MOTIVATION LETTER The Platinum benchmark contains 553 experimentally verified drug-resistance mutations. Structure-based tools can score only about 18 percent of them because most lack a resolved protein-ligand complex. My TOPOLOGIX system scores all 553 from sequence alone, reaching an AUROC of 0.804 plus or minus 0.025, and it does so by combining ESM-2 protein language model delta-embeddings with Morgan fingerprints in a Random Forest classifier. This is the concrete result I bring to the NSF AI Datasets programme. The NSF mission for this call is to unlock value from existing datasets for AI-enabled discovery. My work does exactly that. The Platinum and SKEMPI 2.0 datasets are public, well-curated, and under-exploited. The published literature has applied structure-based predictors to them, but those predictors fail on the majority of mutations that lack structural coverage. TOPOLOGIX demonstrates that a sequence-representation approach extracts predictive signal these datasets already contain but that structure-limited methods cannot reach. On SKEMPI 2.0, TOPOLOGIX achieves 0.634 AUROC, and it covers every mutation in the benchmark, not just the structurally resolved subset. The intellectual merit of the proposal rests on a specific methodological claim: protein language model embeddings, when differenced between wild-type and mutant sequences, encode resistance-relevant biophysical signal that complements small-molecule fingerprints. This is a testable, reproducible claim. I have pre-registered the pipeline, and the code is public on GitHub. The broader impact is direct: drug-resistance prediction is a bottleneck in antimicrobial and anticancer therapy, and a sequence-only method removes the structural-data bottleneck entirely. I am an independent researcher based in Nigeria, currently enrolled in the M.Sc. Digital Health programme at the Hasso Plattner Institute and the University of Potsdam. I hold a B.Pharm from the University of Ibadan and a pharmacist license from the Pharmacists Council of Nigeria. My research record includes a pre-registered replication study on hERG cardiotoxicity topology that settled an open comparison in the literature, and a falsified hypothesis on interface topology for resistance prediction that directly motivated the TOPOLOGIX pivot. I report negative results as negative results. The grant would fund three deliverables: a cleaned, versioned release of the TOPOLOGIX training and evaluation pipeline as a reusable dataset artifact; a benchmark extension that adds resistance annotations from the CBM-AR and CARD databases to the existing Platinum framework; and a documented API so other researchers can run sequence-only resistance prediction on their own mutation lists. Each deliverable is scoped, testable, and aligned with the NSF emphasis on data sharing and community infrastructure. I am applying to the General track because the work spans AI methodology, biomedical data curation, and reproducibility infrastructure. The NSF criteria of intellectual merit, broader impacts, and data value map directly onto the three deliverables above. I request consideration as an early-career, independent researcher from an underrepresented geographic region. RESEARCH STATEMENT TOPOLOGIX: Sequence-Only Prediction of Drug-Resistance Mutations via Protein Language Model Delta-Embeddings Problem and Motivation Drug resistance arises from mutations that alter how proteins bind small molecules. Predicting which mutations confer resistance is a central task in antimicrobial stewardship, anticancer therapy, and drug development. The standard approach requires a resolved protein-ligand complex structure, which exists for only a minority of clinically observed mutations. The Platinum benchmark, a widely used dataset of 553 resistance mutations across diverse protein targets, has structural coverage for roughly 18 percent of its entries. Any method that depends on structure therefore discards the majority of available evidence. The scientific question is whether sequence information alone, combined with drug chemical structure, can predict resistance with accuracy comparable to or better than structure-based methods. The published literature has not answered this question directly. Structure-based tools such as mCSM-lig report AUROC values around 0.70 on the subsets they can score. No published method has reported full-coverage performance on Platinum. This gap is the target of TOPOLOGIX. Methods TOPOLOGIX uses three input modalities. First, ESM-2 protein language model embeddings are computed for the wild-type and mutant protein sequences. The delta-embedding, the vector difference between the two, captures the mutational perturbation in a learned biophysical representation space. Second, Morgan/ECFP circular fingerprints encode the drug molecule's topology and functional groups. Third, a Random Forest classifier learns the mapping from the concatenated delta-embedding and fingerprint vector to a binary resistance label. The pipeline is implemented in Python. Embedding computation uses the ESM-2 model via the HuggingFace transformers library. Fingerprints are generated with RDKit. The classifier is scikit-learn RandomForest with hyperparameter tuning by cross-validation. The full pipeline is version-controlled on GitHub and runs on a standard Linux workstation with a single GPU. No proprietary software or data is required. Results On the Platinum benchmark, TOPOLOGIX achieves an AUROC of 0.804 plus or minus 0.025 across cross-validation folds. This exceeds the reported performance of mCSM-lig, the leading structure-based tool, which scores only the structurally resolved subset. On SKEMPI 2.0, a binding-affinity mutation dataset used as a transfer benchmark, TOPOLOGIX achieves 0.634 AUROC. The performance gap between the two benchmarks is informative: Platinum contains curated resistance labels, while SKEMPI 2.0 contains affinity changes, a related but distinct quantity. The result is a falsification of the hypothesis that interface geometry is the dominant driver of resistance predictability. My prior work tested bipartite persistent homology features, opposition-distance metrics, and element-specific topological descriptors on the Platinum benchmark. Those features produced AUROC values of 0.425 and 0.485, at or below chance. The topological approach failed. The sequence-representation approach succeeds. This negative-to-positive progression is the core intellectual contribution: it rules out one class of features and validates another, using the same benchmark and the same evaluation protocol. Data Value and Reproducibility The TOPOLOGIX contribution to dataset value is threefold. First, it demonstrates that existing public datasets contain predictive signal that current methods cannot access. Second, it provides a reproducible pipeline that other researchers can run on their own mutation lists without structural data. Third, it establishes a benchmark protocol for sequence-only resistance prediction, including train-test split conventions, evaluation metrics, and baseline comparisons. The proposed work under this grant has three deliverables. Deliverable one is a cleaned, versioned release of the TOPOLOGIX dataset artifact, including embeddings, fingerprints, labels, and the exact train-test split used in the reported results. Deliverable two is an extension of the Platinum benchmark with resistance annotations from the CBM-AR and CARD databases, adding approximately 200 to 400 additional mutations with sequence-only coverage. Deliverable three is a documented API, deployed as a containerized service, that accepts a protein sequence and a drug SMILES string and returns a resistance prediction with calibrated uncertainty. Feasibility and Qualifications I am an independent computational researcher with a B.Pharm from the University of Ibadan and current enrollment in the M.Sc. Digital Health programme at the Hasso Plattner Institute and the University of Potsdam. My technical skills include Python, PyMC for Bayesian calibration, RDKit, ESM-2, and HPC workflow management with Nextflow and SLURM. I have built and operated four independent data pipelines from ingestion to analysis, and I self-host local LLM serving infrastructure. The TOPOLOGIX pipeline is already implemented and validated. The proposed work is an extension of an existing, working system, not a new speculative direction. Broader Impacts Drug resistance is a global health problem with disproportionate burden in low- and middle-income countries, including Nigeria. A sequence-only prediction method lowers the barrier to resistance surveillance in settings without structural biology infrastructure. The open-source release of the pipeline and benchmark extension directly serves the NSF goal of advancing scientific infrastructure. As a Nigerian researcher, I bring a perspective on resistance data needs that is underrepresented in the AI-for-science community. EDITOR NOTES - Chosen research line: TOPOLOGIX. This is the only active project that directly uses existing public datasets (Platinum, SKEMPI 2.0) to build an AI model, which is the exact mission of the NSF AI Datasets call. The CCT model, neurocascade, and ergofluids are either not dataset-centric or are behind validation gates; psyche-twin is a personal tool, not a scientific dataset contribution. The hERG topology study is a negative result and is used here only as context for the TOPOLOGIX pivot, not as a current claim. - Eligibility risk: The call is a U.S. NSF grant. The applicant is a Nigerian citizen, independent researcher, and incoming M.Sc. student at a German university. NSF grants typically require U.S. institutional affiliation or a U.S. co-PI. This must be verified before submission. If ineligible, the same materials can be adapted for the NSF IRES or IRAD tracks, or for the applicant's future HPI affiliation. - Facts to verify: The AUROC values for mCSM-lig (approximately 0.70) and the Platinum benchmark structural coverage (approximately 18 percent) are cited from memory of the literature. The applicant must confirm these numbers against the original publications before submission. The CBM-AR and CARD database names and their annotation formats must be checked for current availability and licensing. - Gaps to fill: The proposal mentions a calibrated uncertainty output in the API deliverable. The current TOPOLOGIX Random Forest does not natively produce calibrated probabilities. The applicant must either add a calibration step (e.g., Platt scaling or isotonic regression) to the work plan or remove the calibration claim. The applicant should also specify the exact GPU hardware and compute budget required for the ESM-2 embedding step, as this affects feasibility for reviewers. - Tone and framing: The letter and statement deliberately lead with the falsified topology result as a methodological pivot, not as a failure. This is consistent with the NSF emphasis on rigor and reproducibility. The applicant should be prepared to discuss the hERG cardiotoxicity replication study in the same register if asked, as it is the other major negative result in the portfolio.