Draft — NLnet Foundation / NGI Zero Commons Fund

MOTIVATION LETTER

The NGI Zero Commons Fund exists to build open, trustworthy infrastructure that reduces dependence on centralized technology providers. Antimicrobial resistance kills an estimated 1.27 million people per year, with the highest burden in low- and middle-income countries, yet the tools for predicting drug resistance remain locked behind structural biology expertise and proprietary databases. TOPOLOGIX addresses this gap directly: an open-source, sequence-based drug-resistance prediction tool that requires only a protein sequence and a drug fingerprint to make a prediction, covering 100 percent of mutations in the Platinum benchmark compared to roughly 18 percent for structure-limited tools like mCSM-lig.

TOPOLOGIX achieves an AUROC of 0.804 plus or minus 0.025 on the Platinum benchmark of 553 mutations and 0.634 on SKEMPI 2.0, outperforming structure-based baselines while requiring no crystal structure, no homology model, and no expert structural biology knowledge. The system uses ESM-2 protein language model delta-embeddings combined with Morgan/ECFP drug fingerprints fed into a Random Forest classifier. This design choice is deliberate: it runs on standard Python libraries, trains in minutes on a laptop, and can be deployed by a public health laboratory in Lagos or Nairobi without high-performance computing infrastructure.

The project aligns with NGI Zero's selection criteria on multiple explicit grounds. Open source: the code will be released under an OSI-approved license, with the training pipeline, benchmark scripts, and model weights published alongside the paper. Public interest: drug resistance prediction is a global health commons problem, not a commercial product opportunity. Independence and trust: the tool reduces reliance on commercial structural biology suites and proprietary mutation databases, giving LMIC laboratories autonomous capacity for resistance surveillance. Technical merit: the pre-registered replication study that preceded TOPOLOGIX demonstrated methodological rigor by reporting a negative result, that topological features of protein-ligand interfaces do not predict hERG cardiotoxicity beyond plain descriptors, settling a comparison the literature had never actually run.

The requested budget of 50,000 euros covers one year of development: 30,000 euros for the applicant's time, 10,000 euros for cloud compute for large-scale ESM-2 embedding generation, and 10,000 euros for community building, documentation, and a validation workshop with Nigerian public health partners. The deliverable is a production-ready, documented, benchmarked open-source tool plus a published validation study.

TOPOLOGIX is a public health infrastructure project that uses machine learning methods the applicant has developed across multiple domains. The NGI Zero Commons Fund is the correct home for this work because it funds exactly this category: small, independent, technically rigorous projects that make critical infrastructure accessible to those who need it most.

RESEARCH STATEMENT

TOPOLOGIX addresses a specific, measurable problem: predicting whether a given mutation in a drug target confers resistance to a given drug, using only sequence information. The current state of the art requires either a crystal structure of the mutant protein-ligand complex, which exists for a tiny fraction of clinically relevant mutations, or expensive and slow experimental assays. This creates a structural blind spot. Public health laboratories in LMICs, where resistance surveillance matters most, rarely have structural biology capacity. They have sequencing data. TOPOLOGIX is designed to work with what they have.

The technical approach builds on a falsified hypothesis. Between 2024 and 2025, the applicant tested whether bipartite persistent homology of protein-ligand interface geometry could predict hERG cardiotoxicity. A pre-registered, powered replication found that topological features do not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782. A follow-up study applied the same topological constructs to drug-resistance prediction and found they carry almost no signal, AUROC 0.425 and 0.485 on the Platinum benchmark. These negative results ruled out interface geometry as the driver of resistance and motivated a sequence-representation approach instead. TOPOLOGIX is that pivot, and the negative results are published as preprints and in review, not hidden.

The current TOPOLOGIX pipeline works as follows. A protein sequence is passed through ESM-2 to generate per-residue embeddings. The delta between wild-type and mutant embeddings captures the mutational perturbation. A drug is represented by its Morgan/ECFP fingerprint. The concatenated representation feeds a Random Forest classifier. On the Platinum benchmark of 553 mutations, the system achieves AUROC 0.804 plus or minus 0.025. On SKEMPI 2.0, it achieves 0.634. It beats mCSM-lig, a widely used structure-based tool, which scores approximately 0.70 on the same tasks, while covering 100 percent of mutations versus roughly 18 percent for structure-limited tools.

The development plan for the NGI Zero grant period is concrete. Months one to three: freeze the benchmark suite, reproduce all published results from a clean environment, and containerize the pipeline for reproducible deployment. Months four to six: expand the training set beyond Platinum and SKEMPI 2.0 by curating additional mutation-resistance pairs from published literature, targeting a 30 percent increase in training data. Months seven to nine: build the command-line interface and a simple web API, with documentation aimed at a bench scientist who has never used machine learning. Months ten to twelve: run a validation workshop with a Nigerian public health laboratory partner, collect feedback, and publish the full validation study.

The innovation claim is specific. No existing open-source tool predicts drug resistance from sequence alone with this coverage and accuracy. The novelty is not a new neural architecture; it is the demonstration that a protein language model plus a drug fingerprint, combined in a simple classifier, outperforms structure-based methods while removing the structural data requirement entirely. That is a practical, deployable advance.

The public interest case is direct. Antimicrobial resistance is a global commons problem. The tools to address it should be commons infrastructure. TOPOLOGIX will be released under an OSI-approved license, with all code, benchmarks, and model weights public. The project reduces dependence on commercial structural biology software and proprietary mutation databases, giving independent researchers and LMIC public health laboratories the capacity to run resistance prediction locally, on their own data, without sending their sequences to a foreign cloud service.

The applicant is an independent researcher with a track record of delivering rigorous, pre-registered computational work across multiple domains, including a 1,847-record literature screen for Bayesian prior elicitation, a 14-parameter MCMC calibration of a pharmacological model, and four independent data pipelines built for corpus analysis. The TOPOLOGIX codebase is maintained on GitHub under the applicant's account, with the negative-result predecessor studies archived on OSF and Zenodo. The work is ready for the next stage: production hardening, community adoption, and deployment in the settings where it matters most.

ESSAY: OPEN SOURCE AND TRUST

The trust case for TOPOLOGIX rests on three concrete properties: the code is auditable, the results are reproducible, and the negative results are public. The full pipeline, from raw sequence to prediction, will be released under an OSI-approved license. Every benchmark script, every preprocessing step, and every model weight will be in the repository. A reviewer can run the entire evaluation from a clean environment and verify the reported AUROC values. This is the standard the applicant applied to the predecessor studies, which were pre-registered on OSF before data analysis, and it is the standard TOPOLOGIX will maintain.

Reproducibility is enforced by containerization. The pipeline will ship as a Docker image with pinned dependency versions, so a laboratory in Nigeria runs exactly the same software as the developer. The benchmark suite is frozen at the start of the grant period, and any changes to the pipeline must pass the full suite before release. This prevents the common failure mode where a published tool cannot be re-run by anyone other than the author.

The negative results are part of the trust story. The applicant published the finding that topological features do not predict hERG cardiotoxicity beyond plain descriptors, and the finding that interface topology carries no signal for drug resistance. These results are archived on OSF and Zenodo and are in review at peer-reviewed journals. A project that reports its failures as clearly as its successes is a project that can be trusted with public health infrastructure.

The independence angle is structural. TOPOLOGIX requires no commercial structural biology suite, no proprietary database, and no cloud dependency. It runs on standard Python libraries. A public health laboratory can deploy it on a laptop. This is digital sovereignty in practice: the capacity to make predictions about local resistance patterns without routing data through a foreign commercial service.

ESSAY: TECHNICAL FEASIBILITY AND PLAN

The applicant has already built and evaluated TOPOLOGIX. The core classifier exists, the benchmarks are defined, and the results are reproducible from the public repository. The grant period is therefore not a research proposal; it is a hardening and deployment plan with clear milestones.

Month one to three: freeze the benchmark suite, reproduce all published results from a clean environment, containerize the pipeline. Success criterion: a fresh clone of the repository plus the Docker image reproduces the reported AUROC values within 0.01.

Month four to six: expand the training data by curating additional mutation-resistance pairs from published literature, targeting a 30 percent increase. Success criterion: the expanded training set is publicly released with provenance for every entry.

Month seven to nine: build the command-line interface and a simple web API, with documentation for bench scientists. Success criterion: a user with no machine learning background can run a prediction from a FASTA file and a SMILES string in under ten minutes.

Month ten to twelve: run a validation workshop with a Nigerian public health laboratory partner, collect structured feedback, publish the full validation study. Success criterion: the workshop produces a written report and at least one external user runs the tool independently.

The budget request is 50,000 euros. The applicant's time is 30,000 euros for twelve months of part-time development. Cloud compute for large-scale ESM-2 embedding generation is 10,000 euros. Community building, documentation, and the validation workshop are 10,000 euros. This is proportionate to the work and does not include any equipment, office space, or institutional overhead.

The risk register is honest. The main technical risk is that the expanded training set does not improve performance, which would mean the current benchmark results are close to the ceiling for this method class. The mitigation is that the tool is already useful at current performance, and the grant period would still deliver a production-ready, documented, benchmarked release. The main non-technical risk is community adoption, which the validation workshop directly addresses by building a relationship with a real deployment partner from month one.

CHECKLIST

- [ ] Verify current NLnet Foundation / NGI Zero Commons Fund application form and submission portal access
- [ ] Confirm the 50,000 euro budget request is within the fund's per-project cap for the current round
- [ ] Prepare TOPOLOGIX GitHub repository with README, license (OSI-approved, recommend Apache 2.0 or MIT), and reproduction instructions
- [ ] Archive the current TOPOLOGIX codebase and benchmark results on Zenodo with a DOI
- [ ] Confirm the two negative-result predecessor studies (hERG topology, resistance topology) are publicly archived on OSF/Zenodo with DOIs
- [ ] Draft a one-page project timeline document matching the month-by-month plan in the research statement
- [ ] Draft a one-page budget justification table breaking down the 50,000 euro request
- [ ] Prepare a short bio sketch including B.Pharm, University of Ibadan, M.Sc. Digital Health enrollment at HPI/Potsdam, and ORCID
- [ ] Identify and contact the Nigerian public health laboratory partner for the validation workshop; obtain a letter of intent
- [ ] Confirm whether the fund requires a data management plan; if so, draft one covering the expanded training set release
- [ ] Verify the fund's open-source license requirements and confirm the chosen license is compliant
- [ ] Check whether the fund requires a fiscal sponsor or bank account in a specific jurisdiction; confirm the applicant's independent researcher status is eligible
- [ ] Submit the application through the NLnet portal with all attachments in the required formats

EDITOR NOTES

- Eligibility risk: the applicant is enrolled in an M.Sc. program starting Winter Semester 2026/27 and employed as National Product Manager at Synthcare from March 2026. The fund should be checked for any restriction on funding individuals who are concurrently employed or enrolled; the application should clarify that TOPOLOGIX development is independent of both the M.Sc. coursework and the Synthcare role.
- The Nigerian public health laboratory partner is not named in the profile. The applicant must insert a real partner name and a letter of intent before submission; this is a gap that cannot be filled from the profile.
- The budget of 50,000 euros is an assumption. The fund's per-project cap for the current round must be verified, and the budget adjusted to match the actual cap and the fund's stated expectations for solo researcher grants.
- The profile lists multiple research lines. This application deliberately features TOPOLOGIX only, because it is the best fit for NGI Zero's open infrastructure and public interest criteria. The CCT model, neurocascade, ergofluids, and psyche-twin are not mentioned in the application materials; do not add them, as they would dilute the thematic match.
- The applicant's employment at Synthcare (a commercial entity) and the ergofluids project's explicit "no IP/product claims" stance should be checked for any conflict with the fund's independence criteria. The application should not imply that Synthcare endorses or is involved in TOPOLOGIX.

MOTIVATION LETTER

The NGI Zero Commons Fund exists to build open, trustworthy infrastructure that reduces dependence on centralized technology providers. Antimicrobial resistance kills an estimated 1.27 million people per year, with the highest burden in low- and middle-income countries, yet the tools for predicting drug resistance remain locked behind structural biology expertise and proprietary databases. TOPOLOGIX addresses this gap directly: an open-source, sequence-based drug-resistance prediction tool that requires only a protein sequence and a drug fingerprint to make a prediction, covering 100 percent of mutations in the Platinum benchmark compared to roughly 18 percent for structure-limited tools like mCSM-lig.

TOPOLOGIX achieves an AUROC of 0.804 plus or minus 0.025 on the Platinum benchmark of 553 mutations and 0.634 on SKEMPI 2.0, outperforming structure-based baselines while requiring no crystal structure, no homology model, and no expert structural biology knowledge. The system uses ESM-2 protein language model delta-embeddings combined with Morgan/ECFP drug fingerprints fed into a Random Forest classifier. This design choice is deliberate: it runs on standard Python libraries, trains in minutes on a laptop, and can be deployed by a public health laboratory in Lagos or Nairobi without high-performance computing infrastructure.

The project aligns with NGI Zero's selection criteria on multiple explicit grounds. Open source: the code will be released under an OSI-approved license, with the training pipeline, benchmark scripts, and model weights published alongside the paper. Public interest: drug resistance prediction is a global health commons problem, not a commercial product opportunity. Independence and trust: the tool reduces reliance on commercial structural biology suites and proprietary mutation databases, giving LMIC laboratories autonomous capacity for resistance surveillance. Technical merit: the pre-registered replication study that preceded TOPOLOGIX demonstrated methodological rigor by reporting a negative result, that topological features of protein-ligand interfaces do not predict hERG cardiotoxicity beyond plain descriptors, settling a comparison the literature had never actually run.

The requested budget of 50,000 euros covers one year of development: 30,000 euros for the applicant's time, 10,000 euros for cloud compute for large-scale ESM-2 embedding generation, and 10,000 euros for community building, documentation, and a validation workshop with Nigerian public health partners. The deliverable is a production-ready, documented, benchmarked open-source tool plus a published validation study.

TOPOLOGIX is a public health infrastructure project that uses machine learning methods the applicant has developed across multiple domains. The NGI Zero Commons Fund is the correct home for this work because it funds exactly this category: small, independent, technically rigorous projects that make critical infrastructure accessible to those who need it most.

RESEARCH STATEMENT

TOPOLOGIX addresses a specific, measurable problem: predicting whether a given mutation in a drug target confers resistance to a given drug, using only sequence information. The current state of the art requires either a crystal structure of the mutant protein-ligand complex, which exists for a tiny fraction of clinically relevant mutations, or expensive and slow experimental assays. This creates a structural blind spot. Public health laboratories in LMICs, where resistance surveillance matters most, rarely have structural biology capacity. They have sequencing data. TOPOLOGIX is designed to work with what they have.

The technical approach builds on a falsified hypothesis. Between 2024 and 2025, the applicant tested whether bipartite persistent homology of protein-ligand interface geometry could predict hERG cardiotoxicity. A pre-registered, powered replication found that topological features do not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782. A follow-up study applied the same topological constructs to drug-resistance prediction and found they carry almost no signal, AUROC 0.425 and 0.485 on the Platinum benchmark. These negative results ruled out interface geometry as the driver of resistance and motivated a sequence-representation approach instead. TOPOLOGIX is that pivot, and the negative results are published as preprints and in review, not hidden.

The current TOPOLOGIX pipeline works as follows. A protein sequence is passed through ESM-2 to generate per-residue embeddings. The delta between wild-type and mutant embeddings captures the mutational perturbation. A drug is represented by its Morgan/ECFP fingerprint. The concatenated representation feeds a Random Forest classifier. On the Platinum benchmark of 553 mutations, the system achieves AUROC 0.804 plus or minus 0.025. On SKEMPI 2.0, it achieves 0.634. It beats mCSM-lig, a widely used structure-based tool, which scores approximately 0.70 on the same tasks, while covering 100 percent of mutations versus roughly 18 percent for structure-limited tools.

The development plan for the NGI Zero grant period is concrete. Months one to three: freeze the benchmark suite, reproduce all published results from a clean environment, and containerize the pipeline for reproducible deployment. Months four to six: expand the training set beyond Platinum and SKEMPI 2.0 by curating additional mutation-resistance pairs from published literature, targeting a 30 percent increase in training data. Months seven to nine: build the command-line interface and a simple web API, with documentation aimed at a bench scientist who has never used machine learning. Months ten to twelve: run a validation workshop with a Nigerian public health laboratory partner, collect feedback, and publish the full validation study.

The innovation claim is specific. No existing open-source tool predicts drug resistance from sequence alone with this coverage and accuracy. The novelty is not a new neural architecture; it is the demonstration that a protein language model plus a drug fingerprint, combined in a simple classifier, outperforms structure-based methods while removing the structural data requirement entirely. That is a practical, deployable advance.

The public interest case is direct. Antimicrobial resistance is a global commons problem. The tools to address it should be commons infrastructure. TOPOLOGIX will be released under an OSI-approved license, with all code, benchmarks, and model weights public. The project reduces dependence on commercial structural biology software and proprietary mutation databases, giving independent researchers and LMIC public health laboratories the capacity to run resistance prediction locally, on their own data, without sending their sequences to a foreign cloud service.

The applicant is an independent researcher with a track record of delivering rigorous, pre-registered computational work across multiple domains, including a 1,847-record literature screen for Bayesian prior elicitation, a 14-parameter MCMC calibration of a pharmacological model, and four independent data pipelines built for corpus analysis. The TOPOLOGIX codebase is maintained on GitHub under the applicant's account, with the negative-result predecessor studies archived on OSF and Zenodo. The work is ready for the next stage: production hardening, community adoption, and deployment in the settings where it matters most.

ESSAY: OPEN SOURCE AND TRUST

The trust case for TOPOLOGIX rests on three concrete properties: the code is auditable, the results are reproducible, and the negative results are public. The full pipeline, from raw sequence to prediction, will be released under an OSI-approved license. Every benchmark script, every preprocessing step, and every model weight will be in the repository. A reviewer can run the entire evaluation from a clean environment and verify the reported AUROC values. This is the standard the applicant applied to the predecessor studies, which were pre-registered on OSF before data analysis, and it is the standard TOPOLOGIX will maintain.

Reproducibility is enforced by containerization. The pipeline will ship as a Docker image with pinned dependency versions, so a laboratory in Nigeria runs exactly the same software as the developer. The benchmark suite is frozen at the start of the grant period, and any changes to the pipeline must pass the full suite before release. This prevents the common failure mode where a published tool cannot be re-run by anyone other than the author.

The negative results are part of the trust story. The applicant published the finding that topological features do not predict hERG cardiotoxicity beyond plain descriptors, and the finding that interface topology carries no signal for drug resistance. These results are archived on OSF and Zenodo and are in review at peer-reviewed journals. A project that reports its failures as clearly as its successes is a project that can be trusted with public health infrastructure.

The independence angle is structural. TOPOLOGIX requires no commercial structural biology suite, no proprietary database, and no cloud dependency. It runs on standard Python libraries. A public health laboratory can deploy it on a laptop. This is digital sovereignty in practice: the capacity to make predictions about local resistance patterns without routing data through a foreign commercial service.

ESSAY: TECHNICAL FEASIBILITY AND PLAN

The applicant has already built and evaluated TOPOLOGIX. The core classifier exists, the benchmarks are defined, and the results are reproducible from the public repository. The grant period is therefore not a research proposal; it is a hardening and deployment plan with clear milestones.

Month one to three: freeze the benchmark suite, reproduce all published results from a clean environment, containerize the pipeline. Success criterion: a fresh clone of the repository plus the Docker image reproduces the reported AUROC values within 0.01.

Month four to six: expand the training data by curating additional mutation-resistance pairs from published literature, targeting a 30 percent increase. Success criterion: the expanded training set is publicly released with provenance for every entry.

Month seven to nine: build the command-line interface and a simple web API, with documentation for bench scientists. Success criterion: a user with no machine learning background can run a prediction from a FASTA file and a SMILES string in under ten minutes.

Month ten to twelve: run a validation workshop with a Nigerian public health laboratory partner, collect structured feedback, publish the full validation study. Success criterion: the workshop produces a written report and at least one external user runs the tool independently.

The budget request is 50,000 euros. The applicant's time is 30,000 euros for twelve months of part-time development. Cloud compute for large-scale ESM-2 embedding generation is 10,000 euros. Community building, documentation, and the validation workshop are 10,000 euros. This is proportionate to the work and does not include any equipment, office space, or institutional overhead.

The risk register is honest. The main technical risk is that the expanded training set does not improve performance, which would mean the current benchmark results are close to the ceiling for this method class. The mitigation is that the tool is already useful at current performance, and the grant period would still deliver a production-ready, documented, benchmarked release. The main non-technical risk is community adoption, which the validation workshop directly addresses by building a relationship with a real deployment partner from month one.

CHECKLIST

- [ ] Verify current NLnet Foundation / NGI Zero Commons Fund application form and submission portal access
- [ ] Confirm the 50,000 euro budget request is within the fund's per-project cap for the current round
- [ ] Prepare TOPOLOGIX GitHub repository with README, license (OSI-approved, recommend Apache 2.0 or MIT), and reproduction instructions
- [ ] Archive the current TOPOLOGIX codebase and benchmark results on Zenodo with a DOI
- [ ] Confirm the two negative-result predecessor studies (hERG topology, resistance topology) are publicly archived on OSF/Zenodo with DOIs
- [ ] Draft a one-page project timeline document matching the month-by-month plan in the research statement
- [ ] Draft a one-page budget justification table breaking down the 50,000 euro request
- [ ] Prepare a short bio sketch including B.Pharm, University of Ibadan, M.Sc. Digital Health enrollment at HPI/Potsdam, and ORCID
- [ ] Identify and contact the Nigerian public health laboratory partner for the validation workshop; obtain a letter of intent
- [ ] Confirm whether the fund requires a data management plan; if so, draft one covering the expanded training set release
- [ ] Verify the fund's open-source license requirements and confirm the chosen license is compliant
- [ ] Check whether the fund requires a fiscal sponsor or bank account in a specific jurisdiction; confirm the applicant's independent researcher status is eligible
- [ ] Submit the application through the NLnet portal with all attachments in the required formats

EDITOR NOTES

- Eligibility risk: the applicant is enrolled in an M.Sc. program starting Winter Semester 2026/27 and employed as National Product Manager at Synthcare from March 2026. The fund should be checked for any restriction on funding individuals who are concurrently employed or enrolled; the application should clarify that TOPOLOGIX development is independent of both the M.Sc. coursework and the Synthcare role.
- The Nigerian public health laboratory partner is not named in the profile. The applicant must insert a real partner name and a letter of intent before submission; this is a gap that cannot be filled from the profile.
- The budget of 50,000 euros is an assumption. The fund's per-project cap for the current round must be verified, and the budget adjusted to match the actual cap and the fund's stated expectations for solo researcher grants.
- The profile lists multiple research lines. This application deliberately features TOPOLOGIX only, because it is the best fit for NGI Zero's open infrastructure and public interest criteria. The CCT model, neurocascade, ergofluids, and psyche-twin are not mentioned in the application materials; do not add them, as they would dilute the thematic match.
- The applicant's employment at Synthcare (a commercial entity) and the ergofluids project's explicit "no IP/product claims" stance should be checked for any conflict with the fund's independence criteria. The application should not imply that Synthcare endorses or is involved in TOPOLOGIX.

AI Draft — NLnet Foundation / NGI Zero Commons Fund