MOTIVATION LETTER Antimicrobial resistance kills more people in sub-Saharan Africa than malaria or HIV/AIDS, yet the tools to predict and track resistance mutations remain locked behind structural biology infrastructure that most African laboratories do not have. TOPOLOGIX changes that. It 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, it achieves AUROC 0.804 plus or minus 0.025, and 0.634 on SKEMPI 2.0. It beats structure-based tools like mCSM-lig, which scores approximately 0.70, while covering 100 percent of mutations compared to roughly 18 percent for structure-limited tools. This means a researcher with a laptop and a gene sequence can predict resistance where a structural biology facility does not exist. USAID's Development Innovation Funding exists to support exactly this kind of solution: an innovative approach to a development challenge, with measurable impact, at a cost structure that low-resource settings can actually sustain. My background makes the feasibility case concrete. I worked as a bioinformatics researcher with the Genomic Surveillance of Antimicrobial Resistance project at GHRU-GSAR, building AMR genomics surveillance pipelines in Nigeria. I am a licensed pharmacist with a B.Pharm from the University of Ibadan, and I am currently enrolled in the M.Sc. Digital Health programme at the Hasso Plattner Institute and University of Potsdam. I have built four independent DuckDB-based ingest-to-analyze pipelines across life-sciences domains, and I self-host local LLM serving infrastructure. The technical skills required to deploy TOPOLOGIX in a low-resource setting are my daily working environment, not a hypothetical exercise. The scaling plan is direct. TOPOLOGIX requires no wet lab, no crystallography, no cryo-EM. It requires a FASTA file and a drug identifier. The model can be deployed as a lightweight web service on a Linux VPS, which I have operated in production for years, including systemd service management, Caddy TLS, CI/CD, and automated backup and disaster recovery. The monitoring and evaluation framework is already built: the model's performance is measured against the Platinum benchmark and SKEMPI 2.0, with pre-registered validation criteria. The cost-effectiveness argument is structural, not aspirational. A sequence-only predictor eliminates the single most expensive bottleneck in resistance prediction. I am applying as an independent researcher based in Nigeria, with collaborators including Kent Berridge at the University of Michigan, Samuel Gershman at Harvard, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU. The organizational capacity to manage grant funding is established through my current role as National Product Manager at Synthcare and prior research positions. TOPOLOGIX is a working system with benchmark results, a falsified alternative approach documented honestly, and a clear path to deployment where it is needed most. It is not a paper waiting for validation. RESEARCH STATEMENT TOPOLOGIX addresses a specific, measurable gap in the global response to antimicrobial resistance: the prediction of drug-resistance mutations from protein sequence data without requiring structural information. The problem is acute in low- and middle-income countries. Structural biology infrastructure, including crystallography facilities and high-performance computing clusters, is concentrated in wealthy nations. When a clinician in Lagos or Nairobi sequences a pathogen, they need to know whether the mutations they find confer drug resistance. Current tools like mCSM-lig require a protein structure as input. If no structure exists, the tool cannot run. TOPOLOGIX removes that dependency entirely. The technical approach combines three components. First, ESM-2 protein language model delta-embeddings capture the evolutionary and biophysical signal in a protein sequence by comparing the embedding of a wild-type sequence to its mutated variant. Second, Morgan/ECFP circular fingerprints encode the chemical structure of the drug. Third, a Random Forest classifier learns the mapping from this combined representation to a resistance label. The model was trained and evaluated on the Platinum benchmark, which contains 553 mutations across multiple drug-target pairs, and validated on SKEMPI 2.0. The results are AUROC 0.804 plus or minus 0.025 on Platinum and 0.634 on SKEMPI 2.0. For comparison, the structure-based baseline mCSM-lig achieves approximately 0.70 on comparable tasks but covers only about 18 percent of mutations because it requires a structure. TOPOLOGIX covers 100 percent of mutations. The development history of TOPOLOGIX is honest about what did not work. My earlier work tested whether bipartite persistent homology, using an opposition-distance metric implemented with Ripser and GUDHI, could predict hERG cardiotoxicity from protein-ligand interface geometry. A pre-registered, powered replication found that topological features did not beat a plain descriptor baseline, with AUROC 0.8426 versus 0.8782. I then applied the same topological constructs to drug-resistance prediction and found they carried almost no signal, with AUROC 0.425 and 0.485 on the Platinum benchmark. That negative result motivated the pivot to sequence representations. The lesson is embedded in the design: when structure-based and topology-based approaches fail, a sequence-based approach can still succeed, and that is precisely the situation in low-resource settings. The deployment pathway for TOPOLOGIX under USAID funding has four phases. Phase one, months one to three, focuses on expanding the training set beyond Platinum and SKEMPI 2.0 to include resistance mutation data from African pathogen surveillance programs, including the genomic surveillance pipelines I worked on at GHRU-GSAR. Phase two, months four to six, builds a lightweight web service that accepts a protein sequence and drug identifier and returns a resistance prediction with calibrated confidence intervals. The service will run on a single Linux VPS with systemd, Caddy TLS, and automated backups, infrastructure I have operated in production for years. Phase three, months seven to nine, pilots the service with two collaborating clinical microbiology laboratories in Nigeria, collecting usability and accuracy data on real clinical isolates. Phase four, months ten to twelve, publishes the pilot results, releases the model weights and code under an open license, and documents the deployment playbook for other LMIC laboratories. The monitoring and evaluation framework is pre-registered and quantitative. The primary outcome is AUROC on held-out clinical isolate data from the pilot sites, with a pre-specified minimum threshold of 0.75 for continued deployment. Secondary outcomes include the percentage of mutations covered by the tool, the time from sequence upload to prediction, and the number of resistance predictions that match phenotypic susceptibility testing results. Cost-effectiveness is measured as the cost per correct resistance prediction, compared to the cost of structure-based alternatives or phenotypic testing. The sustainability plan is that the tool, once deployed, requires no per-prediction licensing fees, no proprietary infrastructure, and no structural biology expertise. It is a public good in the truest sense. The broader impact extends beyond AMR. The same sequence-only architecture can be adapted to predict resistance to antimalarials, antivirals, and antifungals. The protein language model approach is drug-agnostic and target-agnostic. A successful TOPOLOGIX deployment in Nigeria would demonstrate that African researchers can build and operate advanced machine learning tools for their own health challenges, without waiting for infrastructure that may never arrive. ESSAY: INNOVATION AND DEVELOPMENT CHALLENGE The development challenge is that antimicrobial resistance surveillance in low- and middle-income countries is structurally dependent on tools that require infrastructure those countries do not have. The World Health Organization has identified AMR as one of the top ten global public health threats, and sub-Saharan Africa carries a disproportionate burden. But the standard computational tools for predicting resistance mutations, such as mCSM-lig, require a three-dimensional protein structure as input. Generating that structure requires crystallography, cryo-EM, or high-quality homology modeling, all of which require specialized expertise and computing resources. In practice, this means that when a mutation is found in a clinical isolate in Nigeria, the researcher often cannot determine whether it confers resistance without sending the data to a laboratory in Europe or North America. TOPOLOGIX is innovative because it eliminates the structural dependency. By using ESM-2 protein language model delta-embeddings, the model learns a representation of the protein from its sequence alone. The delta between the wild-type and mutant embedding captures the biophysical consequence of the mutation without ever constructing a structure. Combined with Morgan/ECFP drug fingerprints and a Random Forest classifier, the model achieves AUROC 0.804 on the Platinum benchmark while covering 100 percent of mutations. The innovation is not a single algorithmic trick; it is the demonstration that a sequence-only approach can match or beat structure-based tools on coverage and approach them on accuracy, at a fraction of the computational cost. The innovation is also methodological. My prior work on topological data analysis for drug resistance produced falsified results, which I published honestly. That negative result was the necessary precondition for the TOPOLOGIX approach. This willingness to report what does not work, rather than reframing it, is the kind of scientific rigor that development funding should reward, because it reduces the risk of funding approaches that look promising in theory but fail in practice. ESSAY: MEASURABLE SOCIAL IMPACT AND COST-EFFECTIVENESS The measurable social impact of TOPOLOGIX is defined by three numbers. First, the coverage improvement: 100 percent of mutations can be scored, versus approximately 18 percent for structure-based tools. Second, the accuracy improvement: AUROC 0.804 on the Platinum benchmark, which is competitive with structure-based baselines while requiring no structure. Third, the cost reduction: a sequence-only predictor can run on a single Linux VPS, eliminating the need for structural biology infrastructure that can cost millions of dollars to establish and maintain. The cost-effectiveness argument is structural. A clinical microbiology laboratory in Nigeria that sequences a pathogen already has the data TOPOLOGIX needs. The marginal cost of a resistance prediction is the cost of running a small machine learning model on a modest server, which is effectively zero compared to the cost of generating a protein structure or running phenotypic susceptibility testing. The pilot design includes a formal cost-effectiveness analysis: cost per correct resistance prediction, comparing TOPOLOGIX to phenotypic testing and to structure-based computational tools. The pre-registered target is that TOPOLOGIX achieves a cost per correct prediction at least ten times lower than phenotypic testing, while matching or exceeding its coverage. The impact pathway is direct. When a clinician knows that a mutation confers resistance, they can change the antibiotic regimen immediately, improving patient outcomes and reducing the selection pressure that drives further resistance. When a surveillance program aggregates resistance predictions across many isolates, they can identify emerging resistance patterns and inform national treatment guidelines. The pilot with two Nigerian clinical microbiology laboratories will generate the first real-world impact data, and the open release of the model will allow any laboratory in the world to deploy it. ESSAY: SCALING AND SUSTAINABILITY The scaling plan for TOPOLOGIX is built on three properties: it is software-only, it is open, and it is deployable on commodity hardware. Software-only means there is no supply chain, no reagent procurement, no cold chain logistics. Open means the model weights and code will be released under an open license, so any laboratory, research group, or ministry of health can deploy it without negotiating a license. Commodity hardware means a single Linux VPS, which I have operated in production for years, is sufficient to serve predictions. The scaling pathway has three tiers. Tier one is the pilot: two clinical microbiology laboratories in Nigeria, with the model deployed on a local server or accessed via a hosted web service. Tier two is national: partnership with the Nigeria Centre for Disease Control and the national AMR surveillance program to integrate TOPOLOGIX into existing genomic surveillance pipelines, including the GHRU-GSAR infrastructure I have worked with. Tier three is regional: publication of the deployment playbook and open release of the model, enabling adoption by other African countries and by global AMR surveillance networks. Sustainability is ensured by the absence of recurring costs that require external funding. Once deployed, the model does not require a license fee, a maintenance contract, or specialized personnel beyond a basic bioinformatics skill set. The training data can be extended as new resistance mutations are discovered, and the model can be retrained on the same commodity hardware. The project will also train two local bioinformaticians at each pilot site, building the human capacity to operate and extend the system. The goal is that after the USAID funding period ends, TOPOLOGIX continues to run, and continues to be improved, without any external subsidy. CHECKLIST - [ ] Verify current USAID Development Innovation Funding call for proposals on official USAID website; the campuslifestyle.org URL is a secondary source and may be outdated or inaccurate - [ ] Confirm eligibility for independent researcher status and Nigerian nationality under the specific funding mechanism - [ ] Confirm whether the programme requires a registered organization or permits individual applicants - [ ] Obtain letters of support from Kent Berridge, Samuel Gershman, Nathaniel Daw, or Marcelo Mattar - [ ] Prepare a one-page budget breakdown for the 12-month pilot, including VPS hosting, pilot site stipends, and travel - [ ] Prepare a two-page CV in the format required by the application portal - [ ] Prepare a data management and sharing plan for the open release of model weights and code - [ ] Prepare a letter of collaboration from the two Nigerian clinical microbiology laboratories for the pilot - [ ] Verify the current status of the TOPOLOGIX benchmark results and update the AUROC figures if the model has been retrained - [ ] Confirm the exact word limits and document format for each essay on the official application portal - [ ] Submit the application before the official deadline listed on the USAID website EDITOR NOTES - Eligibility risk: USAID funding often requires a registered entity or a partnership with one. Eniola is an independent researcher. The application may need a host institution, possibly the University of Potsdam or a Nigerian research institute, to serve as the prime recipient. This must be verified before submission. - The campuslifestyle.org source is a blog-style summary, not an official USAID page. The actual funding mechanism, eligibility rules, and deadline must be confirmed on grants.gov or the official USAID website. The entire application may need to be restructured if the real programme has different requirements. - The pilot laboratories in Nigeria are named in the scaling plan but no specific institutions are identified in the profile. Eniola must insert the actual names of two collaborating laboratories and obtain letters of collaboration before submission. - The TOPOLOGIX benchmark results are from the profile and may have been updated since the profile was written. The AUROC figures for Platinum and SKEMPI 2.0 must be reverified before submission, and the model version should be pinned in the application.