Draft — Journeys Incubator

MOTIVATION LETTER

The Platinum benchmark contains 553 mutations across 179 proteins, and structure-based tools like mCSM-lig can only score roughly 18 percent of them because they require a crystal structure. My venture predicts drug resistance mutations from protein sequence alone, using ESM-2 protein language model delta-embeddings combined with ECFP4 drug fingerprints and a Random Forest classifier. On that benchmark, with protein-grouped cross-validation, the model achieves an AUROC of 0.804 plus or minus 0.025, against a published state-of-the-art of approximately 0.70 for mCSM-lig. It covers 100 percent of mutations, not 18 percent.

I am applying to Journeys Incubator because the Deep Tech track explicitly names AI/ML and synthetic biology as target areas, and because the programme is run by Helsinki University, which gives me access to academic rigor and a health-tech network that a pre-seed founder cannot build alone. The venture is at proof-of-concept stage, not yet incorporated, which matches the programme's stated preference for early-stage teams without significant turnover. I am a pharmacist turned machine learning engineer, and I am the sole author of the codebase and the validation pipeline. That combination means I can speak to both the biology and the model.

The incubator requires commitment to mandatory in-person sessions and bi-monthly mentoring. I can commit to that schedule. The venture's roadmap is specific: fine-tune ESM-2 on the SKEMPI 3K mutation set to reach an AUROC of at least 0.70 on that harder benchmark, then run a pilot with Servier in Suresnes. The current SKEMPI 2.0 score is 0.634, so the target is measurable and the gap is defined. The model already beats published state-of-the-art on Platinum; the SKEMPI work is the next validation gate, not a speculative hope.

The global health case is direct. Antimicrobial resistance and antiviral escape mutations are sequence-level problems. If a clinician or a drug developer can ask whether a given mutation will confer resistance before the mutation appears in a patient, they can pre-empt resistance rather than react to it. That is the product. The business model is a software platform sold to pharma and biotech, starting with a Servier pilot and moving to annual recurring revenue.

I am asking the incubator to accelerate a validated model with a clear benchmark record, a named pilot partner, and a founder who can execute the technical work herself.

RESEARCH STATEMENT

The problem is that drug resistance is usually detected after it emerges in a patient or in a clinical trial. By then, the therapeutic option is already failing. My venture predicts resistance mutations from protein sequence alone, which means the prediction can be made before the mutation is observed, and it can be made for any protein, not just the minority with solved crystal structures.

The method has three components. First, ESM-2, a protein language model pre-trained on millions of sequences, generates embeddings for the wild-type and mutant protein. The delta between those embeddings captures the biophysical effect of the mutation without requiring a structure. Second, ECFP4 fingerprints encode the drug molecule that binds to the protein. Third, a Random Forest classifier learns the interaction between the mutation effect and the drug identity. The output is a binary prediction: resistant or not resistant.

The validation record is concrete. On the Platinum benchmark, 553 mutations with protein-grouped cross-validation, the model achieves AUROC 0.804 plus or minus 0.025. The published state-of-the-art for mCSM-lig is approximately 0.70. On SKEMPI 2.0, a harder benchmark of binding affinity changes, the model scores 0.634. That gap between 0.804 and 0.634 is the roadmap. SKEMPI 2.0 includes mutations that change binding affinity in both directions, and the model was not fine-tuned on that distribution. The next step is to fine-tune ESM-2 on the SKEMPI 3K mutation set, which should raise the AUROC to at least 0.70. That is the technical gate for the Servier pilot.

The commercial path is a pilot with Servier in Suresnes, followed by annual recurring revenue from pharma and biotech customers who need resistance prediction in oncology, antiviral development, and antimicrobial resistance programs. The venture is pre-seed, proof-of-concept validated, and not yet incorporated. The founder, Eniola Olutogun, is a pharmacist and machine learning engineer and the sole author of the codebase.

The fit with Journeys Incubator is specific. The Deep Tech track targets AI/ML and synthetic biology, and this venture sits exactly at that intersection. The programme is open to early-stage ventures without significant turnover, which describes this venture precisely. The mandatory in-person sessions and bi-monthly mentoring are a good match for a founder who needs structured validation of the commercial model while the technical roadmap is already clear.

The scientific claim is narrow and falsifiable. The model predicts resistance from sequence alone without requiring a crystal structure. It covers 100 percent of mutations, not 18 percent. It beats published state-of-the-art on Platinum. The next benchmark is SKEMPI 3K, and the target is an AUROC of at least 0.70. Those are the numbers that will be judged.

ESSAY: WHY THIS VENTURE AND WHY NOW

Drug resistance is a moving target, and the tools to predict it have been structurally limited. Structure-based methods require a crystal structure, which exists for only a fraction of clinically relevant proteins. That is why mCSM-lig, the published state-of-the-art, covers only about 18 percent of Platinum benchmark mutations. My venture removes that constraint by using protein language models, which learn from sequence alone. ESM-2 delta-embeddings capture the effect of a mutation without any structural input. Protein language models reached sufficient accuracy only in the last few years, which is the technical reason the venture exists now.

The timing is also commercial. Pharma companies are investing heavily in AI-driven drug discovery, and resistance prediction is a bottleneck in oncology, antiviral development, and antimicrobial resistance programs. A model that covers 100 percent of mutations, not 18 percent, is a different capability. The venture already has a named pilot partner in Servier and a validation roadmap through SKEMPI 3K.

The founder's background is the third reason for now. Eniola Olutogun is a pharmacist who became a machine learning engineer. She can read a mutation report and debug a Random Forest in the same hour. That dual fluency is rare and it is why the venture is sole-authored: the biology and the code live in one head.

The risk is honest. The SKEMPI 2.0 score of 0.634 shows the model does not yet generalize to binding affinity changes without fine-tuning. The roadmap addresses that directly. The venture is pre-seed and not incorporated, which is appropriate for this stage. The next twelve months are about hitting the SKEMPI 3K target and converting the Servier pilot into a contract.

CHECKLIST

- [ ] Confirm Helsinki Capital Region residency or verify eligibility for non-resident applicants
- [ ] Prepare pitch deck in Dealum format, max 10 slides, including AUROC benchmark chart
- [ ] Record 3-minute video pitch covering problem, method, validation, and roadmap
- [ ] Complete Dealum application form for Journeys Incubator Deep Tech track
- [ ] Verify deadline 2024-08-09 and submit before 23:59 EEST
- [ ] Prepare one-page technical appendix with Platinum and SKEMPI benchmark details
- [ ] Confirm availability for mandatory in-person sessions in Helsinki
- [ ] Draft commitment statement for bi-monthly mentoring schedule
- [ ] Verify current SKEMPI 2.0 score and Platinum AUROC numbers before submission
- [ ] Confirm Servier pilot status and named contact at Suresnes

EDITOR NOTES

- Eligibility risk: Journeys Incubator is open to Helsinki Capital Region residents; the applicant profile does not state Finnish residency. This must be verified before submission, and if not eligible, the application should be withdrawn or the residency question answered honestly.
- The SKEMPI 2.0 score of 0.634 is a weakness that the application acknowledges; do not let a reviewer mistake it for a failure. The roadmap explicitly targets SKEMPI 3K fine-tuning to reach AUROC 0.70, and that gate must be stated as the next milestone, not as achieved.
- The Servier pilot is named in the profile but no contract or letter of intent is confirmed. The application must not imply a signed agreement. Use the phrase "pilot in discussion" or "named pilot partner" and verify the current status before submission.
- The venture is not yet incorporated. The application should state this clearly and frame it as appropriate for the incubator's early-stage preference, not as a deficiency.
- The applicant must insert personal detail about Helsinki residency, availability for in-person sessions, and any prior relationship with Helsinki University or Finnish health-tech ecosystem, as none of this is in the profile.

MOTIVATION LETTER

The Platinum benchmark contains 553 mutations across 179 proteins, and structure-based tools like mCSM-lig can only score roughly 18 percent of them because they require a crystal structure. My venture predicts drug resistance mutations from protein sequence alone, using ESM-2 protein language model delta-embeddings combined with ECFP4 drug fingerprints and a Random Forest classifier. On that benchmark, with protein-grouped cross-validation, the model achieves an AUROC of 0.804 plus or minus 0.025, against a published state-of-the-art of approximately 0.70 for mCSM-lig. It covers 100 percent of mutations, not 18 percent.

I am applying to Journeys Incubator because the Deep Tech track explicitly names AI/ML and synthetic biology as target areas, and because the programme is run by Helsinki University, which gives me access to academic rigor and a health-tech network that a pre-seed founder cannot build alone. The venture is at proof-of-concept stage, not yet incorporated, which matches the programme's stated preference for early-stage teams without significant turnover. I am a pharmacist turned machine learning engineer, and I am the sole author of the codebase and the validation pipeline. That combination means I can speak to both the biology and the model.

The incubator requires commitment to mandatory in-person sessions and bi-monthly mentoring. I can commit to that schedule. The venture's roadmap is specific: fine-tune ESM-2 on the SKEMPI 3K mutation set to reach an AUROC of at least 0.70 on that harder benchmark, then run a pilot with Servier in Suresnes. The current SKEMPI 2.0 score is 0.634, so the target is measurable and the gap is defined. The model already beats published state-of-the-art on Platinum; the SKEMPI work is the next validation gate, not a speculative hope.

The global health case is direct. Antimicrobial resistance and antiviral escape mutations are sequence-level problems. If a clinician or a drug developer can ask whether a given mutation will confer resistance before the mutation appears in a patient, they can pre-empt resistance rather than react to it. That is the product. The business model is a software platform sold to pharma and biotech, starting with a Servier pilot and moving to annual recurring revenue.

I am asking the incubator to accelerate a validated model with a clear benchmark record, a named pilot partner, and a founder who can execute the technical work herself.

RESEARCH STATEMENT

The problem is that drug resistance is usually detected after it emerges in a patient or in a clinical trial. By then, the therapeutic option is already failing. My venture predicts resistance mutations from protein sequence alone, which means the prediction can be made before the mutation is observed, and it can be made for any protein, not just the minority with solved crystal structures.

The method has three components. First, ESM-2, a protein language model pre-trained on millions of sequences, generates embeddings for the wild-type and mutant protein. The delta between those embeddings captures the biophysical effect of the mutation without requiring a structure. Second, ECFP4 fingerprints encode the drug molecule that binds to the protein. Third, a Random Forest classifier learns the interaction between the mutation effect and the drug identity. The output is a binary prediction: resistant or not resistant.

The validation record is concrete. On the Platinum benchmark, 553 mutations with protein-grouped cross-validation, the model achieves AUROC 0.804 plus or minus 0.025. The published state-of-the-art for mCSM-lig is approximately 0.70. On SKEMPI 2.0, a harder benchmark of binding affinity changes, the model scores 0.634. That gap between 0.804 and 0.634 is the roadmap. SKEMPI 2.0 includes mutations that change binding affinity in both directions, and the model was not fine-tuned on that distribution. The next step is to fine-tune ESM-2 on the SKEMPI 3K mutation set, which should raise the AUROC to at least 0.70. That is the technical gate for the Servier pilot.

The commercial path is a pilot with Servier in Suresnes, followed by annual recurring revenue from pharma and biotech customers who need resistance prediction in oncology, antiviral development, and antimicrobial resistance programs. The venture is pre-seed, proof-of-concept validated, and not yet incorporated. The founder, Eniola Olutogun, is a pharmacist and machine learning engineer and the sole author of the codebase.

The fit with Journeys Incubator is specific. The Deep Tech track targets AI/ML and synthetic biology, and this venture sits exactly at that intersection. The programme is open to early-stage ventures without significant turnover, which describes this venture precisely. The mandatory in-person sessions and bi-monthly mentoring are a good match for a founder who needs structured validation of the commercial model while the technical roadmap is already clear.

The scientific claim is narrow and falsifiable. The model predicts resistance from sequence alone without requiring a crystal structure. It covers 100 percent of mutations, not 18 percent. It beats published state-of-the-art on Platinum. The next benchmark is SKEMPI 3K, and the target is an AUROC of at least 0.70. Those are the numbers that will be judged.

ESSAY: WHY THIS VENTURE AND WHY NOW

Drug resistance is a moving target, and the tools to predict it have been structurally limited. Structure-based methods require a crystal structure, which exists for only a fraction of clinically relevant proteins. That is why mCSM-lig, the published state-of-the-art, covers only about 18 percent of Platinum benchmark mutations. My venture removes that constraint by using protein language models, which learn from sequence alone. ESM-2 delta-embeddings capture the effect of a mutation without any structural input. Protein language models reached sufficient accuracy only in the last few years, which is the technical reason the venture exists now.

The timing is also commercial. Pharma companies are investing heavily in AI-driven drug discovery, and resistance prediction is a bottleneck in oncology, antiviral development, and antimicrobial resistance programs. A model that covers 100 percent of mutations, not 18 percent, is a different capability. The venture already has a named pilot partner in Servier and a validation roadmap through SKEMPI 3K.

The founder's background is the third reason for now. Eniola Olutogun is a pharmacist who became a machine learning engineer. She can read a mutation report and debug a Random Forest in the same hour. That dual fluency is rare and it is why the venture is sole-authored: the biology and the code live in one head.

The risk is honest. The SKEMPI 2.0 score of 0.634 shows the model does not yet generalize to binding affinity changes without fine-tuning. The roadmap addresses that directly. The venture is pre-seed and not incorporated, which is appropriate for this stage. The next twelve months are about hitting the SKEMPI 3K target and converting the Servier pilot into a contract.

CHECKLIST

- [ ] Confirm Helsinki Capital Region residency or verify eligibility for non-resident applicants
- [ ] Prepare pitch deck in Dealum format, max 10 slides, including AUROC benchmark chart
- [ ] Record 3-minute video pitch covering problem, method, validation, and roadmap
- [ ] Complete Dealum application form for Journeys Incubator Deep Tech track
- [ ] Verify deadline 2024-08-09 and submit before 23:59 EEST
- [ ] Prepare one-page technical appendix with Platinum and SKEMPI benchmark details
- [ ] Confirm availability for mandatory in-person sessions in Helsinki
- [ ] Draft commitment statement for bi-monthly mentoring schedule
- [ ] Verify current SKEMPI 2.0 score and Platinum AUROC numbers before submission
- [ ] Confirm Servier pilot status and named contact at Suresnes

EDITOR NOTES

- Eligibility risk: Journeys Incubator is open to Helsinki Capital Region residents; the applicant profile does not state Finnish residency. This must be verified before submission, and if not eligible, the application should be withdrawn or the residency question answered honestly.
- The SKEMPI 2.0 score of 0.634 is a weakness that the application acknowledges; do not let a reviewer mistake it for a failure. The roadmap explicitly targets SKEMPI 3K fine-tuning to reach AUROC 0.70, and that gate must be stated as the next milestone, not as achieved.
- The Servier pilot is named in the profile but no contract or letter of intent is confirmed. The application must not imply a signed agreement. Use the phrase "pilot in discussion" or "named pilot partner" and verify the current status before submission.
- The venture is not yet incorporated. The application should state this clearly and frame it as appropriate for the incubator's early-stage preference, not as a deficiency.
- The applicant must insert personal detail about Helsinki residency, availability for in-person sessions, and any prior relationship with Helsinki University or Finnish health-tech ecosystem, as none of this is in the profile.

AI Draft — Journeys Incubator