MOTIVATION LETTER Drug resistance is a sequence problem, not a protein-structure problem. When a mutation arises in a target protein, the structure often shifts too little to explain the loss of drug binding, and structure-based tools like mCSM-lig cover only about 18 percent of clinically observed mutations because they require a resolved crystal structure. My current project, TOPOLOGIX, treats resistance prediction as a language-model task: I feed ESM-2 protein language model delta-embeddings, which capture the semantic shift between wild-type and mutant sequences, together with Morgan/ECFP drug fingerprints, into a Random Forest classifier. On the Platinum benchmark of 553 mutations, TOPOLOGIX achieves an AUROC of 0.804 with a standard deviation of 0.025, and 0.634 on SKEMPI 2.0. It beats the structure-based baseline of roughly 0.70 while covering 100 percent of mutations. This is the project I bring to the Transfyr AI Fellowship Program because it is exactly what the programme's mission describes: AI applied to scientific practice, with a measurable improvement over the prior method and a clear path to clinical use. The Transfyr AI Fellowship Program funds researchers who build AI systems that change how science is done. TOPOLOGIX is a full pipeline I built myself: data ingestion from Platinum and SKEMPI 2.0, sequence embedding generation through the ESM-2 model, fingerprint computation with RDKit, model training and cross-validation, and a pre-registered evaluation protocol. The entire codebase is on my public GitHub. The design decision that matters most is the delta-embedding: instead of classifying a single sequence, I subtract the wild-type embedding from the mutant embedding, isolating the mutation's effect in the language model's representation space. That single choice is what lets the model generalize across protein families without retraining. My path to this work is unusual and I treat it as an advantage. I am a licensed pharmacist with a B.Pharm from the University of Ibadan, where I graduated with a CGPA of 5.1 out of 7.0, equivalent to a German 1.9. I have worked as a clinical pharmacist and as a national product manager. I have also built four independent data pipelines, self-hosted LLM serving infrastructure, and production systems on Linux VPS. This combination means I can take TOPOLOGIX from a research prototype to a deployed tool that a clinical lab or a drug discovery team can actually run. I am currently enrolled in the M.Sc. Digital Health program at the Hasso Plattner Institute and the University of Potsdam, starting in the Winter Semester 2026/27, which gives me a formal institutional base in Germany while I continue independent research. The fellowship's selection criteria emphasize demonstrated research excellence and clear articulation of practical impact. My record includes three sole-authored preprints under review at peer-reviewed journals, a co-authored paper under review at Alcohol (Elsevier), and a pre-registered replication study on hERG cardiotoxicity topology that settled a comparison the literature had never actually run. I also know what a negative result looks like and I report it directly: my ergofluids project on Koopman-operator methods for drug transport failed its first real-data validation gate, and I published that failure rather than reframing it. That rigor is the same standard I apply to TOPOLOGIX. The next step for TOPOLOGIX is to move from benchmark accuracy to prospective validation on a clinical cohort of resistant bacterial isolates, and to extend the model from single mutations to combination resistance. The Transfyr fellowship would fund the compute, the dataset acquisition, and the time to do that work properly. I am asking for the support to build the tool that makes resistance prediction routine. RESEARCH STATEMENT TOPOLOGIX: Predicting Drug Resistance Mutations from Protein Language Model Delta-Embeddings Problem and Motivation Drug resistance is the single largest cause of treatment failure in infectious disease and oncology. When a pathogen or tumor acquires a resistance mutation, the standard clinical response is to switch drugs, but the choice of the next drug is often empirical. Computational tools that predict which mutations cause resistance could guide that choice, but the current generation of tools has a structural bottleneck. Methods like mCSM-lig require a resolved crystal structure of the protein-ligand complex. In practice, only about 18 percent of clinically observed mutations have such a structure available. The remaining 82 percent are simply not analyzable by structure-based tools. This is not a marginal gap; it is the difference between a tool that works in a curated benchmark and a tool that works in a clinic. My prior work established that the structural approach has a hard ceiling. In a pre-registered, powered replication study on hERG cardiotoxicity, I tested whether bipartite persistent homology features, computed with Ripser and GUDHI, could predict cardiotoxicity from protein-ligand interface geometry. The topological features achieved an AUROC of 0.8426, but a plain descriptor baseline achieved 0.8782. The topological approach did not win. I then applied the same topological constructs to drug-resistance prediction on the Platinum benchmark and found the signal was even weaker: AUROC of 0.425 and 0.485. Interface geometry is not the driver of resistance. That negative result is the direct motivation for TOPOLOGIX. If structure does not carry the signal, the signal must be in the sequence. Method TOPOLOGIX uses a three-part architecture. First, I generate protein language model embeddings with ESM-2, a 650-million-parameter transformer trained on 250 million protein sequences. For each mutation, I compute the embedding of the wild-type sequence and the embedding of the mutant sequence, then subtract them to obtain a delta-embedding. This delta-embedding isolates the representation shift caused by the mutation, independent of the surrounding sequence context. Second, I compute Morgan/ECFP fingerprints for the drug molecule using RDKit, capturing the chemical structure of the ligand. Third, I concatenate the delta-embedding and the fingerprint and feed the combined vector into a Random Forest classifier. The model is trained on the Platinum benchmark of 553 mutations and evaluated with cross-validation. The results are as follows. On Platinum, TOPOLOGIX achieves an AUROC of 0.804 with a standard deviation of 0.025. On SKEMPI 2.0, a transfer benchmark of binding affinity changes, it achieves 0.634. The structure-based baseline, mCSM-lig, achieves approximately 0.70 on the subset of mutations for which structures exist, but it covers only 18 percent of the benchmark. TOPOLOGIX covers 100 percent. The comparison is not apples-to-apples in the strict sense, because the coverage differs, but that is precisely the point: a tool that covers every mutation with a slightly lower AUROC is more useful in practice than a tool that covers a fifth of mutations with a slightly higher AUROC. And on the full benchmark, TOPOLOGIX is the better predictor. Why This Is a Frontier ML Contribution The delta-embedding approach is the key technical novelty. Most sequence-based resistance predictors classify the mutant sequence directly, treating the wild-type as background noise. By subtracting the wild-type embedding, I force the model to attend only to the change. This is analogous to how contrastive learning isolates differences between pairs of inputs, but applied to protein language model representations. The approach is also computationally cheap: ESM-2 embeddings are precomputed, and the Random Forest trains in minutes on a laptop. This means the method is accessible to labs in low-resource settings, including in Nigeria, where I am from and where drug resistance surveillance is a critical need. Validation Status and Next Steps The model is validated on two public benchmarks with pre-registered evaluation protocols. The code is public on GitHub. The next phase has three components. First, prospective validation on a clinical cohort of resistant bacterial isolates, which requires acquiring whole-genome sequences with matched phenotypic resistance data. Second, extension from single mutations to combination resistance, where multiple mutations act epistatically. Third, deployment as a web service with a simple interface, so that a clinical microbiologist can paste a sequence and a drug name and get a resistance probability. The Transfyr AI Fellowship Program's support would fund the dataset acquisition, the compute for the extension to combination resistance, and the development of the deployment interface. I am proposing a tool that changes the workflow of resistance prediction, and I have the track record to build it end to end. ESSAY: INDEPENDENT RESEARCH AND REPRODUCIBILITY The Transfyr AI Fellowship Program's selection criteria emphasize demonstrated research excellence and the ability to communicate ideas clearly. My record as an independent researcher demonstrates both, and it also demonstrates something the criteria do not explicitly name but that I consider essential: the willingness to report a result that goes against my own hypothesis. In 2025, I ran a pre-registered replication study on hERG cardiotoxicity. The published literature contained claims that topological data analysis of protein-ligand interfaces could predict cardiotoxicity, but no one had actually run the comparison against a plain descriptor baseline. I ran it. The topological features achieved an AUROC of 0.8426. The plain baseline achieved 0.8782. The topological approach lost. I reported the result as a negative finding, because that is what it was. I then applied the same topological methods to drug-resistance prediction and found they carried almost no signal, with AUROCs of 0.425 and 0.485 on the Platinum benchmark. Those two negative results are what pushed me to try a sequence-representation approach, which became TOPOLOGIX. This pattern, negative result, pivot, better method, is the actual shape of scientific progress. It is not the shape that gets celebrated in press releases, but it is the shape that produces reliable tools. The Transfyr fellowship's mission of transforming scientific practice via AI requires researchers who can distinguish a real signal from a fashionable method. I have done that repeatedly, and I have the pre-registration records and public code to prove it. My independence is also practical. I have built four DuckDB-based data pipelines across life sciences, tech, and social science domains. I self-host local LLM serving with llama.cpp. I run production systems on Linux VPS with systemd, Caddy TLS, and automated backups. I do not need an institution to provide me with infrastructure; I need funding to scale the work I am already doing. The M.Sc. in Digital Health at the Hasso Plattner Institute, starting in Winter Semester 2026/27, gives me an institutional anchor, but my research identity is independent. That independence is why I can move from a failed hypothesis to a working model without institutional friction. I am asking the Transfyr AI Fellowship Program to fund a researcher who has already demonstrated the ability to produce rigorous, reproducible, and useful AI research, and who has a clear plan for the next phase. The track record is public. The code is public. The results are reproducible. That is the standard I hold myself to, and it is the standard the fellowship should hold its fellows to. CHECKLIST - [ ] Confirm Transfyr AI Fellowship Program application deadline and submission portal at https://utibeetim.com/transfyr-ai-fellowship-program/ - [ ] Verify whether the programme requires a CV or resume; if so, prepare a 2-page CV emphasizing TOPOLOGIX, the hERG replication study, and the three sole-authored preprints - [ ] Verify whether the programme requires letters of recommendation; if so, request from Kent Berridge (Michigan), Samuel Gershman (Harvard), or Nathaniel Daw (Princeton), all of whom have endorsed prior work - [ ] Confirm the exact word limits for the motivation letter and research statement on the submission portal; adjust the drafts above to match if limits differ - [ ] Prepare a public link to the TOPOLOGIX GitHub repository (github.com/AmunRaPtah) and verify the README includes the Platinum and SKEMPI 2.0 results with AUROC values - [ ] Prepare a public link to the pre-registration records for the hERG replication study and the TOPOLOGIX evaluation protocol - [ ] Verify the ORCID record (0009-0001-9272-6735) is up to date with all preprints and the Alcohol (Elsevier) co-authored paper - [ ] Confirm enrollment status at Hasso Plattner Institute / University of Potsdam for Winter Semester 2026/27 and have the enrollment letter ready - [ ] Draft a one-paragraph plain-language summary of TOPOLOGIX for the application form, avoiding all technical jargon, in case the portal requires a separate abstract - [ ] Confirm whether the programme requires a budget; if so, prepare a budget covering dataset acquisition for clinical resistance cohorts, compute for ESM-2 embedding generation, and conference travel EDITOR NOTES - Eligibility risk: The Transfyr AI Fellowship Program page does not list explicit eligibility criteria. Eniola is a Nigerian citizen enrolled in a German M.Sc. program, which should satisfy most early-career and LMIC-track criteria, but the applicant must verify whether the programme restricts applicants to specific countries or institutional affiliations. This is the single largest risk and must be checked before submission. - Fact verification: The profile lists the hERG replication AUROC as 0.8426 for topological features and 0.8782 for the descriptor baseline. These numbers appear in the motivation letter and research statement. The applicant must confirm these match the pre-registration and the preprint exactly, as any discrepancy will be caught by reviewers. - Fact verification: The Platinum benchmark AUROC for TOPOLOGIX is stated as 0.804 with a standard deviation of 0.025, and SKEMPI 2.0 as 0.634. The mCSM-lig baseline is stated as approximately 0.70 with 18 percent coverage. These numbers must be verified against the actual benchmark results and the mCSM-lig publication before submission. - Gap: The motivation letter references "a clinical cohort of resistant bacterial isolates" as the next validation step, but the profile does not specify which bacterial species, which drugs, or which dataset the applicant intends to acquire. The applicant must insert a concrete plan, even a provisional one, or the proposal will read as vague. - Gap: The essay mentions the M.Sc. at Hasso Plattner Institute as starting in Winter Semester 2026/27. The applicant must confirm the exact start date and whether the fellowship period would overlap with the first semester of the M.Sc., as this could affect time availability. - Tone check: The motivation letter opens with "Drug resistance is not a protein-structure problem," which is a strong claim. The applicant should be prepared to defend this in an interview, especially against researchers who work on structure-based methods. The claim is supported by the 18 percent coverage figure, but the applicant should have the citation ready.