MOTIVATION LETTER Antimicrobial resistance is projected to cause ten million deaths per year by 2050, with sub-Saharan Africa carrying a disproportionate share of that burden. In Nigeria, where I am licensed as a pharmacist and where routine clinical microbiology is scarce, resistance mutations are often detected only after treatment failure. My research addresses this gap directly: I built TOPOLOGIX, a machine learning pipeline that predicts drug-resistance mutations from protein sequence alone, using 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, TOPOLOGIX achieves an AUROC of 0.804 plus or minus 0.025, outperforming structure-based tools such as mCSM-lig at approximately 0.70 while covering 100 percent of mutations, compared to roughly 18 percent for structure-limited methods. This means a clinician or researcher with only a DNA sequence, no crystal structure, no purified protein, can ask whether a given mutation will confer resistance to a given drug. The Google Data Center Community AI Fellowship is the right home for this work because the project depends on exactly the kind of large-scale AI infrastructure Google operates. ESM-2 embeddings require substantial GPU compute for inference across whole proteomes; my current pipeline runs on a single workstation, which limits throughput. Scaling TOPOLOGIX to screen entire bacterial genomes, or to scan emerging variants in real time, requires distributed inference and efficient model serving. Google's data center expertise in large-scale machine learning operations is directly relevant to that scaling problem. The community dimension matters equally. I plan to deploy TOPOLOGIX as an open, queryable service for researchers in LMICs who lack access to structural biology facilities. A researcher in Lagos or Nairobi should be able to paste a mutation and a drug name and receive a resistance prediction without needing a cluster or a crystallography lab. My technical qualifications are documented across multiple independent projects. I have built and published three sole-authored preprints on the CCT model, a Bayesian-calibrated pharmacological framework for addiction, with all five pre-registered hypotheses confirmed. I have also conducted a pre-registered, powered replication study on topological descriptors for hERG cardiotoxicity prediction, which found that persistent homology features do not beat a plain descriptor baseline (AUROC 0.8426 versus 0.8782), a negative result that settles a comparison the literature had never actually run. That experience taught me to report what the data say, not what I hoped they would say. The ergofluids project, a Koopman-operator method for drug transport modeling, failed its first real-data validation gate; I reported that failure directly rather than reframing it. This is the rigor that a fellowship supporting community AI should reward. I am currently enrolled in the M.Sc. Digital Health program at the Hasso Plattner Institute and University of Potsdam, and I hold a B.Pharm from the University of Ibadan with a German equivalent grade of 1.9. I have endorsements from Kent Berridge at Michigan, Samuel Gershman at Harvard, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU. My ORCID is 0009-0001-9272-6735 and my code is public on GitHub at github.com/AmunRaPtah. TOPOLOGIX is the project that best matches this fellowship's mission: it is AI-driven, it addresses a pressing global health challenge, and it is designed for communities that need it most. RESEARCH STATEMENT TOPOLOGIX: SEQUENCE-ONLY PREDICTION OF DRUG-RESISTANCE MUTATIONS USING PROTEIN LANGUAGE MODEL EMBEDDINGS Problem. Drug resistance is a primary cause of treatment failure in infectious disease and oncology. Predicting whether a specific mutation confers resistance to a specific drug is therefore a high-value clinical question. The dominant computational tools, such as mCSM-lig, require a protein-ligand crystal structure as input. This is a severe limitation: for most proteins of clinical interest, no such structure exists. On the Platinum benchmark of 553 resistance mutations, structure-limited tools cover only about 18 percent of cases. The remaining 82 percent are simply unanswerable with current methods. A sequence-only approach removes this bottleneck entirely. Methods. TOPOLOGIX uses ESM-2 protein language model delta-embeddings to represent the mutation's effect on the protein sequence, combined with Morgan/ECFP drug fingerprints to represent the ligand, and a Random Forest classifier to predict resistance. The delta-embedding approach captures the change in the protein's learned representation induced by the amino acid substitution, which encodes biophysical and evolutionary information that structure-based features miss. The pipeline is implemented in Python using standard scientific computing libraries, with RDKit for fingerprint generation and scikit-learn for the classifier. The full codebase is public on GitHub. Results. On the Platinum benchmark of 553 mutations, TOPOLOGIX achieves an AUROC of 0.804 plus or minus 0.025. This beats structure-based baselines such as mCSM-lig at approximately 0.70. On SKEMPI 2.0, a binding-affinity benchmark, the AUROC is 0.634. Critically, TOPOLOGIX covers 100 percent of mutations in the benchmark, because it requires only sequence data. The performance gap between TOPOLOGIX and structure-based tools is not merely statistical; it is a coverage gap. For the majority of clinically relevant mutations, structure-based tools cannot even produce a prediction. Context and rigor. This project is the direct successor to two earlier studies that tested a different hypothesis. In the first, I tested whether bipartite persistent homology features of the protein-ligand interface could predict hERG cardiotoxicity. The pre-registered, powered replication found they could not: topological features achieved AUROC 0.8426 versus 0.8782 for a plain descriptor baseline. In the second, I applied the same topological constructs to drug-resistance prediction on the Platinum benchmark and found they carried almost no signal, with AUROCs of 0.425 and 0.485. These negative results ruled out interface geometry as the driver of resistance and motivated the sequence-representation approach that became TOPOLOGIX. I report these results because they are the evidence base for the design choice that made TOPOLOGIX work. Scaling plan. The current pipeline runs on a single workstation. The fellowship would support three scaling steps. First, distributed inference: ESM-2 embedding generation across whole bacterial genomes using Google-scale compute. Second, a public query interface: a web service where a researcher submits a protein sequence, a mutation, and a drug name, and receives a resistance prediction. Third, continuous retraining: as new resistance data are published, the model is updated and revalidated. The infrastructure for the third step, automated retraining pipelines with versioned datasets, is already part of my workflow from building four independent DuckDB-based ingest-to-analyze pipelines across life-sciences and other domains. Community impact. Antimicrobial resistance is a documented crisis in Nigeria and across sub-Saharan Africa. Routine clinical microbiology is scarce; resistance is often inferred from treatment failure. A sequence-only prediction tool lowers the barrier to actionable resistance information. It requires only a DNA sequencer, which is increasingly available, and an internet connection. The Google Data Center Community AI Fellowship's emphasis on community benefit in underserved regions aligns directly with this deployment model. The tool is open-source, the model weights will be public, and the query service will be free for non-commercial research use. Validation status. TOPOLOGIX has been validated on two independent benchmarks, Platinum and SKEMPI 2.0. It has not yet been validated on prospective clinical isolates from Nigerian hospitals. That is the next step, and it requires the kind of collaborative network and compute resources that a fellowship can provide. I am not claiming clinical deployment readiness; I am claiming a validated research tool with a clear path to deployment. SHORT ESSAY: COMMUNITY IMPACT The community I want to serve is Nigerian clinical microbiology and infectious disease research. In Nigeria, the standard of care for bacterial infection is empiric antibiotic therapy, because culture and susceptibility testing are often unavailable or unaffordable. When resistance emerges, it is detected late, after treatment failure, and the patient has often already progressed to severe disease. This is a structural problem, not a knowledge problem. The genomic data to detect resistance exist, but the computational tools to interpret them are locked behind structural biology infrastructure that most Nigerian institutions do not have. TOPOLOGIX changes the equation. It predicts drug-resistance mutations from sequence alone. A Nigerian researcher with a DNA sequencer and an internet connection can get a resistance prediction without a crystallography lab, without a high-performance computing cluster, and without a license fee. The model covers 100 percent of mutations in the Platinum benchmark, compared to 18 percent for structure-based tools. That is not an incremental improvement; it is a change in what is possible. The Google Data Center Community AI Fellowship is specifically designed for projects that use AI infrastructure for community benefit. My project fits that description precisely. The compute required for ESM-2 inference across whole genomes is substantial, and Google's data center expertise is directly relevant to making that inference fast and affordable. I am not asking for a grant to build a product; I am asking for support to build a public good. The model weights will be open, the code is already on GitHub, and the query service will be free for non-commercial research use. I also bring a track record of honest reporting. My ergofluids project, a Koopman-operator method for drug transport modeling, failed its first real-data validation gate; I reported that failure directly rather than reframing it. My hERG cardiotoxicity study found that topological features do not beat a plain descriptor baseline; I published that negative result. This matters because community AI projects fail when researchers overclaim. I will not overclaim. TOPOLOGIX is validated on two benchmarks, and I will report its performance on prospective Nigerian clinical isolates exactly as it is, whether that performance is good or bad. SHORT ESSAY: TECHNICAL EXPERTISE AND INFRASTRUCTURE ALIGNMENT The Google Data Center Community AI Fellowship emphasizes alignment with Google's infrastructure themes. My technical background maps to that emphasis in three concrete ways. First, large-scale ML inference. TOPOLOGIX uses ESM-2, a 650-million-parameter protein language model, to generate delta-embeddings for mutation representation. Generating embeddings for a single mutation is cheap; generating them for a whole bacterial genome, or for a panel of emerging variants, is not. The scaling problem is real, and it is the kind of problem Google's data center expertise solves. I have experience with HPC workflows through Nextflow and SLURM, and I have built production systems operations on Linux VPS with systemd, Caddy TLS, CI/CD, and automated backup and disaster recovery. I know how to deploy and maintain a service; the fellowship would let me do it at scale. Second, data engineering. I have built four independent DuckDB-based ingest-to-analyze pipelines across life-sciences, tech/AI/security, and social-science domains. These pipelines handle the full lifecycle from raw ingestion to structured analysis. TOPOLOGIX needs exactly this kind of infrastructure to move from a research script to a reliable service: versioned datasets, reproducible preprocessing, automated retraining, and a queryable interface. I also run self-hosted local LLM serving with llama.cpp and on-demand model swapping, which gives me hands-on experience with model serving infrastructure. Third, honest validation. My pre-registered hERG cardiotoxicity study found that topological features do not beat a plain descriptor baseline (AUROC 0.8426 versus 0.8782). My ergofluids project failed its first real-data validation gate, and I reported that failure directly. This matters for infrastructure projects because the cost of overclaiming is borne by the community that relies on the tool. I have a documented record of reporting what the data say. The alignment with Google's mission is not superficial. The fellowship supports AI projects that benefit communities; TOPOLOGIX is an AI project that predicts drug resistance from sequence alone, which is a direct benefit to communities in Nigeria and other LMICs that lack structural biology infrastructure. The compute requirements are real, the deployment model is open, and the validation status is honest. This is the project I am best positioned to execute, and it is the project that best matches this fellowship's stated criteria. CHECKLIST - [ ] Confirm exact application deadline and submission portal from the Watson.is programme page - [ ] Verify whether the fellowship requires a formal research proposal document in addition to the essays - [ ] Verify whether the fellowship requires letters of recommendation; if so, request from Kent Berridge, Samuel Gershman, Nathaniel Daw, or Marcelo Mattar - [ ] Confirm whether the fellowship requires a CV or resume upload; prepare a two-page academic CV - [ ] Confirm whether the fellowship requires a project budget or funding breakdown; prepare a one-page budget for TOPOLOGIX scaling (compute, data storage, travel for collaboration) - [ ] Confirm whether the fellowship requires proof of enrollment in the HPI/Potsdam M.Sc. Digital Health program; obtain enrollment certificate - [ ] Confirm whether the fellowship requires proof of Nigerian citizenship or LMIC status; prepare passport copy - [ ] Confirm whether the fellowship requires a GitHub portfolio review; ensure TOPOLOGIX repository is public, documented, and includes a README with benchmark results - [ ] Confirm whether the fellowship requires a video pitch or interview; prepare a 3-minute presentation of TOPOLOGIX results - [ ] Verify the ORCID record is up to date with all preprints and publications - [ ] Verify the zyco.org personal site links to the TOPOLOGIX repository and preprints - [ ] Confirm the Platinum benchmark and SKEMPI 2.0 results are reproducible from the public codebase; run the pipeline once more before submission - [ ] Draft a one-paragraph plain-language summary of TOPOLOGIX for non-specialist reviewers - [ ] Confirm the fellowship's eligibility criteria for independent researchers and pre-PhD applicants; if unclear, email the programme contact before submitting EDITOR NOTES - Eligibility risk: the programme page on Watson.is lists no explicit eligibility conditions, but Google fellowships often restrict to enrolled students or employees of specific institutions. Verify whether independent researchers and pre-PhD applicants are eligible before investing time in the full application. If restricted, consider applying through the HPI/Potsdam institutional channel instead. - The TOPOLOGIX results on Platinum (AUROC 0.804 plus or minus 0.025) and SKEMPI 2.0 (0.634) must be reproducible from the public GitHub repository. Run the pipeline end-to-end once more before submission and pin dependency versions in the README. A reviewer will check. - The SKEMPI 2.0 AUROC of 0.634 is notably lower than the Platinum result. Be prepared to explain this gap in an interview: SKEMPI measures binding affinity change, not resistance, so the task is different. Do not let a reviewer mistake this for a weakness. - The profile lists endorsements from Berridge, Gershman, Daw, and Mattar, but these are for the CCT and neuroscience work, not for TOPOLOGIX. If the fellowship requires letters, confirm whether any of these endorsers can speak to the TOPOLOGIX work specifically, or whether a letter from a computational biology collaborator is needed. - The employment history lists a National Product Manager role at Synthcare starting March 2026. Confirm this does not create a conflict of interest with the fellowship's independent researcher track, and confirm the fellowship does not require full-time dedication that would conflict with the M.Sc. enrollment.