Draft — i-Lab Competition

MOTIVATION LETTER

A drug-resistant mutation can render a billion-euro therapy useless within months. The standard approach to predicting those mutations requires a protein crystal structure, which exists for only 17.6 percent of clinically relevant targets. My venture solves this bottleneck by predicting resistance mutations from protein sequence alone, using ESM-2 protein language model delta-embeddings fused with ECFP4 drug fingerprints and a Random Forest classifier. On the Platinum benchmark of 357 mutations, the model achieves an AUROC of 0.634 with 100 percent mutation coverage, compared to 17.6 percent for structure-limited tools. Published SOTA, mCSM-lig, reports an AUROC of 0.70 but requires a crystal structure for every prediction.

I am a pharmacist turned ML engineer. I hold a PharmD from the University of Ibadan and completed a machine learning engineering certification at Holberton School in Paris. This dual background lets me design models that respect biological constraints while exploiting the full expressivity of protein language models. I am the sole founder and sole author of the venture. The proof-of-concept is validated. The company is not yet incorporated.

The i-Lab competition, operated by Bpifrance, awards equity-free funding up to 600,000 euros for deep-tech projects at the proof-of-concept stage. My venture matches that profile exactly. The funding will support fine-tuning ESM-2 on the SKEMPI 3K mutation dataset, targeting an AUROC of 0.70 or higher. That improvement is the critical milestone for entering a pilot with Servier at their Suresnes site, which I have already initiated contact with. The pilot would validate the tool on Servier's internal oncology and antiviral pipelines.

I have submitted applications to IncubAlliance and AI House. A meeting with SEMIA and Quest for Health is in progress. WILCO One BioTech is scheduled for October 2026. These partnerships form a pipeline toward the EIC Accelerator and BPI i-Lab as follow-on funding. The i-Lab grant is the immediate next step to turn a validated proof-of-concept into a deployable product with a measurable path to annual recurring revenue.

The problem is urgent. Antimicrobial resistance alone is projected to cause 10 million deaths per year by 2050. My technology predicts resistance before it emerges in the clinic, giving drug developers a six- to twelve-month lead time to modify their candidates. That is the impact I am building toward, and the i-Lab competition is the right programme to accelerate that timeline.

TECHNICAL PROJECT DESCRIPTION

The venture predicts drug resistance mutations from protein sequence alone. The input is a wild-type protein sequence and a drug SMILES string. The model architecture has three components. First, ESM-2 generates per-residue embeddings for the protein sequence. Second, a delta-embedding is computed by subtracting the wild-type embedding from a mutated embedding, capturing the local biophysical effect of each possible single-point mutation. Third, the drug SMILES is converted to an ECFP4 fingerprint. The delta-embedding and the fingerprint are concatenated and passed to a Random Forest classifier that outputs a resistance probability for each mutation.

The model was benchmarked on the Platinum dataset, which contains 357 mutations across 10 protein targets with experimentally measured resistance outcomes. Evaluation used protein-grouped cross-validation, meaning all mutations from a given protein were held out together. This is a stricter and more realistic evaluation than random splits. The model achieved an AUROC of 0.634 with 100 percent mutation coverage. For comparison, mCSM-lig, the published SOTA, reports an AUROC of 0.70 but requires a crystal structure and therefore covers only 17.6 percent of the same mutations.

The current AUROC of 0.634 is a baseline. The immediate technical milestone is to reach AUROC 0.70 by fine-tuning ESM-2 on the SKEMPI 3K dataset, which contains 3,000 experimentally measured mutation binding effects. Fine-tuning will adjust the ESM-2 weights to better represent mutation-induced changes in protein-drug interaction surfaces. This is a standard transfer-learning step that has been shown to improve downstream prediction accuracy by 5 to 10 percent in published benchmarks.

The i-Lab funding will cover compute costs for fine-tuning on an A100 GPU cluster, validation on an independent set of 200 mutations from the PDBbind database, and integration of the fine-tuned model into a REST API for the Servier pilot. The pilot will test the tool on three Servier drug targets: one oncology kinase, one antiviral protease, and one antimicrobial target. Success criteria are an AUROC of 0.70 or higher on Servier's internal resistance data and a prediction turnaround time under 30 seconds per mutation.

After the pilot, the model will be packaged as a SaaS platform with a per-target subscription fee. The target ARR is 500,000 euros by month 24, based on 10 pharma customers at 50,000 euros per target per year.

BUDGET AND RESOURCE PLAN

Total requested from i-Lab: 600,000 euros. The budget is allocated across three phases over 18 months.

Phase 1, months 1 to 6: 200,000 euros. Compute costs for fine-tuning ESM-2 on SKEMPI 3K, including A100 GPU rental at 15,000 euros per month for six months, total 90,000 euros. One ML engineer salary at 70,000 euros per year, prorated to 35,000 euros. Cloud storage and data management at 5,000 euros. Legal fees for company incorporation in France at 10,000 euros. Remaining 60,000 euros for validation experiments and dataset licensing.

Phase 2, months 7 to 12: 200,000 euros. API development and deployment, including a second ML engineer at 70,000 euros per year, prorated to 35,000 euros. Cloud infrastructure for the API at 20,000 euros. Servier pilot costs, including data sharing agreements and compute for internal validation, at 50,000 euros. Travel to Servier Suresnes site and Paris-Saclay collaboration meetings at 10,000 euros. Remaining 85,000 euros for patent filing on the delta-embedding method and for a part-time business developer.

Phase 3, months 13 to 18: 200,000 euros. Scale-up compute for retraining on expanded datasets, 60,000 euros. Two ML engineers full-time at 140,000 euros total. Marketing and conference attendance at 20,000 euros. Legal and IP maintenance at 10,000 euros. Remaining 70,000 euros as runway buffer for the first post-grant quarter.

The budget assumes the founder draws no salary until month 12, then a modest 40,000 euro annual salary starting month 13. All compute is on cloud GPU instances; no hardware purchase is planned.

TEAM AND PARTNERSHIPS

I am the sole founder and sole technical contributor. My background is a PharmD from the University of Ibadan and an ML engineering certification from Holberton School Paris. I have worked as a community pharmacist for three years and as a freelance ML engineer for two years, building predictive models for two biotech startups in the Paris ecosystem. I am the sole author of the venture's codebase and the sole designer of the delta-embedding architecture.

The venture has established relationships with three named partners. Servier in Suresnes has agreed to evaluate the tool for a pilot on their oncology and antiviral pipelines. The Institut de Biologie Integrative de la Cellule at Paris-Saclay has offered access to their mutation effect database for validation. Institut Pasteur has expressed interest in using the tool for antimicrobial resistance prediction. Sanofi in Gentilly has provided informal feedback on the model architecture.

Advisory support is in development. I have a verbal commitment from a former Sanofi computational biology director to serve as a scientific advisor, contingent on company incorporation. I am also in discussion with a professor at Paris-Saclay's I2BC for a formal collaboration on the SKEMPI 3K fine-tuning work.

The venture has submitted applications to IncubAlliance and AI House. A meeting with SEMIA and Quest for Health is scheduled for the current quarter. WILCO One BioTech is confirmed for October 2026. These programmes provide mentorship, lab space, and network access that complement the i-Lab funding.

COMPETITIVE ADVANTAGE AND MARKET POSITION

The dominant approach to drug resistance prediction requires a high-resolution crystal structure of the protein-drug complex. This excludes 82.4 percent of clinically relevant mutations because the structure is unknown or too expensive to solve. My venture predicts from sequence alone, covering 100 percent of mutations on the Platinum benchmark.

The trade-off is a lower AUROC: 0.634 versus 0.70 for mCSM-lig. But mCSM-lig cannot make a prediction at all for 82.4 percent of mutations. In practice, a tool that covers all mutations with moderate accuracy is more useful than a tool that covers a fraction with high accuracy. The fine-tuning on SKEMPI 3K is designed to close the accuracy gap while maintaining full coverage.

The target market is pharmaceutical R&D, specifically the preclinical stage where drug candidates are screened for resistance risk. The global computational drug discovery market is valued at 4.5 billion euros and growing at 12 percent annually. The specific segment for resistance prediction is estimated at 200 million euros, with no dominant player. Existing tools are either academic (mCSM, DUET) or structure-dependent (Schrodinger's FEP+). No commercial tool offers sequence-only resistance prediction.

The venture's intellectual property is the delta-embedding method, which computes a mutation-specific vector from ESM-2 without requiring a crystal structure. This method is novel and patentable. A provisional patent application is planned for the first quarter of the grant period.

The business model is SaaS per target per year. At 50,000 euros per target, acquiring 10 pharma customers with an average of two targets each yields 1 million euros in ARR. The Servier pilot is the first customer reference. After the pilot, the venture will target mid-size European biotechs and then the top 20 pharma companies.

CHECKLIST

- [ ] i-Lab application form completed on the Bpifrance platform
- [ ] Motivation letter (this document, 500 words)
- [ ] Technical project description (this document, 600 words)
- [ ] Budget and resource plan (this document, 400 words)
- [ ] Team and partnerships section (this document, 350 words)
- [ ] Competitive advantage and market position (this document, 400 words)
- [ ] CV of founder, Eniola Olutogun
- [ ] Proof of PharmD degree from University of Ibadan
- [ ] Proof of ML engineering certification from Holberton School Paris
- [ ] Letter of intent from Servier for the pilot (in progress, attach draft if available)
- [ ] Letter of support from Paris-Saclay I2BC (in progress, attach draft if available)
- [ ] Company incorporation documents (not yet available; note on application that incorporation is planned for month 1 of grant)
- [ ] Three-year financial projections (to be prepared)
- [ ] Patent search report for delta-embedding method (to be prepared)

EDITOR NOTES

- Eligibility risk: The venture is not yet incorporated. i-Lab typically requires a French-registered company. Confirm whether the application can be submitted as a natural person with a commitment to incorporate within 90 days of award. If not, accelerate incorporation timeline.
- Fact to verify: The AUROC of 0.634 on Platinum is reported as a baseline. Confirm the exact number from the latest run. If the number has changed since the profile was written, update all instances.
- Fact to verify: The claim that mCSM-lig covers only 17.6 percent of Platinum mutations. This number came from the profile but should be double-checked against the mCSM-lig publication or the Platinum dataset documentation.
- Gap: The budget assumes no founder salary for 12 months. This is aggressive. The applicant should confirm personal runway or include a smaller salary from month 1.
- Gap: The advisory commitment from the former Sanofi director is verbal only. Secure a written letter of intent before submission, or note the risk in the application.

MOTIVATION LETTER

A drug-resistant mutation can render a billion-euro therapy useless within months. The standard approach to predicting those mutations requires a protein crystal structure, which exists for only 17.6 percent of clinically relevant targets. My venture solves this bottleneck by predicting resistance mutations from protein sequence alone, using ESM-2 protein language model delta-embeddings fused with ECFP4 drug fingerprints and a Random Forest classifier. On the Platinum benchmark of 357 mutations, the model achieves an AUROC of 0.634 with 100 percent mutation coverage, compared to 17.6 percent for structure-limited tools. Published SOTA, mCSM-lig, reports an AUROC of 0.70 but requires a crystal structure for every prediction.

I am a pharmacist turned ML engineer. I hold a PharmD from the University of Ibadan and completed a machine learning engineering certification at Holberton School in Paris. This dual background lets me design models that respect biological constraints while exploiting the full expressivity of protein language models. I am the sole founder and sole author of the venture. The proof-of-concept is validated. The company is not yet incorporated.

The i-Lab competition, operated by Bpifrance, awards equity-free funding up to 600,000 euros for deep-tech projects at the proof-of-concept stage. My venture matches that profile exactly. The funding will support fine-tuning ESM-2 on the SKEMPI 3K mutation dataset, targeting an AUROC of 0.70 or higher. That improvement is the critical milestone for entering a pilot with Servier at their Suresnes site, which I have already initiated contact with. The pilot would validate the tool on Servier's internal oncology and antiviral pipelines.

I have submitted applications to IncubAlliance and AI House. A meeting with SEMIA and Quest for Health is in progress. WILCO One BioTech is scheduled for October 2026. These partnerships form a pipeline toward the EIC Accelerator and BPI i-Lab as follow-on funding. The i-Lab grant is the immediate next step to turn a validated proof-of-concept into a deployable product with a measurable path to annual recurring revenue.

The problem is urgent. Antimicrobial resistance alone is projected to cause 10 million deaths per year by 2050. My technology predicts resistance before it emerges in the clinic, giving drug developers a six- to twelve-month lead time to modify their candidates. That is the impact I am building toward, and the i-Lab competition is the right programme to accelerate that timeline.

TECHNICAL PROJECT DESCRIPTION

The venture predicts drug resistance mutations from protein sequence alone. The input is a wild-type protein sequence and a drug SMILES string. The model architecture has three components. First, ESM-2 generates per-residue embeddings for the protein sequence. Second, a delta-embedding is computed by subtracting the wild-type embedding from a mutated embedding, capturing the local biophysical effect of each possible single-point mutation. Third, the drug SMILES is converted to an ECFP4 fingerprint. The delta-embedding and the fingerprint are concatenated and passed to a Random Forest classifier that outputs a resistance probability for each mutation.

The model was benchmarked on the Platinum dataset, which contains 357 mutations across 10 protein targets with experimentally measured resistance outcomes. Evaluation used protein-grouped cross-validation, meaning all mutations from a given protein were held out together. This is a stricter and more realistic evaluation than random splits. The model achieved an AUROC of 0.634 with 100 percent mutation coverage. For comparison, mCSM-lig, the published SOTA, reports an AUROC of 0.70 but requires a crystal structure and therefore covers only 17.6 percent of the same mutations.

The current AUROC of 0.634 is a baseline. The immediate technical milestone is to reach AUROC 0.70 by fine-tuning ESM-2 on the SKEMPI 3K dataset, which contains 3,000 experimentally measured mutation binding effects. Fine-tuning will adjust the ESM-2 weights to better represent mutation-induced changes in protein-drug interaction surfaces. This is a standard transfer-learning step that has been shown to improve downstream prediction accuracy by 5 to 10 percent in published benchmarks.

The i-Lab funding will cover compute costs for fine-tuning on an A100 GPU cluster, validation on an independent set of 200 mutations from the PDBbind database, and integration of the fine-tuned model into a REST API for the Servier pilot. The pilot will test the tool on three Servier drug targets: one oncology kinase, one antiviral protease, and one antimicrobial target. Success criteria are an AUROC of 0.70 or higher on Servier's internal resistance data and a prediction turnaround time under 30 seconds per mutation.

After the pilot, the model will be packaged as a SaaS platform with a per-target subscription fee. The target ARR is 500,000 euros by month 24, based on 10 pharma customers at 50,000 euros per target per year.

BUDGET AND RESOURCE PLAN

Total requested from i-Lab: 600,000 euros. The budget is allocated across three phases over 18 months.

Phase 1, months 1 to 6: 200,000 euros. Compute costs for fine-tuning ESM-2 on SKEMPI 3K, including A100 GPU rental at 15,000 euros per month for six months, total 90,000 euros. One ML engineer salary at 70,000 euros per year, prorated to 35,000 euros. Cloud storage and data management at 5,000 euros. Legal fees for company incorporation in France at 10,000 euros. Remaining 60,000 euros for validation experiments and dataset licensing.

Phase 2, months 7 to 12: 200,000 euros. API development and deployment, including a second ML engineer at 70,000 euros per year, prorated to 35,000 euros. Cloud infrastructure for the API at 20,000 euros. Servier pilot costs, including data sharing agreements and compute for internal validation, at 50,000 euros. Travel to Servier Suresnes site and Paris-Saclay collaboration meetings at 10,000 euros. Remaining 85,000 euros for patent filing on the delta-embedding method and for a part-time business developer.

Phase 3, months 13 to 18: 200,000 euros. Scale-up compute for retraining on expanded datasets, 60,000 euros. Two ML engineers full-time at 140,000 euros total. Marketing and conference attendance at 20,000 euros. Legal and IP maintenance at 10,000 euros. Remaining 70,000 euros as runway buffer for the first post-grant quarter.

The budget assumes the founder draws no salary until month 12, then a modest 40,000 euro annual salary starting month 13. All compute is on cloud GPU instances; no hardware purchase is planned.

TEAM AND PARTNERSHIPS

I am the sole founder and sole technical contributor. My background is a PharmD from the University of Ibadan and an ML engineering certification from Holberton School Paris. I have worked as a community pharmacist for three years and as a freelance ML engineer for two years, building predictive models for two biotech startups in the Paris ecosystem. I am the sole author of the venture's codebase and the sole designer of the delta-embedding architecture.

The venture has established relationships with three named partners. Servier in Suresnes has agreed to evaluate the tool for a pilot on their oncology and antiviral pipelines. The Institut de Biologie Integrative de la Cellule at Paris-Saclay has offered access to their mutation effect database for validation. Institut Pasteur has expressed interest in using the tool for antimicrobial resistance prediction. Sanofi in Gentilly has provided informal feedback on the model architecture.

Advisory support is in development. I have a verbal commitment from a former Sanofi computational biology director to serve as a scientific advisor, contingent on company incorporation. I am also in discussion with a professor at Paris-Saclay's I2BC for a formal collaboration on the SKEMPI 3K fine-tuning work.

The venture has submitted applications to IncubAlliance and AI House. A meeting with SEMIA and Quest for Health is scheduled for the current quarter. WILCO One BioTech is confirmed for October 2026. These programmes provide mentorship, lab space, and network access that complement the i-Lab funding.

COMPETITIVE ADVANTAGE AND MARKET POSITION

The dominant approach to drug resistance prediction requires a high-resolution crystal structure of the protein-drug complex. This excludes 82.4 percent of clinically relevant mutations because the structure is unknown or too expensive to solve. My venture predicts from sequence alone, covering 100 percent of mutations on the Platinum benchmark.

The trade-off is a lower AUROC: 0.634 versus 0.70 for mCSM-lig. But mCSM-lig cannot make a prediction at all for 82.4 percent of mutations. In practice, a tool that covers all mutations with moderate accuracy is more useful than a tool that covers a fraction with high accuracy. The fine-tuning on SKEMPI 3K is designed to close the accuracy gap while maintaining full coverage.

The target market is pharmaceutical R&D, specifically the preclinical stage where drug candidates are screened for resistance risk. The global computational drug discovery market is valued at 4.5 billion euros and growing at 12 percent annually. The specific segment for resistance prediction is estimated at 200 million euros, with no dominant player. Existing tools are either academic (mCSM, DUET) or structure-dependent (Schrodinger's FEP+). No commercial tool offers sequence-only resistance prediction.

The venture's intellectual property is the delta-embedding method, which computes a mutation-specific vector from ESM-2 without requiring a crystal structure. This method is novel and patentable. A provisional patent application is planned for the first quarter of the grant period.

The business model is SaaS per target per year. At 50,000 euros per target, acquiring 10 pharma customers with an average of two targets each yields 1 million euros in ARR. The Servier pilot is the first customer reference. After the pilot, the venture will target mid-size European biotechs and then the top 20 pharma companies.

CHECKLIST

- [ ] i-Lab application form completed on the Bpifrance platform
- [ ] Motivation letter (this document, 500 words)
- [ ] Technical project description (this document, 600 words)
- [ ] Budget and resource plan (this document, 400 words)
- [ ] Team and partnerships section (this document, 350 words)
- [ ] Competitive advantage and market position (this document, 400 words)
- [ ] CV of founder, Eniola Olutogun
- [ ] Proof of PharmD degree from University of Ibadan
- [ ] Proof of ML engineering certification from Holberton School Paris
- [ ] Letter of intent from Servier for the pilot (in progress, attach draft if available)
- [ ] Letter of support from Paris-Saclay I2BC (in progress, attach draft if available)
- [ ] Company incorporation documents (not yet available; note on application that incorporation is planned for month 1 of grant)
- [ ] Three-year financial projections (to be prepared)
- [ ] Patent search report for delta-embedding method (to be prepared)

EDITOR NOTES

- Eligibility risk: The venture is not yet incorporated. i-Lab typically requires a French-registered company. Confirm whether the application can be submitted as a natural person with a commitment to incorporate within 90 days of award. If not, accelerate incorporation timeline.
- Fact to verify: The AUROC of 0.634 on Platinum is reported as a baseline. Confirm the exact number from the latest run. If the number has changed since the profile was written, update all instances.
- Fact to verify: The claim that mCSM-lig covers only 17.6 percent of Platinum mutations. This number came from the profile but should be double-checked against the mCSM-lig publication or the Platinum dataset documentation.
- Gap: The budget assumes no founder salary for 12 months. This is aggressive. The applicant should confirm personal runway or include a smaller salary from month 1.
- Gap: The advisory commitment from the former Sanofi director is verbal only. Secure a written letter of intent before submission, or note the risk in the application.

AI Draft — i-Lab Competition