supabase/Postgres, JavaScript/Node.js.
- DuckDB-based RAG pipelines, local LLM serving (llama.cpp), production systems ops.

Endorsements
- Kent Berridge, University of Michigan.
- Samuel Gershman, Harvard University, arXiv endorsement.
- Nathaniel Daw, Princeton University.
- Marcelo Mattar, NYU.

EMAIL DRAFT

Subject: Application for Complexity Cat Postdoctoral Fellowship, TOPOLOGIX project

Dear Complexity Cat selection committee,

I am applying for the Complexity Cat postdoctoral fellowship at AIMS and Imperial College London. I do not hold a PhD, but I am an independent computational researcher with a publication record and a project that fits the AI-in-Science remit.

My current project, TOPOLOGIX, predicts drug-resistance mutations from protein sequence alone using ESM-2 protein language model delta-embeddings and Morgan fingerprints. On the Platinum benchmark, it achieves an AUROC of 0.804, covers 100 percent of mutations versus roughly 18 percent for structure-based tools, and outperforms the mCSM-lig baseline of approximately 0.70. This is a basic molecular biology result, not a clinical application. It addresses antimicrobial resistance, a critical problem in Africa.

My prior work includes a pre-registered replication that overturned a published claim about topological features predicting hERG cardiotoxicity, and a follow-up showing the same constructs carry no signal for resistance prediction. These negative results motivated the sequence-based approach in TOPOLOGIX.

I have three sole-authored preprints under review, a co-authored paper under review at Alcohol (Elsevier), and endorsements from Kent Berridge, Samuel Gershman, Nathaniel Daw, and Marcelo Mattar. My code is fully open-source.

I understand the PhD requirement. I am asking you to consider whether my research record and the fit of this project justify an exception. I am applying to both AIMS and Imperial College fellowships as required.

Full application materials are attached.

Sincerely,
Eniola Ayodele Olutogun

CHECKLIST

- [ ] Verify the PhD requirement and whether an exception is possible; if not, confirm whether the programme accepts pre-PhD applicants.
- [ ] Confirm the application deadline on the programme website.
- [ ] Confirm the requirement to apply to both AIMS and Imperial College fellowships, and prepare separate applications for each.
- [ ] Prepare full CV in the programme's required format.
- [ ] Prepare 1-page publication elaboration.
- [ ] Prepare 1-page research proposal summary.
- [ ] Prepare 3-page research proposal.
- [ ] Prepare quarter-page kindness statement.
- [ ] Prepare email draft and send to the address specified on the programme website.
- [ ] Verify all facts in the application: AUROC values, benchmark names, journal names, endorsement names.
- [ ] Confirm the status of the M.Sc. enrollment at HPI/Potsdam and whether it affects eligibility.
- [ ] Confirm the status of the three preprints under review and the co-authored paper under review at Alcohol.

EDITOR NOTES

- Eligibility risk: The programme requires a PhD by 1 September 2027. The applicant does not hold a PhD and is enrolled in an M.Sc. This is the single largest risk. The motivation letter addresses it head-on, but the applicant should verify whether the programme has any flexibility or whether a PhD-equivalent record is accepted.
- Fact verification: All numbers in the application (AUROC values, benchmark names, journal names, endorsement names) must be verified against the applicant's actual records before submission. The profile lists specific values, but the applicant should confirm they are current and accurate.
- Gap: The applicant's employment history and current role at Synthcare are not described in detail. If the programme asks about current employment or time commitment, the applicant must insert that detail.
- Gap: The applicant's relationship with the named endorsers (Berridge, Gershman, Daw, Mattar) is described as endorsements, but the nature and depth of those relationships is not specified. The applicant should be prepared to describe them if asked.
- The research proposal is framed around TOPOLOGIX, consistent with the recommended framing angle. The negative results from the topology studies are presented as motivation, not as current work, which is the correct framing.

AI Draft — African Institute for Mathematical Sciences | Complexity Cat