MOTIVATION LETTER The gap between what sequence databases contain and what they mean is widening faster than manual curation can close it. The Platinum benchmark holds 553 drug-resistance mutations; structure-based tools like mCSM-lig can only score roughly 18 percent of them because the remaining 82 percent lack experimentally resolved protein structures. My TOPOLOGIX system closes that gap by predicting drug-resistance mutations from sequence alone, using ESM-2 protein-language-model delta-embeddings combined with Morgan/ECFP drug fingerprints and a Random Forest classifier. It achieves an AUROC of 0.804 plus or minus 0.025 on Platinum, outperforming structure-based baselines while covering every mutation in the benchmark. This is exactly the kind of scalable curation the National Institutes of Health's Computational Approaches to Curation at Scale programme exists to fund. The NIH R01 mechanism demands strong preliminary data, clear hypotheses, and high potential impact. TOPOLOGIX delivers all three. The preliminary data comes from my pre-registered replication work on hERG cardiotoxicity topology, where I tested whether bipartite persistent homology predicts toxicity from protein-ligand interface geometry. The powered replication found topological features do not beat a plain descriptor baseline (AUROC 0.8426 versus 0.8782), a comparison the published literature had never actually run. That negative result taught me which computational shortcuts fail and motivated the pivot to sequence-representation methods. The follow-up study on interface-topology-for-resistance confirmed the same lesson: topological constructs carry almost no signal for resistance prediction (AUROC 0.425 and 0.485 on Platinum), ruling out interface geometry as the driver. TOPOLOGIX is the positive result that emerged from those falsifications. The proposed project extends TOPOLOGIX from drug-resistance mutations to a broader class of biomedical curation assets: variant effect prediction across protein families, automated literature mining for mutation-effect annotations, and continuous retraining as new sequence data enters public repositories. The method's core advantage is that it requires no structural data, no expensive molecular dynamics simulations, and no manual feature engineering. It ingests raw sequence and drug fingerprint data and outputs calibrated resistance predictions. That makes it deployable at database scale, which is precisely the curation bottleneck the NIH programme targets. My training supports this work. I hold a B.Pharm from the University of Ibadan with a German-equivalent grade of 1.9, I am enrolled in the M.Sc. Digital Health programme at the Hasso Plattner Institute and University of Potsdam, and I have built four independent DuckDB-based ingest-to-analyze pipelines across life-sciences, tech, and social-science domains. I am an independent researcher, which means the NIH environment criterion requires careful handling; I address that in the research statement with a concrete computing and collaboration plan. The problem is important, the method is novel, the preliminary data is strong, and the scalability is demonstrated. I am applying to the NIH to build the curation infrastructure that sequence databases urgently need. RESEARCH STATEMENT Project title: Scalable Curation of Drug-Resistance Mutation Data Using Protein Language Models Specific aims Aim 1: Extend TOPOLOGIX from the Platinum benchmark to a multi-dataset variant effect prediction framework. The current system predicts drug-resistance mutations from sequence alone using ESM-2 delta-embeddings, Morgan/ECFP fingerprints, and a Random Forest classifier. On Platinum (553 mutations) it achieves AUROC 0.804 plus or minus 0.025; on SKEMPI 2.0 it achieves 0.634. The extension will add training and evaluation on ClinVar, CIViC, and OncoKB variant sets, with standardized train-test splits and calibration curves reported for every dataset. Aim 2: Build an automated literature-mining pipeline that extracts mutation-effect annotations from PubMed abstracts and full-text open-access articles, using the same ESM-2 embedding space to link textual mentions to sequence positions. The pipeline will use a named-entity recognition layer trained on the CIViC and OncoKB corpora, followed by a relation-extraction step that binds mutations to drugs and resistance outcomes. Output will be structured as machine-readable annotation records with confidence scores. Aim 3: Deploy a continuous retraining and validation loop that ingests new sequence and annotation data monthly, retrains the classifier, and publishes updated performance metrics on a public dashboard. The loop will use the pre-registered gated validation protocol I developed for the ergofluids project: each new data batch must pass a pre-specified performance threshold on held-out sets before it is merged into the training corpus. This protocol prevents silent drift and keeps every published number reproducible. Significance Drug resistance is a moving target. Pathogens and tumors mutate continuously, and each new mutation changes the clinical meaning of a sequence record. Manual curation cannot keep pace: the 18 percent structural coverage problem on Platinum is one instance of a general failure mode. The NIH mission to make biomedical research assets FAIR, reproducible, and discoverable depends on automated methods that can annotate sequences at the rate they are deposited. TOPOLOGIX addresses that need directly. It also addresses reproducibility: every component, from the ESM-2 embedding extraction to the Random Forest training and evaluation, is versioned and runnable from a single command-line interface. Innovation The innovation is the combination of protein language model embeddings with drug fingerprints for resistance prediction, and the demonstration that this combination beats structure-based tools while covering 100 percent of mutations. No published system achieves both coverage and accuracy on the Platinum benchmark. The extension to literature mining is also novel: linking textual mutation mentions to sequence positions through the same embedding space creates a unified representation that supports both prediction and retrieval. Approach Data sources: Platinum benchmark (553 mutations), SKEMPI 2.0, ClinVar, CIViC, OncoKB, PubMed, and PubMed Central open-access subset. All are public and freely accessible. Methods: ESM-2 embeddings extracted per protein sequence, delta-embeddings computed between wild-type and mutant sequences, Morgan fingerprints (radius 2, 2048 bits) for drug molecules, concatenated feature vectors fed to a Random Forest classifier with cross-validated hyperparameter tuning. Literature mining uses a transformer-based NER model fine-tuned on CIViC and OncoKB annotations, with relation extraction via a span-based classifier. All code is Python, using scipy, numpy, pandas, RDKit, and PyTorch. Validation: Pre-registered evaluation protocols for every dataset. Primary metrics are AUROC and coverage. Calibration is assessed with reliability diagrams. The gated retraining loop uses the same protocol I applied in the ergofluids project, where a real-data gate that failed its primary criterion was reported directly rather than reframed. Feasibility and timeline Month 1-3: Assemble multi-dataset training and evaluation framework. Month 4-6: Train and evaluate extended TOPOLOGIX on ClinVar, CIViC, and OncoKB. Month 7-9: Build and validate the literature-mining pipeline. Month 10-12: Deploy the retraining loop and public dashboard. Month 12-18: Prepare publications and release all code and trained models under an open license. Environment I am an independent researcher. Computing resources are self-hosted: a Linux VPS with systemd, Caddy TLS, CI/CD, and automated backup and disaster recovery. Model training runs on CPU and modest GPU resources; ESM-2 embedding extraction for the datasets in this proposal is feasible within that constraint. I have secured arXiv endorsement from Samuel Gershman at Harvard and maintain collaborations with Kent Berridge at Michigan, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU. For this project, I will establish a data-sharing agreement with the maintainers of the Platinum benchmark and the CIViC database to ensure access to the latest curated annotations. I am also enrolled in the M.Sc. Digital Health programme at the Hasso Plattner Institute and University of Potsdam, which provides institutional affiliation and access to academic computing infrastructure. Budget narrative The requested funds support computing resources (GPU rental for ESM-2 embedding extraction and model training), open-access publication fees, and travel to one conference for presenting results. No personnel costs are requested; all work is performed by the applicant. The budget is modest relative to the NIH R01 scale because the method is computationally efficient and the data sources are public. CHECKLIST - [ ] Confirm eligibility for NIH R01 as an independent researcher without a sponsoring institution; contact NIH program officer listed on grants.gov detail page 356888 before drafting final submission - [ ] Verify that the M.Sc. Digital Health enrollment at HPI/Potsdam provides sufficient institutional affiliation for the R01 environment criterion; if not, secure a letter of support from a faculty member at HPI or a collaborating institution - [ ] Obtain letters of support from Kent Berridge, Samuel Gershman, Nathaniel Daw, and Marcelo Mattar confirming the nature of the collaboration and their willingness to be named - [ ] Prepare the specific aims page (one page, 30 lines maximum, NIH format) - [ ] Prepare the research strategy section (12 pages maximum, NIH format) using the research statement above as the core - [ ] Prepare the bibliography and references cited section - [ ] Prepare the biosketch using the NIH format, including ORCID 0009-0001-9272-6735 and GitHub github.com/AmunRaPtah - [ ] Prepare the budget and budget justification using the NIH SF424 R&R form - [ ] Prepare the facilities and other resources section, documenting self-hosted computing infrastructure and HPI affiliation - [ ] Prepare the authentication of key biological and/or chemical resources section if required by the NIH R01 instructions - [ ] Prepare the project summary and project narrative (public abstract, one page each) - [ ] Verify all preprints and publications are listed with DOIs or Zenodo/OSF links: CCT model preprints (three sole-authored), Alcohol co-authored paper, hERG topology study, interface-topology-for-resistance study, TOPOLOGIX results - [ ] Confirm the deadline of 04/15/2027 and submit through grants.gov at least 48 hours before the deadline - [ ] Register or verify eRA Commons account and link to the grants.gov submission EDITOR NOTES - Eligibility risk: The NIH R01 is typically awarded to institutions, not individuals. An independent researcher without a formal institutional affiliation may face rejection at the eligibility stage. The application must either secure a sponsoring institution (HPI/Potsdam is the most plausible) or contact the NIH program officer early to confirm whether independent applicants are accepted. This is the single largest risk to the submission. - The TOPOLOGIX results (AUROC 0.804 on Platinum, 0.634 on SKEMPI 2.0) are reported from the applicant profile but the underlying preprint or publication link is not listed. Verify that these numbers are published or available as a preprint with a DOI before including them in the final application. NIH reviewers will check. - The budget narrative states no personnel costs are requested. If the applicant intends to pay themselves a salary from the grant, this must be revised. Independent researchers often need salary support; the current draft assumes all work is voluntary, which may be unrealistic for a multi-year project. - The environment section claims access to HPI computing infrastructure through the M.Sc. program. This needs verification: enrolled students may not have the same resource access as faculty or research staff. Obtain a written statement from HPI confirming available compute resources before submission. - The literature-mining pipeline in Aim 2 is described as feasible but no preliminary data exists for the NER or relation-extraction components. Consider adding a small pilot study to the preliminary data section, or reframe Aim 2 as exploratory with clear success criteria, to avoid reviewer criticism about overreach.