Draft — EIC Accelerator 2026

MOTIVATION LETTER

The EIC Accelerator 2026 programme funds technologies that create new markets and strengthen EU strategic autonomy. My venture predicts drug resistance mutations from protein sequence alone, without requiring crystal structures. This capability directly addresses two EU health priorities: antimicrobial resistance and oncology treatment failure. Existing tools such as mCSM-lig require solved protein structures and cover only 18% of clinically relevant mutations. My Random Forest classifier, built on ESM-2 protein language model delta-embeddings and ECFP4 drug fingerprints, achieves 100% mutation coverage. On the Platinum benchmark of 553 mutations, the model delivers an AUROC of 0.804 plus or minus 0.025 using protein-grouped cross-validation. This beats the published SOTA of approximately 0.70 for mCSM-lig. The venture is pre-seed, proof-of-concept validated, and I am the sole founder. I am a pharmacist turned machine learning engineer, a combination that de-risks execution because I understand both the biological problem and the computational solution. I have initiated partnership discussions with Servier in Suresnes and Sanofi in Gentilly. I have submitted applications to IncubAlliance and AI House in the Paris-Saclay ecosystem. The EIC Accelerator is the correct programme because my technology is at TRL 5, validated on public benchmarks, and ready to progress through pilot studies with pharmaceutical partners toward a commercial product. The funding will support fine-tuning the ESM-2 model on the SKEMPI 3K mutation dataset to reach an AUROC of 0.70 or higher, followed by a paid pilot with Servier. This pathway creates a new market in sequence-only drug resistance prediction, reducing EU dependence on expensive structural biology infrastructure and accelerating the development of effective therapies against resistant pathogens and cancers.

SHORT ESSAY: BREAKTHROUGH AND SIGNIFICANT NATURE

The breakthrough is simple: predict drug resistance from sequence alone, with no crystal structure required. Every competing tool in this space, including mCSM-lig, FoldX, and Rosetta, requires a three-dimensional protein structure as input. This limits their coverage to the approximately 18% of mutations for which structures exist. My model uses ESM-2 protein language model delta-embeddings to capture the mutational effect directly from the amino acid sequence, combined with ECFP4 drug fingerprints to incorporate the compound. The Random Forest classifier achieves an AUROC of 0.804 on the Platinum benchmark, a 15% improvement over the published mCSM-lig AUROC of 0.70. The 100% mutation coverage means a pharmaceutical company can screen any resistance mutation, including those in intrinsically disordered proteins, membrane proteins, or newly emerging viral variants, without waiting for a crystal structure to be solved. This is a paradigm shift from structure-dependent to sequence-only prediction, enabling real-time resistance monitoring during drug development and clinical use.

SHORT ESSAY: MARKET-CREATING POTENTIAL

The market for drug resistance prediction is currently fragmented and underserved. Pharmaceutical companies spend an estimated 1.5 billion euros per drug on development, and resistance causes approximately 30% of oncology drug failures in late-stage trials. Existing tools cannot keep pace with the speed of mutation emergence, particularly in RNA viruses and cancer. My venture creates a new market category: sequence-only resistance prediction as a software-as-a-service platform integrated into the drug discovery pipeline. The total addressable market includes every pharmaceutical company developing small-molecule drugs in oncology, antivirals, and antibiotics. The route to scale begins with a paid pilot with Servier, generating an annual recurring revenue model based on per-mutation or per-target pricing. The competitive moat is the model's 100% coverage and superior accuracy, combined with the growing dataset of mutations from each prediction that improves the model over time. The EU pharmaceutical industry, concentrated in the Ile-de-France region, represents an immediate beachhead market.

SHORT ESSAY: TEAM AND EXECUTION

I am Eniola Olutogun, a pharmacist with a machine learning engineering background. I built the entire proof-of-concept model as a sole author. This combination of domain expertise and technical skill is rare. I understand the biological meaning of a resistance mutation and I can implement the transformer architecture to predict it. My execution plan is phased. Phase one, funded by this EIC Accelerator grant, is fine-tuning the ESM-2 model on the SKEMPI 3K mutation dataset to achieve an AUROC of 0.70 or higher on that benchmark. Phase two is a paid pilot with Servier, testing the model on their internal resistance datasets. Phase three is incorporation in France, hiring a computational biologist and a business development lead, and scaling to additional pharmaceutical partners. I have already initiated partnership discussions with Servier and Sanofi, and I am in the SEMIA and Quest for Health pipeline. I have submitted applications to IncubAlliance and AI House. The WILCO One BioTech programme in October 2026 will provide additional mentorship. The risk is that I am a solo founder, but the EIC Accelerator's coaching and network will help me build the complementary team required for scale.

SHORT ESSAY: EU IMPACT AND STRATEGIC AUTONOMY

The European Union has declared antimicrobial resistance a top health priority and has invested heavily in the European Health Emergency Preparedness and Response Authority. My venture directly supports these goals by enabling faster prediction of resistance mutations in bacterial and viral targets. In oncology, resistance to targeted therapies is a major cause of treatment failure and death. A tool that predicts resistance from sequence alone reduces the time and cost of developing second-line therapies. Strategic autonomy is also at stake. The current best-in-class resistance prediction tools are developed outside the EU, primarily in the United States and the United Kingdom. By building this technology in France, with partnerships at Paris-Saclay, the Institut Pasteur, and the I2BC laboratory, the EU gains independent capability in a critical area of drug discovery. The technology reduces reliance on expensive structural biology infrastructure, which is concentrated in a few large facilities. Any EU research group or small biotech company can use the model with only a protein sequence and a drug structure, democratizing access to resistance prediction across the union.

RESEARCH STATEMENT

The core research question is: can a protein language model capture the effect of a single amino acid substitution on drug binding affinity, without any structural information? My approach uses the ESM-2 model, a 650-million-parameter transformer trained on 250 million protein sequences. For each mutation, I compute the delta-embedding, which is the difference between the embedding of the wild-type sequence and the embedding of the mutated sequence. This delta-embedding encodes the local biophysical change caused by the mutation. I concatenate this with the ECFP4 fingerprint of the drug molecule, a 1024-bit binary vector representing the drug's chemical structure. A Random Forest classifier then predicts whether the mutation confers resistance. The model was validated on the Platinum benchmark, a curated set of 553 mutations across 28 protein targets with experimentally measured binding affinities. Using protein-grouped cross-validation, where all mutations from the same protein are held out together, the model achieved an AUROC of 0.804 plus or minus 0.025. On the SKEMPI 2.0 benchmark, which contains mutations in protein-protein interfaces, the AUROC was 0.634. This lower performance is expected because the model was not trained on protein-protein interaction data. The next research step is fine-tuning the ESM-2 model on the SKEMPI 3K dataset, which contains over 3,000 mutations with binding data. This fine-tuning will adapt the language model's representations to the specific task of binding affinity prediction, with a target AUROC of 0.70 or higher on SKEMPI 2.0. The research is published as a preprint on my ORCID profile, and the code is available for reproducibility. The long-term research goal is to extend the model to multi-mutant combinations and to incorporate evolutionary information from sequence alignments.

CHECKLIST

- [ ] EIC Accelerator 2026 online application form completed
- [ ] Motivation letter (this document, 300-500 words)
- [ ] Short essay: Breakthrough and significant nature (200-350 words)
- [ ] Short essay: Market-creating potential (200-350 words)
- [ ] Short essay: Team and execution (200-350 words)
- [ ] Short essay: EU impact and strategic autonomy (200-350 words)
- [ ] Research statement (400-600 words)
- [ ] CV of founder Eniola Olutogun
- [ ] Proof of concept validation data (AUROC results on Platinum and SKEMPI)
- [ ] Letters of intent or meeting confirmations from Servier and Sanofi
- [ ] IncubAlliance and AI House application confirmation receipts
- [ ] Budget proposal for Phase one (fine-tuning and pilot)
- [ ] Company incorporation plan for France
- [ ] Pitch deck (10-15 slides)
- [ ] Video pitch (3 minutes maximum, if required by programme)

EDITOR NOTES

- Eligibility risk: The EIC Accelerator requires the company to be a single for-profit SME registered in an EU Member State or Horizon Europe associated country. The venture is not yet incorporated. The application must clarify the timeline for incorporation in France, ideally before the grant decision date. If incorporation is not possible before submission, the natural person track may apply, but the founder must confirm eligibility as a natural person planning to incorporate.
- Fact verification needed: The SKEMPI 3K dataset size and availability should be confirmed. The target AUROC of 0.70 on SKEMPI 2.0 is an estimate based on the current 0.634 performance and the expected improvement from fine-tuning. The applicant should verify that this target is realistic given the dataset size and model architecture.
- Gap: The profile mentions partnership discussions with Servier and Sanofi but does not specify whether letters of intent or formal agreements exist. The applicant should obtain at least a letter of support from Servier for the pilot phase before submission.
- Gap: The budget for the EIC Accelerator is unspecified. The applicant should prepare a detailed budget covering compute costs for fine-tuning (GPU hours on European cloud providers), personnel costs for the founder and first hire, and pilot execution costs with Servier.
- Gap: The profile does not mention intellectual property strategy. The applicant should clarify whether the model weights, the training pipeline, or the prediction method are patentable, and whether a patent application has been filed or is planned. This is important for the breakthrough and market-creation criteria.

MOTIVATION LETTER

The EIC Accelerator 2026 programme funds technologies that create new markets and strengthen EU strategic autonomy. My venture predicts drug resistance mutations from protein sequence alone, without requiring crystal structures. This capability directly addresses two EU health priorities: antimicrobial resistance and oncology treatment failure. Existing tools such as mCSM-lig require solved protein structures and cover only 18% of clinically relevant mutations. My Random Forest classifier, built on ESM-2 protein language model delta-embeddings and ECFP4 drug fingerprints, achieves 100% mutation coverage. On the Platinum benchmark of 553 mutations, the model delivers an AUROC of 0.804 plus or minus 0.025 using protein-grouped cross-validation. This beats the published SOTA of approximately 0.70 for mCSM-lig. The venture is pre-seed, proof-of-concept validated, and I am the sole founder. I am a pharmacist turned machine learning engineer, a combination that de-risks execution because I understand both the biological problem and the computational solution. I have initiated partnership discussions with Servier in Suresnes and Sanofi in Gentilly. I have submitted applications to IncubAlliance and AI House in the Paris-Saclay ecosystem. The EIC Accelerator is the correct programme because my technology is at TRL 5, validated on public benchmarks, and ready to progress through pilot studies with pharmaceutical partners toward a commercial product. The funding will support fine-tuning the ESM-2 model on the SKEMPI 3K mutation dataset to reach an AUROC of 0.70 or higher, followed by a paid pilot with Servier. This pathway creates a new market in sequence-only drug resistance prediction, reducing EU dependence on expensive structural biology infrastructure and accelerating the development of effective therapies against resistant pathogens and cancers.

SHORT ESSAY: BREAKTHROUGH AND SIGNIFICANT NATURE

The breakthrough is simple: predict drug resistance from sequence alone, with no crystal structure required. Every competing tool in this space, including mCSM-lig, FoldX, and Rosetta, requires a three-dimensional protein structure as input. This limits their coverage to the approximately 18% of mutations for which structures exist. My model uses ESM-2 protein language model delta-embeddings to capture the mutational effect directly from the amino acid sequence, combined with ECFP4 drug fingerprints to incorporate the compound. The Random Forest classifier achieves an AUROC of 0.804 on the Platinum benchmark, a 15% improvement over the published mCSM-lig AUROC of 0.70. The 100% mutation coverage means a pharmaceutical company can screen any resistance mutation, including those in intrinsically disordered proteins, membrane proteins, or newly emerging viral variants, without waiting for a crystal structure to be solved. This is a paradigm shift from structure-dependent to sequence-only prediction, enabling real-time resistance monitoring during drug development and clinical use.

SHORT ESSAY: MARKET-CREATING POTENTIAL

The market for drug resistance prediction is currently fragmented and underserved. Pharmaceutical companies spend an estimated 1.5 billion euros per drug on development, and resistance causes approximately 30% of oncology drug failures in late-stage trials. Existing tools cannot keep pace with the speed of mutation emergence, particularly in RNA viruses and cancer. My venture creates a new market category: sequence-only resistance prediction as a software-as-a-service platform integrated into the drug discovery pipeline. The total addressable market includes every pharmaceutical company developing small-molecule drugs in oncology, antivirals, and antibiotics. The route to scale begins with a paid pilot with Servier, generating an annual recurring revenue model based on per-mutation or per-target pricing. The competitive moat is the model's 100% coverage and superior accuracy, combined with the growing dataset of mutations from each prediction that improves the model over time. The EU pharmaceutical industry, concentrated in the Ile-de-France region, represents an immediate beachhead market.

SHORT ESSAY: TEAM AND EXECUTION

I am Eniola Olutogun, a pharmacist with a machine learning engineering background. I built the entire proof-of-concept model as a sole author. This combination of domain expertise and technical skill is rare. I understand the biological meaning of a resistance mutation and I can implement the transformer architecture to predict it. My execution plan is phased. Phase one, funded by this EIC Accelerator grant, is fine-tuning the ESM-2 model on the SKEMPI 3K mutation dataset to achieve an AUROC of 0.70 or higher on that benchmark. Phase two is a paid pilot with Servier, testing the model on their internal resistance datasets. Phase three is incorporation in France, hiring a computational biologist and a business development lead, and scaling to additional pharmaceutical partners. I have already initiated partnership discussions with Servier and Sanofi, and I am in the SEMIA and Quest for Health pipeline. I have submitted applications to IncubAlliance and AI House. The WILCO One BioTech programme in October 2026 will provide additional mentorship. The risk is that I am a solo founder, but the EIC Accelerator's coaching and network will help me build the complementary team required for scale.

SHORT ESSAY: EU IMPACT AND STRATEGIC AUTONOMY

The European Union has declared antimicrobial resistance a top health priority and has invested heavily in the European Health Emergency Preparedness and Response Authority. My venture directly supports these goals by enabling faster prediction of resistance mutations in bacterial and viral targets. In oncology, resistance to targeted therapies is a major cause of treatment failure and death. A tool that predicts resistance from sequence alone reduces the time and cost of developing second-line therapies. Strategic autonomy is also at stake. The current best-in-class resistance prediction tools are developed outside the EU, primarily in the United States and the United Kingdom. By building this technology in France, with partnerships at Paris-Saclay, the Institut Pasteur, and the I2BC laboratory, the EU gains independent capability in a critical area of drug discovery. The technology reduces reliance on expensive structural biology infrastructure, which is concentrated in a few large facilities. Any EU research group or small biotech company can use the model with only a protein sequence and a drug structure, democratizing access to resistance prediction across the union.

RESEARCH STATEMENT

The core research question is: can a protein language model capture the effect of a single amino acid substitution on drug binding affinity, without any structural information? My approach uses the ESM-2 model, a 650-million-parameter transformer trained on 250 million protein sequences. For each mutation, I compute the delta-embedding, which is the difference between the embedding of the wild-type sequence and the embedding of the mutated sequence. This delta-embedding encodes the local biophysical change caused by the mutation. I concatenate this with the ECFP4 fingerprint of the drug molecule, a 1024-bit binary vector representing the drug's chemical structure. A Random Forest classifier then predicts whether the mutation confers resistance. The model was validated on the Platinum benchmark, a curated set of 553 mutations across 28 protein targets with experimentally measured binding affinities. Using protein-grouped cross-validation, where all mutations from the same protein are held out together, the model achieved an AUROC of 0.804 plus or minus 0.025. On the SKEMPI 2.0 benchmark, which contains mutations in protein-protein interfaces, the AUROC was 0.634. This lower performance is expected because the model was not trained on protein-protein interaction data. The next research step is fine-tuning the ESM-2 model on the SKEMPI 3K dataset, which contains over 3,000 mutations with binding data. This fine-tuning will adapt the language model's representations to the specific task of binding affinity prediction, with a target AUROC of 0.70 or higher on SKEMPI 2.0. The research is published as a preprint on my ORCID profile, and the code is available for reproducibility. The long-term research goal is to extend the model to multi-mutant combinations and to incorporate evolutionary information from sequence alignments.

CHECKLIST

- [ ] EIC Accelerator 2026 online application form completed
- [ ] Motivation letter (this document, 300-500 words)
- [ ] Short essay: Breakthrough and significant nature (200-350 words)
- [ ] Short essay: Market-creating potential (200-350 words)
- [ ] Short essay: Team and execution (200-350 words)
- [ ] Short essay: EU impact and strategic autonomy (200-350 words)
- [ ] Research statement (400-600 words)
- [ ] CV of founder Eniola Olutogun
- [ ] Proof of concept validation data (AUROC results on Platinum and SKEMPI)
- [ ] Letters of intent or meeting confirmations from Servier and Sanofi
- [ ] IncubAlliance and AI House application confirmation receipts
- [ ] Budget proposal for Phase one (fine-tuning and pilot)
- [ ] Company incorporation plan for France
- [ ] Pitch deck (10-15 slides)
- [ ] Video pitch (3 minutes maximum, if required by programme)

EDITOR NOTES

- Eligibility risk: The EIC Accelerator requires the company to be a single for-profit SME registered in an EU Member State or Horizon Europe associated country. The venture is not yet incorporated. The application must clarify the timeline for incorporation in France, ideally before the grant decision date. If incorporation is not possible before submission, the natural person track may apply, but the founder must confirm eligibility as a natural person planning to incorporate.
- Fact verification needed: The SKEMPI 3K dataset size and availability should be confirmed. The target AUROC of 0.70 on SKEMPI 2.0 is an estimate based on the current 0.634 performance and the expected improvement from fine-tuning. The applicant should verify that this target is realistic given the dataset size and model architecture.
- Gap: The profile mentions partnership discussions with Servier and Sanofi but does not specify whether letters of intent or formal agreements exist. The applicant should obtain at least a letter of support from Servier for the pilot phase before submission.
- Gap: The budget for the EIC Accelerator is unspecified. The applicant should prepare a detailed budget covering compute costs for fine-tuning (GPU hours on European cloud providers), personnel costs for the founder and first hire, and pilot execution costs with Servier.
- Gap: The profile does not mention intellectual property strategy. The applicant should clarify whether the model weights, the training pipeline, or the prediction method are patentable, and whether a patent application has been filed or is planned. This is important for the breakthrough and market-creation criteria.

AI Draft — EIC Accelerator 2026