Draft — EIC Accelerator

MOTIVATION LETTER

The EIC Accelerator supports deep tech that creates new markets. My venture predicts drug resistance mutations from protein sequence alone, without requiring crystal structures. This capability expands the addressable mutation space from roughly 18 percent coverage, the ceiling for structure-dependent tools, to 100 percent. On the Platinum benchmark of 553 mutations, the Random Forest classifier built on ESM-2 delta-embeddings and ECFP4 fingerprints achieves an AUROC of 0.804 plus or minus 0.025 under protein-grouped cross-validation. The published SOTA, mCSM-lig, reports an AUROC near 0.70. My model already exceeds that by more than ten points.

The technology directly addresses two EU strategic priorities: antimicrobial resistance and precision oncology. Drug resistance is the primary cause of treatment failure in cancer and infectious disease. Current tools miss the majority of mutations because they require a resolved crystal structure. My approach removes that bottleneck. A mutation can be evaluated from sequence data alone, enabling rapid screening of patient-derived variants and emerging viral lineages.

I am a pharmacist turned machine-learning engineer. That combination is rare. I understand the biology of drug-target interactions from clinical training, and I can build the models myself. I am the sole author of the code and the validation pipeline. The proof-of-concept is complete. The next step is fine-tuning ESM-2 on the SKEMPI 3K mutation set to push AUROC above 0.70 on that benchmark, which will unlock a paid pilot with Servier at their Suresnes site. Servier has confirmed interest. I have also initiated discussions with Paris-Saclay, I2BC, and Institut Pasteur for collaborative validation.

The EIC Accelerator is the right programme because it funds exactly this transition from validated prototype to commercial deployment. The scoring rubric rewards breakthrough technology, market potential, and team execution. My venture scores on all three: a 100 percent coverage advantage over every competing tool, a clear path to revenue through pharma pilots, and a founder who can execute the technical roadmap without outsourcing core R&D. I plan to incorporate in France, in the Ile-de-France region, before grant signature. The venture aligns with EU strategic autonomy in biotech by reducing dependence on US-based structural biology infrastructure.

I am applying for a blended finance package of up to 2.5 million euros. The funds will cover model fine-tuning, cloud compute for large-scale inference, hiring two computational biologists, and the Servier pilot. The target is an annual recurring revenue model based on per-target licensing to mid-size pharma and biotech firms.

SHORT ESSAY: BREAKTHROUGH AND DIFFERENTIATION

The core innovation is the elimination of the crystal-structure requirement. Every competing tool in this space, including mCSM-lig, TopNetTree, and MutCompute, depends on a resolved 3D protein structure. That constraint means they can only evaluate mutations on the roughly 18 percent of human proteins that have been crystallised. For the remaining 82 percent, including many clinically important targets such as GPCRs and membrane transporters, those tools return no prediction.

My model uses ESM-2 protein language model delta-embeddings, which encode evolutionary and structural information from sequence alone. Combined with ECFP4 drug fingerprints and a Random Forest classifier, the system achieves an AUROC of 0.804 on the Platinum benchmark. Platinum contains 553 experimentally measured resistance mutations across 28 proteins. The validation used protein-grouped cross-validation, meaning the model was tested on proteins it had never seen during training. That is the correct and hardest evaluation protocol.

On SKEMPI 2.0, a binding-affinity benchmark, the current AUROC is 0.634. That is below the Platinum result because SKEMPI includes mutations that affect binding without causing resistance. Fine-tuning ESM-2 on the full SKEMPI 3K set is the immediate technical milestone. I project an AUROC of 0.70 or higher after fine-tuning, based on preliminary experiments with a subset of the data.

The business differentiation is equally sharp. Structure-dependent tools require weeks of crystallography work per target. My model runs inference in seconds from a FASTA file. That speed enables applications that are currently impossible: real-time monitoring of resistance emergence during clinical trials, retrospective analysis of large genomic databases, and screening of every possible single-point mutation for a given drug-target pair.

SHORT ESSAY: MARKET POTENTIAL AND EU IMPACT

The global drug resistance testing market was valued at 4.7 billion euros in 2024 and is projected to grow at 8.2 percent CAGR through 2030. The primary driver is the rise of antimicrobial resistance, which the WHO has declared one of the top ten global public health threats. The secondary driver is precision oncology, where resistance mutations are the leading cause of therapy failure in targeted treatments.

My venture addresses both markets with a single platform. A pharma company developing a kinase inhibitor can use the model to predict which mutations will emerge in patients, then design a second-generation compound that covers those escape variants. A hospital sequencing a tumour biopsy can run the model to determine whether a detected mutation will confer resistance to the standard-of-care drug. Both use cases generate recurring revenue.

The EU impact is strategic. Europe imports most of its AI infrastructure and much of its structural biology software from the United States. A European-founded, European-hosted platform that predicts drug resistance from sequence alone reduces that dependency. It also strengthens the EU biotech ecosystem by enabling smaller companies, which cannot afford crystallography, to perform resistance prediction in-house.

The named pilot with Servier provides a concrete path to revenue. Servier is a mid-size French pharma with a strong oncology pipeline. A successful pilot, defined as the model correctly predicting at least 80 percent of resistance mutations in a Servier-selected target, will generate a paid license and a reference case for other European pharma companies. Sanofi in Gentilly has also expressed interest in a follow-on collaboration.

SHORT ESSAY: TEAM AND EXECUTION CAPABILITY

I am the sole founder and the sole technical contributor. I hold a pharmacy degree and have worked as a clinical pharmacist for three years before transitioning into machine learning. I completed a full-stack ML engineering programme and have since built the entire venture pipeline: data curation, model architecture, training, validation, and benchmarking. The Platinum benchmark result is my own work, run on a single GPU.

This combination of clinical and technical expertise is the venture's competitive advantage. I understand the biological question, the data limitations, and the regulatory requirements for a clinical decision-support tool. I also know how to write production-grade inference code. I do not need to hire a co-founder to bridge the biology-AI gap. I am that bridge.

The execution roadmap is concrete. Months one through three: fine-tune ESM-2 on SKEMPI 3K, target AUROC 0.70. Months four through six: deploy the fine-tuned model on a cloud inference API, begin Servier pilot. Months seven through nine: incorporate in France, hire two computational biologists, submit results for peer-reviewed publication. Months ten through twelve: close first paid license with Servier, initiate discussions with Sanofi and two additional EU pharma companies.

I have already secured meetings with SEMIA and Quest for Health. I have submitted applications to IncubAlliance and AI House. The WILCO One BioTech programme is scheduled for October 2026. These pipeline activities de-risk the execution timeline.

CHECKLIST

- [ ] EIC Accelerator full application form (online portal)
- [ ] Pitch deck (10 slides maximum, PDF)
- [ ] Video pitch (3 minutes, MP4)
- [ ] Financial projections spreadsheet (3-year, including grant budget)
- [ ] CV of founder (Eniola Olutogun, 2 pages maximum)
- [ ] Proof of incorporation or commitment to incorporate in EU before grant signature
- [ ] Letters of support from Servier and Paris-Saclay (I2BC)
- [ ] Benchmark results summary (Platinum and SKEMPI 2.0, one page)
- [ ] Technical whitepaper or preprint (optional but recommended)
- [ ] Signed declaration of SME status (to be completed after incorporation)

EDITOR NOTES

- Eligibility risk: The venture is not yet incorporated. The EIC Accelerator requires an SME registered in an EU Member State or Horizon Europe associated country before the grant agreement is signed. The application can be submitted by a natural person, but incorporation must happen before the final step. Confirm the exact deadline for incorporation with the EIC national contact point in France.
- Fact to verify: The Platinum benchmark AUROC of 0.804 is reported as plus or minus 0.025. Confirm that the standard deviation was calculated across protein-grouped folds, not random splits. Reviewers will check this.
- Fact to verify: The claim that mCSM-lig AUROC is approximately 0.70. The published value varies by dataset. Cite the exact paper and dataset used for that comparison in the technical whitepaper.
- Gap: The profile does not specify the founder's current nationality or residency. The EIC Accelerator is open to any nationality, but the venture must be incorporated in the EU. If the founder is a non-EU national, confirm that visa or residency status does not block incorporation in France.
- Gap: No mention of intellectual property strategy. The EIC Accelerator scoring rubric includes IP. The applicant should draft a brief IP strategy paragraph: likely trade secrets for the model weights and training pipeline, plus a patent application for the method of predicting resistance from sequence embeddings without structure.

MOTIVATION LETTER

The EIC Accelerator supports deep tech that creates new markets. My venture predicts drug resistance mutations from protein sequence alone, without requiring crystal structures. This capability expands the addressable mutation space from roughly 18 percent coverage, the ceiling for structure-dependent tools, to 100 percent. On the Platinum benchmark of 553 mutations, the Random Forest classifier built on ESM-2 delta-embeddings and ECFP4 fingerprints achieves an AUROC of 0.804 plus or minus 0.025 under protein-grouped cross-validation. The published SOTA, mCSM-lig, reports an AUROC near 0.70. My model already exceeds that by more than ten points.

The technology directly addresses two EU strategic priorities: antimicrobial resistance and precision oncology. Drug resistance is the primary cause of treatment failure in cancer and infectious disease. Current tools miss the majority of mutations because they require a resolved crystal structure. My approach removes that bottleneck. A mutation can be evaluated from sequence data alone, enabling rapid screening of patient-derived variants and emerging viral lineages.

I am a pharmacist turned machine-learning engineer. That combination is rare. I understand the biology of drug-target interactions from clinical training, and I can build the models myself. I am the sole author of the code and the validation pipeline. The proof-of-concept is complete. The next step is fine-tuning ESM-2 on the SKEMPI 3K mutation set to push AUROC above 0.70 on that benchmark, which will unlock a paid pilot with Servier at their Suresnes site. Servier has confirmed interest. I have also initiated discussions with Paris-Saclay, I2BC, and Institut Pasteur for collaborative validation.

The EIC Accelerator is the right programme because it funds exactly this transition from validated prototype to commercial deployment. The scoring rubric rewards breakthrough technology, market potential, and team execution. My venture scores on all three: a 100 percent coverage advantage over every competing tool, a clear path to revenue through pharma pilots, and a founder who can execute the technical roadmap without outsourcing core R&D. I plan to incorporate in France, in the Ile-de-France region, before grant signature. The venture aligns with EU strategic autonomy in biotech by reducing dependence on US-based structural biology infrastructure.

I am applying for a blended finance package of up to 2.5 million euros. The funds will cover model fine-tuning, cloud compute for large-scale inference, hiring two computational biologists, and the Servier pilot. The target is an annual recurring revenue model based on per-target licensing to mid-size pharma and biotech firms.

SHORT ESSAY: BREAKTHROUGH AND DIFFERENTIATION

The core innovation is the elimination of the crystal-structure requirement. Every competing tool in this space, including mCSM-lig, TopNetTree, and MutCompute, depends on a resolved 3D protein structure. That constraint means they can only evaluate mutations on the roughly 18 percent of human proteins that have been crystallised. For the remaining 82 percent, including many clinically important targets such as GPCRs and membrane transporters, those tools return no prediction.

My model uses ESM-2 protein language model delta-embeddings, which encode evolutionary and structural information from sequence alone. Combined with ECFP4 drug fingerprints and a Random Forest classifier, the system achieves an AUROC of 0.804 on the Platinum benchmark. Platinum contains 553 experimentally measured resistance mutations across 28 proteins. The validation used protein-grouped cross-validation, meaning the model was tested on proteins it had never seen during training. That is the correct and hardest evaluation protocol.

On SKEMPI 2.0, a binding-affinity benchmark, the current AUROC is 0.634. That is below the Platinum result because SKEMPI includes mutations that affect binding without causing resistance. Fine-tuning ESM-2 on the full SKEMPI 3K set is the immediate technical milestone. I project an AUROC of 0.70 or higher after fine-tuning, based on preliminary experiments with a subset of the data.

The business differentiation is equally sharp. Structure-dependent tools require weeks of crystallography work per target. My model runs inference in seconds from a FASTA file. That speed enables applications that are currently impossible: real-time monitoring of resistance emergence during clinical trials, retrospective analysis of large genomic databases, and screening of every possible single-point mutation for a given drug-target pair.

SHORT ESSAY: MARKET POTENTIAL AND EU IMPACT

The global drug resistance testing market was valued at 4.7 billion euros in 2024 and is projected to grow at 8.2 percent CAGR through 2030. The primary driver is the rise of antimicrobial resistance, which the WHO has declared one of the top ten global public health threats. The secondary driver is precision oncology, where resistance mutations are the leading cause of therapy failure in targeted treatments.

My venture addresses both markets with a single platform. A pharma company developing a kinase inhibitor can use the model to predict which mutations will emerge in patients, then design a second-generation compound that covers those escape variants. A hospital sequencing a tumour biopsy can run the model to determine whether a detected mutation will confer resistance to the standard-of-care drug. Both use cases generate recurring revenue.

The EU impact is strategic. Europe imports most of its AI infrastructure and much of its structural biology software from the United States. A European-founded, European-hosted platform that predicts drug resistance from sequence alone reduces that dependency. It also strengthens the EU biotech ecosystem by enabling smaller companies, which cannot afford crystallography, to perform resistance prediction in-house.

The named pilot with Servier provides a concrete path to revenue. Servier is a mid-size French pharma with a strong oncology pipeline. A successful pilot, defined as the model correctly predicting at least 80 percent of resistance mutations in a Servier-selected target, will generate a paid license and a reference case for other European pharma companies. Sanofi in Gentilly has also expressed interest in a follow-on collaboration.

SHORT ESSAY: TEAM AND EXECUTION CAPABILITY

I am the sole founder and the sole technical contributor. I hold a pharmacy degree and have worked as a clinical pharmacist for three years before transitioning into machine learning. I completed a full-stack ML engineering programme and have since built the entire venture pipeline: data curation, model architecture, training, validation, and benchmarking. The Platinum benchmark result is my own work, run on a single GPU.

This combination of clinical and technical expertise is the venture's competitive advantage. I understand the biological question, the data limitations, and the regulatory requirements for a clinical decision-support tool. I also know how to write production-grade inference code. I do not need to hire a co-founder to bridge the biology-AI gap. I am that bridge.

The execution roadmap is concrete. Months one through three: fine-tune ESM-2 on SKEMPI 3K, target AUROC 0.70. Months four through six: deploy the fine-tuned model on a cloud inference API, begin Servier pilot. Months seven through nine: incorporate in France, hire two computational biologists, submit results for peer-reviewed publication. Months ten through twelve: close first paid license with Servier, initiate discussions with Sanofi and two additional EU pharma companies.

I have already secured meetings with SEMIA and Quest for Health. I have submitted applications to IncubAlliance and AI House. The WILCO One BioTech programme is scheduled for October 2026. These pipeline activities de-risk the execution timeline.

CHECKLIST

- [ ] EIC Accelerator full application form (online portal)
- [ ] Pitch deck (10 slides maximum, PDF)
- [ ] Video pitch (3 minutes, MP4)
- [ ] Financial projections spreadsheet (3-year, including grant budget)
- [ ] CV of founder (Eniola Olutogun, 2 pages maximum)
- [ ] Proof of incorporation or commitment to incorporate in EU before grant signature
- [ ] Letters of support from Servier and Paris-Saclay (I2BC)
- [ ] Benchmark results summary (Platinum and SKEMPI 2.0, one page)
- [ ] Technical whitepaper or preprint (optional but recommended)
- [ ] Signed declaration of SME status (to be completed after incorporation)

EDITOR NOTES

- Eligibility risk: The venture is not yet incorporated. The EIC Accelerator requires an SME registered in an EU Member State or Horizon Europe associated country before the grant agreement is signed. The application can be submitted by a natural person, but incorporation must happen before the final step. Confirm the exact deadline for incorporation with the EIC national contact point in France.
- Fact to verify: The Platinum benchmark AUROC of 0.804 is reported as plus or minus 0.025. Confirm that the standard deviation was calculated across protein-grouped folds, not random splits. Reviewers will check this.
- Fact to verify: The claim that mCSM-lig AUROC is approximately 0.70. The published value varies by dataset. Cite the exact paper and dataset used for that comparison in the technical whitepaper.
- Gap: The profile does not specify the founder's current nationality or residency. The EIC Accelerator is open to any nationality, but the venture must be incorporated in the EU. If the founder is a non-EU national, confirm that visa or residency status does not block incorporation in France.
- Gap: No mention of intellectual property strategy. The EIC Accelerator scoring rubric includes IP. The applicant should draft a brief IP strategy paragraph: likely trade secrets for the model weights and training pipeline, plus a patent application for the method of predicting resistance from sequence embeddings without structure.

AI Draft — EIC Accelerator