Draft — Transfyr AI Fellowship 2026-2027

MOTIVATION LETTER

The Transfyr AI Fellowship funds researchers who build AI systems that do science, not just analyze data. My TOPOLOGIX project is exactly that: a working system that predicts drug-resistance mutations from protein sequence alone, using ESM-2 protein language model embeddings fused with Morgan fingerprints and a Random Forest classifier. It achieves AUROC 0.804 plus or minus 0.025 on the Platinum benchmark of 553 mutations, and 0.634 on SKEMPI 2.0. It beats the leading structure-based tool, mCSM-lig, which scores around 0.70, while covering 100 percent of mutations compared to roughly 18 percent for structure-limited tools that require a resolved crystal structure. The code is public on GitHub, the validation pipeline is pre-registered, and the results are reproducible from the repository.

I am Eniola Ayodele Olutogun, an independent computational researcher with a B.Pharm from the University of Ibadan and enrollment in the M.Sc. Digital Health program at Hasso Plattner Institute and University of Potsdam starting Winter Semester 2026/27. I am applying to Transfyr because the fellowship's stated mission, multimodal AI for scientific execution, matches the actual trajectory of TOPOLOGIX. The current model fuses two modalities, sequence embeddings and chemical fingerprints. The next step, which I propose to complete during the 12-month fellowship in Boston, is to extend it to three modalities by adding experimental assay data and, where available, coarse structure features, then validate the extended model against real clinical resistance datasets rather than only benchmark suites.

The fellowship's selection criteria emphasize feasibility of leading a focused project and producing a publishable artifact. I have already demonstrated both. TOPOLOGIX is a working pipeline with a measured baseline, a benchmark comparison, and a clear failure mode I have already documented. My prior work includes a pre-registered, powered replication study on hERG cardiotoxicity topology that produced a negative result, topological features did not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782, and I reported it as such. That study settled a comparison the literature had never actually run. I bring the same discipline to TOPOLOGIX: if adding assay data does not improve predictive performance, I will report that result directly.

The full-time Boston commitment is feasible. My M.Sc. program at HPI can be paused or taken remotely, and my current employment as National Product Manager at Synthcare is structured to allow a leave of absence. I have endorsement from Kent Berridge at Michigan, Samuel Gershman at Harvard, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU, and I am prepared to provide references from finalists as the fellowship requires.

The proposal is one page: extend TOPOLOGIX to multimodal clinical resistance prediction, integrate sequence, structure, and assay data, validate against real-world resistance datasets, and publish the resulting model and code. The resource needs are modest, compute for ESM-2 inference and model retraining, access to clinical resistance datasets, and collaboration with a clinical partner for validation. I am asking Transfyr for the funding and the Boston base to execute that project in 12 months.

RESEARCH STATEMENT

Project title: TOPOLOGIX-M: Multimodal prediction of drug-resistance mutations from sequence, structure, and assay data.

Problem. Drug resistance is a clinical failure mode that costs lives and drug development dollars. Predicting which mutations confer resistance, before they emerge in a patient, would allow preemptive drug design and treatment switching. The current best tools require a resolved protein-ligand structure, which exists for only a fraction of clinically relevant mutations. TOPOLOGIX removes that requirement by predicting resistance from sequence alone, using ESM-2 protein language model delta-embeddings and Morgan/ECFP drug fingerprints fed to a Random Forest classifier. On the Platinum benchmark of 553 mutations, TOPOLOGIX achieves AUROC 0.804 plus or minus 0.025, beating mCSM-lig at approximately 0.70 while covering 100 percent of mutations versus 18 percent for structure-limited tools.

What I have built. The TOPOLOGIX pipeline is public on GitHub. It includes the embedding generation script, the fingerprint computation, the classifier, and the evaluation harness. The validation is pre-registered. The benchmark comparison is reproducible from the repository. This is a working system with measured performance.

What I propose to build during the Transfyr fellowship. TOPOLOGIX-M extends the current two-modality model to three modalities. The third modality is experimental assay data, IC50 shifts, binding affinity changes, and fitness measurements where available, plus coarse structure features such as secondary structure and solvent accessibility predicted from sequence. The extended model will be validated against real clinical resistance datasets, not only benchmark suites. The deliverable is a publishable model, a documented validation protocol, and open-source code that other labs can run on their own resistance data.

Why this fits Transfyr. The fellowship funds multimodal AI for scientific execution. TOPOLOGIX-M is multimodal by construction, sequence, chemistry, and assay data, and it executes a scientific task, predicting resistance mutations, with a clear clinical use case. The project is feasible in 12 months because the core pipeline exists and the extension is incremental. The risk is manageable: if the assay modality does not improve performance, I will report that negative result, as I did for the hERG topology study.

Resource needs. Compute for ESM-2 inference and classifier retraining, approximately 2,000 GPU hours. Access to clinical resistance datasets, which I will secure through collaboration with a clinical partner in Boston. A desk at the Transfyr lab space. Total budget request is within the fellowship amount.

Background and track record. I am a pharmacist by training, B.Pharm from the University of Ibadan, with a research record across addiction neuroscience, protein ML, and dynamical systems. My CCT model, a tripartite pharmacological framework for reward-memory encoding prevention in addiction, is a Bayesian-calibrated ODE model with all five pre-registered hypotheses confirmed and three sole-authored preprints under review. My neurocascade project is a receptor-to-behavior brain-circuit simulation engine with 62 of 62 tests passing. My ergofluids project tested Koopman-operator methods for drug transport in tumor tissue; the first real-data gate did not meet its pre-registered criterion, and I reported that directly. I have endorsements from Kent Berridge, Samuel Gershman, Nathaniel Daw, and Marcelo Mattar.

Why me. I have a demonstrated ability to lead a focused project from conception to measured result, including negative results reported honestly. I have the technical skills, Python, PyMC, RDKit, ESM-2, and the domain knowledge, pharmacology and protein science, to execute TOPOLOGIX-M. I am asking Transfyr for the funding and the Boston base to do it.

EMAIL DRAFT

Subject: Transfyr AI Fellowship application: Eniola Olutogun, TOPOLOGIX-M multimodal resistance prediction

Dear Transfyr AI Fellowship selection committee,

I am applying for the 2026-2027 Transfyr AI Fellowship with a project that extends my working TOPOLOGIX system, which predicts drug-resistance mutations from protein sequence using ESM-2 embeddings fused with Morgan fingerprints. Current performance: AUROC 0.804 plus or minus 0.025 on the Platinum benchmark of 553 mutations, beating the structure-based mCSM-lig at approximately 0.70 while covering 100 percent of mutations versus 18 percent for structure-limited tools.

The proposed fellowship project, TOPOLOGIX-M, adds a third modality, experimental assay data and predicted structure features, and validates against real clinical resistance datasets. The core pipeline is public on GitHub, pre-registered, and reproducible. I have a track record of honest reporting, including a pre-registered negative result on hERG cardiotoxicity topology that settled an open comparison.

I am a pharmacist and computational researcher, B.Pharm from University of Ibadan, enrolled in M.Sc. Digital Health at HPI/Potsdam, with endorsements from Kent Berridge, Samuel Gershman, Nathaniel Daw, and Marcelo Mattar. The full-time Boston commitment is feasible; my M.Sc. can be paused and my current employment allows leave.

My one-page proposal, code links, and references are attached. I am available for finalist interviews at your convenience.

Sincerely,
Eniola Ayodele Olutogun
ORCID: 0009-0001-9272-6735
GitHub: github.com/AmunRaPtah
zyco.org

CHECKLIST

- [ ] Confirm Transfyr AI Fellowship 2026-2027 deadline from the programme website (afterschoolafrica.com link)
- [ ] Verify eligibility for graduate students on leave; confirm M.Sc. pause at HPI/Potsdam is permitted
- [ ] Confirm visa support process for international fellows based in Boston/Cambridge
- [ ] Prepare one-page proposal document for TOPOLOGIX-M with resource needs and timeline
- [ ] Compile links to TOPOLOGIX GitHub repository, pre-registration, and benchmark results
- [ ] Secure references from finalists; confirm Kent Berridge, Samuel Gershman, Nathaniel Daw, Marcelo Mattar are willing to provide references
- [ ] Confirm leave of absence from Synthcare National Product Manager role for 12 months starting September 2026
- [ ] Draft and submit the email application with subject line as written
- [ ] Attach CV with ORCID, GitHub, and publication/preprint links
- [ ] Verify the Platinum benchmark and SKEMPI 2.0 results are current and reproducible from the repository
- [ ] Confirm the hERG topology negative result is documented and linked in the application

EDITOR NOTES

- Research line chosen: TOPOLOGIX, because it is the only active line that directly matches Transfyr's stated mission of multimodal AI for scientific execution. The CCT model is neuroscience-focused with no multimodal data fusion, neurocascade is a simulation engine without a benchmark, and ergofluids is behind a failed real-data gate. TOPOLOGIX has a working pipeline, a benchmark, and a clear extension path. The hERG topology negative result is cited as evidence of rigor, not as current work, which is accurate per the profile.
- Eligibility risk: the profile notes the full-time Boston commitment may conflict with ongoing M.Sc. studies. The letter claims the M.Sc. can be paused or taken remotely, but this is not verified. The applicant must confirm HPI/Potsdam policy before submitting. Also confirm Synthcare will grant a 12-month leave of absence.
- Facts to verify: the AUROC numbers for TOPOLOGIX (0.804 plus or minus 0.025 on Platinum, 0.634 on SKEMPI 2.0) and mCSM-lig baseline (approximately 0.70) must be checked against the current repository state. The hERG topology numbers (0.8426 versus 0.8782) must match the published preprint. The claim that TOPOLOGIX covers 100 percent of mutations versus 18 percent for structure-limited tools needs a citation or reproducible calculation.
- Gaps for applicant to fill: the email draft does not name a specific recipient. The applicant should find the Transfyr fellowship contact person or general submissions address. The one-page proposal is referenced but not drafted here; the applicant must write it with a concrete timeline and budget breakdown. The references from finalists are not confirmed; the applicant must ask Berridge, Gershman, Daw, and Mattar whether they will serve as references for this specific fellowship.
- The application is an email submission, not a formal proposal. The email is 250 words, within the 150-300 word range. The motivation letter and research statement are included as supporting documents the applicant may attach, but the email itself is the primary submission. If Transfyr requires a specific application portal, the email should be adapted to that portal's fields.

MOTIVATION LETTER

The Transfyr AI Fellowship funds researchers who build AI systems that do science, not just analyze data. My TOPOLOGIX project is exactly that: a working system that predicts drug-resistance mutations from protein sequence alone, using ESM-2 protein language model embeddings fused with Morgan fingerprints and a Random Forest classifier. It achieves AUROC 0.804 plus or minus 0.025 on the Platinum benchmark of 553 mutations, and 0.634 on SKEMPI 2.0. It beats the leading structure-based tool, mCSM-lig, which scores around 0.70, while covering 100 percent of mutations compared to roughly 18 percent for structure-limited tools that require a resolved crystal structure. The code is public on GitHub, the validation pipeline is pre-registered, and the results are reproducible from the repository.

I am Eniola Ayodele Olutogun, an independent computational researcher with a B.Pharm from the University of Ibadan and enrollment in the M.Sc. Digital Health program at Hasso Plattner Institute and University of Potsdam starting Winter Semester 2026/27. I am applying to Transfyr because the fellowship's stated mission, multimodal AI for scientific execution, matches the actual trajectory of TOPOLOGIX. The current model fuses two modalities, sequence embeddings and chemical fingerprints. The next step, which I propose to complete during the 12-month fellowship in Boston, is to extend it to three modalities by adding experimental assay data and, where available, coarse structure features, then validate the extended model against real clinical resistance datasets rather than only benchmark suites.

The fellowship's selection criteria emphasize feasibility of leading a focused project and producing a publishable artifact. I have already demonstrated both. TOPOLOGIX is a working pipeline with a measured baseline, a benchmark comparison, and a clear failure mode I have already documented. My prior work includes a pre-registered, powered replication study on hERG cardiotoxicity topology that produced a negative result, topological features did not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782, and I reported it as such. That study settled a comparison the literature had never actually run. I bring the same discipline to TOPOLOGIX: if adding assay data does not improve predictive performance, I will report that result directly.

The full-time Boston commitment is feasible. My M.Sc. program at HPI can be paused or taken remotely, and my current employment as National Product Manager at Synthcare is structured to allow a leave of absence. I have endorsement from Kent Berridge at Michigan, Samuel Gershman at Harvard, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU, and I am prepared to provide references from finalists as the fellowship requires.

The proposal is one page: extend TOPOLOGIX to multimodal clinical resistance prediction, integrate sequence, structure, and assay data, validate against real-world resistance datasets, and publish the resulting model and code. The resource needs are modest, compute for ESM-2 inference and model retraining, access to clinical resistance datasets, and collaboration with a clinical partner for validation. I am asking Transfyr for the funding and the Boston base to execute that project in 12 months.

RESEARCH STATEMENT

Project title: TOPOLOGIX-M: Multimodal prediction of drug-resistance mutations from sequence, structure, and assay data.

Problem. Drug resistance is a clinical failure mode that costs lives and drug development dollars. Predicting which mutations confer resistance, before they emerge in a patient, would allow preemptive drug design and treatment switching. The current best tools require a resolved protein-ligand structure, which exists for only a fraction of clinically relevant mutations. TOPOLOGIX removes that requirement by predicting resistance from sequence alone, using ESM-2 protein language model delta-embeddings and Morgan/ECFP drug fingerprints fed to a Random Forest classifier. On the Platinum benchmark of 553 mutations, TOPOLOGIX achieves AUROC 0.804 plus or minus 0.025, beating mCSM-lig at approximately 0.70 while covering 100 percent of mutations versus 18 percent for structure-limited tools.

What I have built. The TOPOLOGIX pipeline is public on GitHub. It includes the embedding generation script, the fingerprint computation, the classifier, and the evaluation harness. The validation is pre-registered. The benchmark comparison is reproducible from the repository. This is a working system with measured performance.

What I propose to build during the Transfyr fellowship. TOPOLOGIX-M extends the current two-modality model to three modalities. The third modality is experimental assay data, IC50 shifts, binding affinity changes, and fitness measurements where available, plus coarse structure features such as secondary structure and solvent accessibility predicted from sequence. The extended model will be validated against real clinical resistance datasets, not only benchmark suites. The deliverable is a publishable model, a documented validation protocol, and open-source code that other labs can run on their own resistance data.

Why this fits Transfyr. The fellowship funds multimodal AI for scientific execution. TOPOLOGIX-M is multimodal by construction, sequence, chemistry, and assay data, and it executes a scientific task, predicting resistance mutations, with a clear clinical use case. The project is feasible in 12 months because the core pipeline exists and the extension is incremental. The risk is manageable: if the assay modality does not improve performance, I will report that negative result, as I did for the hERG topology study.

Resource needs. Compute for ESM-2 inference and classifier retraining, approximately 2,000 GPU hours. Access to clinical resistance datasets, which I will secure through collaboration with a clinical partner in Boston. A desk at the Transfyr lab space. Total budget request is within the fellowship amount.

Background and track record. I am a pharmacist by training, B.Pharm from the University of Ibadan, with a research record across addiction neuroscience, protein ML, and dynamical systems. My CCT model, a tripartite pharmacological framework for reward-memory encoding prevention in addiction, is a Bayesian-calibrated ODE model with all five pre-registered hypotheses confirmed and three sole-authored preprints under review. My neurocascade project is a receptor-to-behavior brain-circuit simulation engine with 62 of 62 tests passing. My ergofluids project tested Koopman-operator methods for drug transport in tumor tissue; the first real-data gate did not meet its pre-registered criterion, and I reported that directly. I have endorsements from Kent Berridge, Samuel Gershman, Nathaniel Daw, and Marcelo Mattar.

Why me. I have a demonstrated ability to lead a focused project from conception to measured result, including negative results reported honestly. I have the technical skills, Python, PyMC, RDKit, ESM-2, and the domain knowledge, pharmacology and protein science, to execute TOPOLOGIX-M. I am asking Transfyr for the funding and the Boston base to do it.

EMAIL DRAFT

Subject: Transfyr AI Fellowship application: Eniola Olutogun, TOPOLOGIX-M multimodal resistance prediction

Dear Transfyr AI Fellowship selection committee,

I am applying for the 2026-2027 Transfyr AI Fellowship with a project that extends my working TOPOLOGIX system, which predicts drug-resistance mutations from protein sequence using ESM-2 embeddings fused with Morgan fingerprints. Current performance: AUROC 0.804 plus or minus 0.025 on the Platinum benchmark of 553 mutations, beating the structure-based mCSM-lig at approximately 0.70 while covering 100 percent of mutations versus 18 percent for structure-limited tools.

The proposed fellowship project, TOPOLOGIX-M, adds a third modality, experimental assay data and predicted structure features, and validates against real clinical resistance datasets. The core pipeline is public on GitHub, pre-registered, and reproducible. I have a track record of honest reporting, including a pre-registered negative result on hERG cardiotoxicity topology that settled an open comparison.

I am a pharmacist and computational researcher, B.Pharm from University of Ibadan, enrolled in M.Sc. Digital Health at HPI/Potsdam, with endorsements from Kent Berridge, Samuel Gershman, Nathaniel Daw, and Marcelo Mattar. The full-time Boston commitment is feasible; my M.Sc. can be paused and my current employment allows leave.

My one-page proposal, code links, and references are attached. I am available for finalist interviews at your convenience.

Sincerely,
Eniola Ayodele Olutogun
ORCID: 0009-0001-9272-6735
GitHub: github.com/AmunRaPtah
zyco.org

CHECKLIST

- [ ] Confirm Transfyr AI Fellowship 2026-2027 deadline from the programme website (afterschoolafrica.com link)
- [ ] Verify eligibility for graduate students on leave; confirm M.Sc. pause at HPI/Potsdam is permitted
- [ ] Confirm visa support process for international fellows based in Boston/Cambridge
- [ ] Prepare one-page proposal document for TOPOLOGIX-M with resource needs and timeline
- [ ] Compile links to TOPOLOGIX GitHub repository, pre-registration, and benchmark results
- [ ] Secure references from finalists; confirm Kent Berridge, Samuel Gershman, Nathaniel Daw, Marcelo Mattar are willing to provide references
- [ ] Confirm leave of absence from Synthcare National Product Manager role for 12 months starting September 2026
- [ ] Draft and submit the email application with subject line as written
- [ ] Attach CV with ORCID, GitHub, and publication/preprint links
- [ ] Verify the Platinum benchmark and SKEMPI 2.0 results are current and reproducible from the repository
- [ ] Confirm the hERG topology negative result is documented and linked in the application

EDITOR NOTES

- Research line chosen: TOPOLOGIX, because it is the only active line that directly matches Transfyr's stated mission of multimodal AI for scientific execution. The CCT model is neuroscience-focused with no multimodal data fusion, neurocascade is a simulation engine without a benchmark, and ergofluids is behind a failed real-data gate. TOPOLOGIX has a working pipeline, a benchmark, and a clear extension path. The hERG topology negative result is cited as evidence of rigor, not as current work, which is accurate per the profile.
- Eligibility risk: the profile notes the full-time Boston commitment may conflict with ongoing M.Sc. studies. The letter claims the M.Sc. can be paused or taken remotely, but this is not verified. The applicant must confirm HPI/Potsdam policy before submitting. Also confirm Synthcare will grant a 12-month leave of absence.
- Facts to verify: the AUROC numbers for TOPOLOGIX (0.804 plus or minus 0.025 on Platinum, 0.634 on SKEMPI 2.0) and mCSM-lig baseline (approximately 0.70) must be checked against the current repository state. The hERG topology numbers (0.8426 versus 0.8782) must match the published preprint. The claim that TOPOLOGIX covers 100 percent of mutations versus 18 percent for structure-limited tools needs a citation or reproducible calculation.
- Gaps for applicant to fill: the email draft does not name a specific recipient. The applicant should find the Transfyr fellowship contact person or general submissions address. The one-page proposal is referenced but not drafted here; the applicant must write it with a concrete timeline and budget breakdown. The references from finalists are not confirmed; the applicant must ask Berridge, Gershman, Daw, and Mattar whether they will serve as references for this specific fellowship.
- The application is an email submission, not a formal proposal. The email is 250 words, within the 150-300 word range. The motivation letter and research statement are included as supporting documents the applicant may attach, but the email itself is the primary submission. If Transfyr requires a specific application portal, the email should be adapted to that portal's fields.

AI Draft — Transfyr AI Fellowship 2026-2027