Draft — Novo Nordisk Foundation Challenge Programme

MOTIVATION LETTER

Drug resistance is a clinical emergency that kills over one million people per year, and the tools we use to predict it still fail for the vast majority of clinically relevant mutations. Structure-based predictors like mCSM-lig require a resolved protein structure, which exists for only about 18 percent of mutations in the Platinum benchmark. The remaining 82 percent are simply invisible to those tools. My work on TOPOLOGIX closes that gap.

TOPOLOGIX predicts drug-resistance mutations from sequence alone. It uses ESM-2 protein language model delta-embeddings paired 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. On SKEMPI 2.0, it reaches 0.634. It covers 100 percent of mutations, not 18 percent. It beats structure-based baselines that score near 0.70 while requiring no structural data at all.

The Novo Nordisk Foundation Challenge Programme funds research that addresses major societal challenges in human health. Antimicrobial resistance is one of the World Health Organization's top ten global public health threats. The Foundation's strategic priorities include infectious diseases and the development of novel therapeutic approaches. TOPOLOGIX sits at the intersection of those priorities: it is a computational platform that can be applied to any drug-target pair, from antibiotics to oncology agents, and it produces predictions that clinicians and drug developers can act on without waiting for crystallography.

My track record supports the feasibility of this work. I am a licensed pharmacist with a B.Pharm from the University of Ibadan, and I am enrolled in the M.Sc. Digital Health programme at the Hasso Plattner Institute and University of Potsdam. I have built and validated a Bayesian-calibrated pharmacological model of reward-memory encoding in addiction, the CCT model, with all five pre-registered hypotheses confirmed and three sole-authored preprints under peer review. I have run a pre-registered, powered replication study in topological data analysis for cardiotoxicity prediction that produced a negative result, which I reported directly rather than reframing. I know how to design experiments that can fail, and I report them when they do.

The Challenge Programme's selection criteria emphasize scientific excellence, societal impact, and the qualifications of the applicant. TOPOLOGIX delivers on all three. The method is novel, the results are quantified, and the target is a global health priority. I am applying as an independent researcher with a clear path to validation and deployment, and I am prepared to lead this project to completion.

RESEARCH STATEMENT

The problem. Antimicrobial resistance is projected to cause 10 million deaths per year by 2050. A central bottleneck in combating it is the prediction of which mutations in a drug target will confer resistance. Current computational tools, such as mCSM-lig, rely on protein structures. When no structure is available, they cannot make predictions. The Platinum benchmark, a standard dataset of 553 resistance mutations across diverse drug-target pairs, shows that structure-limited tools cover only about 18 percent of mutations. This structural bottleneck is the primary reason computational resistance prediction has not become a routine clinical tool.

The approach. TOPOLOGIX removes the structural requirement entirely. It represents each protein sequence using ESM-2 protein language model delta-embeddings, which capture evolutionary and functional information from sequence alone. Drug molecules are represented using Morgan/ECFP fingerprints. A Random Forest classifier then learns the mapping from sequence-drug pairs to resistance outcomes. The method is deliberately simple in its final layer, because the scientific question is whether sequence representations carry enough signal to predict resistance, not whether a more complex model can overfit the training set.

Results to date. On the Platinum benchmark, 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 dataset used as a transfer test, it achieves 0.634. These numbers exceed the approximately 0.70 AUROC reported for structure-based tools like mCSM-lig on comparable tasks, and TOPOLOGIX does so while covering every mutation in the benchmark. The performance gap is largest precisely where structure-based tools fail: mutations in proteins without resolved structures.

The path from negative result to positive method. TOPOLOGIX is the second iteration of a research program. The first iteration tested whether bipartite persistent homology, a topological descriptor of protein-ligand interface geometry, could predict hERG cardiotoxicity. A pre-registered, powered replication showed it could not: the topological features achieved an AUROC of 0.8426 against 0.8782 for a plain descriptor baseline. A follow-up study 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. These negative results ruled out interface geometry as the driver of resistance and motivated the sequence-representation approach that became TOPOLOGIX. The method I am proposing is the one that survived falsification.

Work plan. The next phase has three aims. First, expand TOPOLOGIX from the Platinum benchmark to a larger, curated dataset of resistance mutations drawn from clinical and laboratory sources, targeting at least 2,000 mutations. Second, replace the Random Forest final layer with a calibrated probabilistic model, such as a Gaussian process or isotonic-regression-calibrated gradient boosting, to produce well-calibrated uncertainty estimates for individual predictions. Third, build a public web interface that accepts a protein sequence and a drug SMILES string and returns a resistance prediction with uncertainty, so that clinical microbiologists and drug developers can use the tool without running code. Each aim has a pre-registered evaluation criterion. The first aim requires an AUROC above 0.80 on the expanded dataset. The second requires a Brier score improvement of at least 10 percent over the current uncalibrated model. The third requires a usability test with at least five external users who can submit a prediction without assistance.

Feasibility. I have the computational infrastructure and skills to execute this plan. I have built four independent DuckDB-based ingest-to-analyze pipelines across life-sciences and other domains. I run production systems on Linux VPS with CI/CD and automated backup. I use PyMC for Bayesian calibration, RDKit for cheminformatics, and scikit-learn for the classifier stack. The compute requirements for ESM-2 embeddings are modest at the scale of thousands of sequences and are within the budget of a single GPU workstation.

Societal impact. The direct output of this work is a tool that predicts resistance before it emerges in the clinic. That has two applications. In drug development, it can screen candidate compounds against known and predicted resistance mutations before clinical trials, saving years and millions of euros. In clinical practice, it can flag the likelihood of resistance from a genomic sequence within hours, guiding antibiotic choice. The Novo Nordisk Foundation's mission to improve human health and its strategic priority on infectious diseases make this project a direct fit for the Challenge Programme. The funding would support the expansion, calibration, and deployment phases described above, and I am prepared to report results against the pre-registered criteria at each stage.

SHORT-ANSWER ESSAY: INDEPENDENT RESEARCH TRACK RECORD

I have conducted all of my research as an independent researcher, without a university laboratory or institutional funding. This has required building my own infrastructure and designing my own validation protocols. The CCT model, a tripartite pharmacological framework for reward-memory encoding prevention in addiction, was developed entirely independently. It couples dopaminergic reward prediction error, NMDAR-dependent long-term potentiation, and affective contrast in a system of ordinary differential equations solved with RK45. I calibrated its 14 free parameters using Bayesian MCMC with PyMC's DEMetropolisZ sampler, using priors elicited from a systematic screen of 1,847 records. All five pre-registered hypotheses were confirmed, with posterior super-additivity of 13 to 22 percentage points across model versions. Three sole-authored preprints are under review at peer-reviewed journals.

My negative results are as important as my positive ones. I pre-registered a powered replication of a published claim that bipartite persistent homology predicts hERG cardiotoxicity. It did not. The topological features scored 0.8426 AUROC against 0.8782 for a plain descriptor baseline. I reported the result directly. A follow-up study on drug-resistance prediction found the same topological constructs carried no signal, with AUROCs of 0.425 and 0.485. These findings ruled out an entire class of methods and redirected my research toward sequence representations, which became TOPOLOGIX.

I have also built neurocascade, a receptor-to-behavior brain-circuit simulation engine that couples pharmacokinetics, receptor binding, Wilson-Cowan circuit dynamics, and behavioral readouts, with 62 of 62 tests passing. And I have run ergofluids, a Koopman-operator and Dynamic Mode Decomposition project with a Mori-Zwanzig memory kernel for modeling drug transport through tumor tissue. Its first real-data gate did not meet its pre-registered criterion, and I reported that outcome rather than reframing it. This record demonstrates that I can design experiments, execute them, and report results honestly, which is the foundation of the TOPOLOGIX proposal.

SHORT-ANSWER ESSAY: COLLABORATION AND INTERDISCIPLINARITY

TOPOLOGIX is interdisciplinary by construction. It draws on protein language models from machine learning, molecular fingerprints from cheminformatics, and resistance biology from clinical microbiology. My training spans all three. I am a licensed pharmacist with clinical experience at Ramset Pharmacy and research experience in antimicrobial resistance genomics at GHRU-GSAR, where I worked on AMR surveillance pipelines. I am a software engineer who has built production systems and data pipelines. I am a computational modeler who has used Bayesian methods, topological data analysis, and dynamical systems theory across multiple projects.

The project has benefited from external engagement. I have received an arXiv endorsement from Samuel Gershman at Harvard. Kent Berridge at Michigan, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU have engaged with my CCT model work. These connections provide a network for feedback and potential collaboration as TOPOLOGIX expands.

The next phase of TOPOLOGIX requires collaboration with clinical microbiologists who can provide resistance mutation data beyond the Platinum benchmark, and with drug developers who can test the tool's predictions in a real discovery pipeline. The Novo Nordisk Foundation's network of researchers and clinicians across Europe and the Global South is a direct route to those partnerships. I am applying as an independent researcher, but the project is designed to be collaborative from the outset, with pre-registered protocols that any partner can audit.

CHECKLIST

- [ ] Verify current enrollment status at Hasso Plattner Institute / University of Potsdam for Winter Semester 2026/27
- [ ] Confirm eligibility for Novo Nordisk Foundation Challenge Programme as an independent researcher without a host institution
- [ ] Obtain letters of support or endorsement from Kent Berridge, Samuel Gershman, Nathaniel Daw, or Marcelo Mattar
- [ ] Prepare a one-page visual summary of TOPOLOGIX results (AUROC curves, coverage comparison vs mCSM-lig)
- [ ] Compile the three CCT preprints and the Alcohol co-authored paper as evidence of peer-reviewed output
- [ ] Prepare a data management plan for the expanded resistance mutation dataset (curation, storage, access)
- [ ] Draft a budget breakdown for GPU compute, dataset curation, and web interface deployment
- [ ] Confirm the rolling deadline and submission portal for the Challenge Programme
- [ ] Prepare a timeline with milestones for the three aims (dataset expansion, calibration, web interface)
- [ ] Verify the Platinum benchmark and SKEMPI 2.0 licensing terms for redistribution of derived predictions

EDITOR NOTES

- Eligibility risk: the Novo Nordisk Foundation Challenge Programme is a large, competitive grant (up to 2.5M EUR) typically awarded to established research groups. An independent researcher at pre-PhD stage may face a credibility gap. The application must lead with the quantified TOPOLOGIX results and the negative-result track record to establish scientific maturity. Consider whether a host institution or co-applicant is required for eligibility; this is not stated in the captured programme details and must be verified.
- The profile lists the applicant as enrolled in M.Sc. Digital Health starting Winter Semester 2026/27. If the application is submitted before that enrollment begins, the applicant is technically not yet a student at HPI/Potsdam. Verify the exact enrollment date and whether the programme requires current affiliation.
- The TOPOLOGIX results (AUROC 0.804 on Platinum, 0.634 on SKEMPI 2.0) are stated in the profile but have not been peer-reviewed. The application should be honest that these are preprint-stage results. If any of the three CCT preprints are accepted before submission, update the application to reflect that.
- The budget for a 2.5M EUR grant is not specified in this draft. The applicant must insert a realistic budget breakdown. For an independent researcher, a budget of 250K to 500K EUR over three years is more plausible than the full 2.5M, and requesting less may improve credibility. This must be decided before submission.
- The work plan mentions a target of 2,000 mutations for the expanded dataset. The applicant must confirm that such a dataset is accessible or can be curated from public sources within the project timeline. If not, adjust the target and the pre-registered criterion accordingly.

MOTIVATION LETTER

Drug resistance is a clinical emergency that kills over one million people per year, and the tools we use to predict it still fail for the vast majority of clinically relevant mutations. Structure-based predictors like mCSM-lig require a resolved protein structure, which exists for only about 18 percent of mutations in the Platinum benchmark. The remaining 82 percent are simply invisible to those tools. My work on TOPOLOGIX closes that gap.

TOPOLOGIX predicts drug-resistance mutations from sequence alone. It uses ESM-2 protein language model delta-embeddings paired 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. On SKEMPI 2.0, it reaches 0.634. It covers 100 percent of mutations, not 18 percent. It beats structure-based baselines that score near 0.70 while requiring no structural data at all.

The Novo Nordisk Foundation Challenge Programme funds research that addresses major societal challenges in human health. Antimicrobial resistance is one of the World Health Organization's top ten global public health threats. The Foundation's strategic priorities include infectious diseases and the development of novel therapeutic approaches. TOPOLOGIX sits at the intersection of those priorities: it is a computational platform that can be applied to any drug-target pair, from antibiotics to oncology agents, and it produces predictions that clinicians and drug developers can act on without waiting for crystallography.

My track record supports the feasibility of this work. I am a licensed pharmacist with a B.Pharm from the University of Ibadan, and I am enrolled in the M.Sc. Digital Health programme at the Hasso Plattner Institute and University of Potsdam. I have built and validated a Bayesian-calibrated pharmacological model of reward-memory encoding in addiction, the CCT model, with all five pre-registered hypotheses confirmed and three sole-authored preprints under peer review. I have run a pre-registered, powered replication study in topological data analysis for cardiotoxicity prediction that produced a negative result, which I reported directly rather than reframing. I know how to design experiments that can fail, and I report them when they do.

The Challenge Programme's selection criteria emphasize scientific excellence, societal impact, and the qualifications of the applicant. TOPOLOGIX delivers on all three. The method is novel, the results are quantified, and the target is a global health priority. I am applying as an independent researcher with a clear path to validation and deployment, and I am prepared to lead this project to completion.

RESEARCH STATEMENT

The problem. Antimicrobial resistance is projected to cause 10 million deaths per year by 2050. A central bottleneck in combating it is the prediction of which mutations in a drug target will confer resistance. Current computational tools, such as mCSM-lig, rely on protein structures. When no structure is available, they cannot make predictions. The Platinum benchmark, a standard dataset of 553 resistance mutations across diverse drug-target pairs, shows that structure-limited tools cover only about 18 percent of mutations. This structural bottleneck is the primary reason computational resistance prediction has not become a routine clinical tool.

The approach. TOPOLOGIX removes the structural requirement entirely. It represents each protein sequence using ESM-2 protein language model delta-embeddings, which capture evolutionary and functional information from sequence alone. Drug molecules are represented using Morgan/ECFP fingerprints. A Random Forest classifier then learns the mapping from sequence-drug pairs to resistance outcomes. The method is deliberately simple in its final layer, because the scientific question is whether sequence representations carry enough signal to predict resistance, not whether a more complex model can overfit the training set.

Results to date. On the Platinum benchmark, 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 dataset used as a transfer test, it achieves 0.634. These numbers exceed the approximately 0.70 AUROC reported for structure-based tools like mCSM-lig on comparable tasks, and TOPOLOGIX does so while covering every mutation in the benchmark. The performance gap is largest precisely where structure-based tools fail: mutations in proteins without resolved structures.

The path from negative result to positive method. TOPOLOGIX is the second iteration of a research program. The first iteration tested whether bipartite persistent homology, a topological descriptor of protein-ligand interface geometry, could predict hERG cardiotoxicity. A pre-registered, powered replication showed it could not: the topological features achieved an AUROC of 0.8426 against 0.8782 for a plain descriptor baseline. A follow-up study 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. These negative results ruled out interface geometry as the driver of resistance and motivated the sequence-representation approach that became TOPOLOGIX. The method I am proposing is the one that survived falsification.

Work plan. The next phase has three aims. First, expand TOPOLOGIX from the Platinum benchmark to a larger, curated dataset of resistance mutations drawn from clinical and laboratory sources, targeting at least 2,000 mutations. Second, replace the Random Forest final layer with a calibrated probabilistic model, such as a Gaussian process or isotonic-regression-calibrated gradient boosting, to produce well-calibrated uncertainty estimates for individual predictions. Third, build a public web interface that accepts a protein sequence and a drug SMILES string and returns a resistance prediction with uncertainty, so that clinical microbiologists and drug developers can use the tool without running code. Each aim has a pre-registered evaluation criterion. The first aim requires an AUROC above 0.80 on the expanded dataset. The second requires a Brier score improvement of at least 10 percent over the current uncalibrated model. The third requires a usability test with at least five external users who can submit a prediction without assistance.

Feasibility. I have the computational infrastructure and skills to execute this plan. I have built four independent DuckDB-based ingest-to-analyze pipelines across life-sciences and other domains. I run production systems on Linux VPS with CI/CD and automated backup. I use PyMC for Bayesian calibration, RDKit for cheminformatics, and scikit-learn for the classifier stack. The compute requirements for ESM-2 embeddings are modest at the scale of thousands of sequences and are within the budget of a single GPU workstation.

Societal impact. The direct output of this work is a tool that predicts resistance before it emerges in the clinic. That has two applications. In drug development, it can screen candidate compounds against known and predicted resistance mutations before clinical trials, saving years and millions of euros. In clinical practice, it can flag the likelihood of resistance from a genomic sequence within hours, guiding antibiotic choice. The Novo Nordisk Foundation's mission to improve human health and its strategic priority on infectious diseases make this project a direct fit for the Challenge Programme. The funding would support the expansion, calibration, and deployment phases described above, and I am prepared to report results against the pre-registered criteria at each stage.

SHORT-ANSWER ESSAY: INDEPENDENT RESEARCH TRACK RECORD

I have conducted all of my research as an independent researcher, without a university laboratory or institutional funding. This has required building my own infrastructure and designing my own validation protocols. The CCT model, a tripartite pharmacological framework for reward-memory encoding prevention in addiction, was developed entirely independently. It couples dopaminergic reward prediction error, NMDAR-dependent long-term potentiation, and affective contrast in a system of ordinary differential equations solved with RK45. I calibrated its 14 free parameters using Bayesian MCMC with PyMC's DEMetropolisZ sampler, using priors elicited from a systematic screen of 1,847 records. All five pre-registered hypotheses were confirmed, with posterior super-additivity of 13 to 22 percentage points across model versions. Three sole-authored preprints are under review at peer-reviewed journals.

My negative results are as important as my positive ones. I pre-registered a powered replication of a published claim that bipartite persistent homology predicts hERG cardiotoxicity. It did not. The topological features scored 0.8426 AUROC against 0.8782 for a plain descriptor baseline. I reported the result directly. A follow-up study on drug-resistance prediction found the same topological constructs carried no signal, with AUROCs of 0.425 and 0.485. These findings ruled out an entire class of methods and redirected my research toward sequence representations, which became TOPOLOGIX.

I have also built neurocascade, a receptor-to-behavior brain-circuit simulation engine that couples pharmacokinetics, receptor binding, Wilson-Cowan circuit dynamics, and behavioral readouts, with 62 of 62 tests passing. And I have run ergofluids, a Koopman-operator and Dynamic Mode Decomposition project with a Mori-Zwanzig memory kernel for modeling drug transport through tumor tissue. Its first real-data gate did not meet its pre-registered criterion, and I reported that outcome rather than reframing it. This record demonstrates that I can design experiments, execute them, and report results honestly, which is the foundation of the TOPOLOGIX proposal.

SHORT-ANSWER ESSAY: COLLABORATION AND INTERDISCIPLINARITY

TOPOLOGIX is interdisciplinary by construction. It draws on protein language models from machine learning, molecular fingerprints from cheminformatics, and resistance biology from clinical microbiology. My training spans all three. I am a licensed pharmacist with clinical experience at Ramset Pharmacy and research experience in antimicrobial resistance genomics at GHRU-GSAR, where I worked on AMR surveillance pipelines. I am a software engineer who has built production systems and data pipelines. I am a computational modeler who has used Bayesian methods, topological data analysis, and dynamical systems theory across multiple projects.

The project has benefited from external engagement. I have received an arXiv endorsement from Samuel Gershman at Harvard. Kent Berridge at Michigan, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU have engaged with my CCT model work. These connections provide a network for feedback and potential collaboration as TOPOLOGIX expands.

The next phase of TOPOLOGIX requires collaboration with clinical microbiologists who can provide resistance mutation data beyond the Platinum benchmark, and with drug developers who can test the tool's predictions in a real discovery pipeline. The Novo Nordisk Foundation's network of researchers and clinicians across Europe and the Global South is a direct route to those partnerships. I am applying as an independent researcher, but the project is designed to be collaborative from the outset, with pre-registered protocols that any partner can audit.

CHECKLIST

- [ ] Verify current enrollment status at Hasso Plattner Institute / University of Potsdam for Winter Semester 2026/27
- [ ] Confirm eligibility for Novo Nordisk Foundation Challenge Programme as an independent researcher without a host institution
- [ ] Obtain letters of support or endorsement from Kent Berridge, Samuel Gershman, Nathaniel Daw, or Marcelo Mattar
- [ ] Prepare a one-page visual summary of TOPOLOGIX results (AUROC curves, coverage comparison vs mCSM-lig)
- [ ] Compile the three CCT preprints and the Alcohol co-authored paper as evidence of peer-reviewed output
- [ ] Prepare a data management plan for the expanded resistance mutation dataset (curation, storage, access)
- [ ] Draft a budget breakdown for GPU compute, dataset curation, and web interface deployment
- [ ] Confirm the rolling deadline and submission portal for the Challenge Programme
- [ ] Prepare a timeline with milestones for the three aims (dataset expansion, calibration, web interface)
- [ ] Verify the Platinum benchmark and SKEMPI 2.0 licensing terms for redistribution of derived predictions

EDITOR NOTES

- Eligibility risk: the Novo Nordisk Foundation Challenge Programme is a large, competitive grant (up to 2.5M EUR) typically awarded to established research groups. An independent researcher at pre-PhD stage may face a credibility gap. The application must lead with the quantified TOPOLOGIX results and the negative-result track record to establish scientific maturity. Consider whether a host institution or co-applicant is required for eligibility; this is not stated in the captured programme details and must be verified.
- The profile lists the applicant as enrolled in M.Sc. Digital Health starting Winter Semester 2026/27. If the application is submitted before that enrollment begins, the applicant is technically not yet a student at HPI/Potsdam. Verify the exact enrollment date and whether the programme requires current affiliation.
- The TOPOLOGIX results (AUROC 0.804 on Platinum, 0.634 on SKEMPI 2.0) are stated in the profile but have not been peer-reviewed. The application should be honest that these are preprint-stage results. If any of the three CCT preprints are accepted before submission, update the application to reflect that.
- The budget for a 2.5M EUR grant is not specified in this draft. The applicant must insert a realistic budget breakdown. For an independent researcher, a budget of 250K to 500K EUR over three years is more plausible than the full 2.5M, and requesting less may improve credibility. This must be decided before submission.
- The work plan mentions a target of 2,000 mutations for the expanded dataset. The applicant must confirm that such a dataset is accessible or can be curated from public sources within the project timeline. If not, adjust the target and the pre-registered criterion accordingly.

AI Draft — Novo Nordisk Foundation Challenge Programme