Draft — MEST AI Startup Programme

MOTIVATION LETTER

Drug resistance is a present, measurable crisis. In 2019, bacterial antimicrobial resistance directly caused 1.27 million deaths worldwide, and resistant mutations in oncology render otherwise effective therapies useless within months of approval. The tools we use to predict these mutations are structurally blind. They require a crystal structure of the target protein, which exists for only a fraction of clinically relevant proteins, and they miss roughly 82 percent of the mutations that matter. My venture removes that requirement entirely.

I am Eniola Olutogun, a licensed pharmacist and machine learning engineer. My venture predicts drug resistance mutations from protein sequence alone, using ESM-2 protein language model delta-embeddings combined with ECFP4 drug fingerprints, classified by a Random Forest model. On the Platinum benchmark, a set of 553 mutations with protein-grouped cross-validation, the model achieves an AUROC of 0.804 plus or minus 0.025. That beats the published state of the art, mCSM-lig, which sits near 0.70. On SKEMPI 2.0, the model reaches 0.634. Critically, the method achieves 100 percent mutation coverage because it never needs a structure. Structure-limited tools cover roughly 18 percent.

MEST AI Startup Programme exists to support African founders building AI-first companies. I am a Nigerian founder. The venture is AI-first by construction, not by label. The core innovation is a protein language model, a deep learning architecture trained on 65 million protein sequences, adapted to a drug resistance prediction task. This is not a conventional bioinformatics pipeline with a machine learning wrapper. The entire product thesis depends on the representational power of large language models applied to biology.

The programme emphasizes early-stage startups with a prototype and proof of concept. The venture meets both. The model is built, validated on independent benchmarks, and outperforms existing tools. The business model targets pharmaceutical partnerships. The roadmap is concrete: 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, then convert that pilot into recurring revenue. Named partners include Servier, Paris-Saclay I2BC, Institut Pasteur, and Sanofi in Gentilly.

MEST values coachability and openness to structure. I am a solo founder, which means I have built the entire technical stack alone, but it also means I need the operational and go-to-market discipline that a programme like MEST provides. The network across African tech ecosystems and the AI focus are the specific reasons I am applying here rather than to a general biotech incubator.

The global health angle is direct. Drug resistance kills over a million people per year, and the burden falls disproportionately on African health systems. A tool that predicts resistance from sequence data, without expensive structural biology infrastructure, is a tool built for the constraints of real-world clinical settings. That is the venture I am building, and that is the venture I want to build inside MEST.

RESEARCH STATEMENT

The scientific problem is precise: given a protein sequence and a drug molecule, predict which mutations in that protein will confer resistance to that drug. This is a supervised learning problem with a clear input-output structure. The input is a protein sequence, typically 300 to 1,000 amino acids, and a drug represented as a molecular fingerprint. The output is a binary label indicating whether a given mutation confers resistance.

The technical approach has three components. First, the protein sequence is passed through ESM-2, a protein language model trained on 65 million unique protein sequences. The model produces per-position embeddings. The delta embedding, the difference between the wild-type and mutant embeddings at the mutation site, captures the biophysical effect of the amino acid substitution without requiring a folded structure. Second, the drug is encoded as an ECFP4 extended connectivity fingerprint, a standard molecular representation that captures local chemical environments. Third, the concatenated delta embedding and drug fingerprint are passed to a Random Forest classifier.

The validation is already done. On the Platinum benchmark, which contains 553 mutations across 29 proteins with protein-grouped cross-validation to prevent data leakage, the model achieves an AUROC of 0.804 with a standard deviation of 0.025. The published state of the art, mCSM-lig, reports approximately 0.70 on comparable tasks. On SKEMPI 2.0, a more challenging benchmark with a different distribution of mutation types, the model achieves 0.634. The key operational advantage is coverage. Structure-based tools require a crystal structure of the protein-ligand complex, which exists for only a fraction of clinically relevant targets. The sequence-based approach covers 100 percent of mutations in any protein with a known sequence.

The current limitation is clear. The model was trained on a relatively small dataset of experimentally validated resistance mutations. The next step is to fine-tune ESM-2 on the SKEMPI 3K dataset, which contains roughly 3,000 mutation stability and binding affinity measurements. This fine-tuning should improve the quality of the embeddings for the resistance prediction task, targeting an AUROC of at least 0.70 on SKEMPI 2.0. That target is realistic because fine-tuning a language model on domain-specific data is a well-established transfer learning technique, and the SKEMPI 3K dataset is an order of magnitude larger than the current training set.

The commercial application is a software-as-a-service platform for pharmaceutical R&D. The customer is a drug discovery team that needs to assess resistance risk for a candidate molecule before committing to clinical development. The workflow is: input a target protein sequence and a candidate drug, receive a ranked list of likely resistance mutations with confidence scores. The pilot target is Servier in Suresnes, with named relationships already in progress through SEMIA and Quest for Health. The revenue model is an annual subscription per therapeutic program.

The broader scientific contribution is a general method for sequence-based resistance prediction. The same architecture applies to antiviral resistance, antibacterial resistance, and oncology resistance. The only change is the training data. This is a platform technology, not a single-use model.

The African relevance is not incidental. Antimicrobial resistance rates in sub-Saharan Africa are among the highest in the world, and structural biology infrastructure is scarce. A tool that works from sequence data alone is deployable in settings where a cryo-EM facility is not available. The venture is registered to be built with this deployment context in mind.

ESSAY: AFRICAN FOUNDER AND AI INNOVATION

MEST selects African founders building AI startups. I am a Nigerian founder, and the AI is not an add-on to the venture, it is the entire venture. The product is a machine learning model. The data is protein sequences and drug fingerprints. The output is a resistance prediction. There is no wet lab component, no hardware, no manual curation step in the product itself. The company is an AI company that happens to operate in the biotech sector.

The African founder angle is substantive, not symbolic. Drug resistance is a first-order public health crisis in Nigeria and across the continent. The WHO African region carries the highest burden of antimicrobial resistance per capita in the world. A prediction tool that works from sequence data, without requiring expensive structural biology infrastructure, is designed for the constraints of African research and clinical settings. I built this tool with that constraint in mind because I am from that context.

My background is the bridge. I trained as a pharmacist, which means I understand the clinical stakes of a resistance prediction. I then trained as a machine learning engineer, which means I can build the model myself. I am the sole author of the codebase. This combination is rare, and it is the reason the venture exists at all. Most computational biology startups are founded by PhDs in bioinformatics who have never administered a drug. Most pharma companies are led by people who have never written a neural network. I have done both.

The AI innovation is the use of protein language model delta embeddings. ESM-2 is a transformer architecture trained on 65 million protein sequences. The delta embedding, the difference between wild-type and mutant representations, is a novel way to encode the effect of a single amino acid substitution. This is not a standard feature engineering approach. It is a direct application of the representational power of large language models to a structured biological prediction task. The AUROC of 0.804 on the Platinum benchmark, beating the published state of the art by a meaningful margin, is evidence that the approach works.

MEST asks whether the startup can scale. The answer is yes, because the marginal cost of a prediction is near zero once the model is trained. The constraint is data, not compute. The roadmap is to fine-tune on larger datasets, which requires funding for compute and for data acquisition. The MEST programme provides the structure, the network, and the compute resources to take the venture from proof of concept to pilot.

ESSAY: BUSINESS MODEL AND MARKET FIT

The customer is a pharmaceutical company with a drug candidate in preclinical or early clinical development. The problem that customer faces is resistance risk. A drug that works for six months and then fails because of a resistance mutation is a drug that loses its market entirely. The cost of discovering that failure late in clinical development is measured in hundreds of millions of dollars. The venture sells a prediction that identifies resistance risk early, when the cost of changing the molecule is low.

The business model is a software subscription. The unit of sale is a therapeutic program. A customer subscribes to the platform for a specific target protein and a specific drug candidate. The platform returns a ranked list of likely resistance mutations with confidence scores. The subscription is annual and priced per program. The target price point is in the range of 50,000 to 150,000 euros per program per year, which is small relative to the cost of a single failed clinical trial.

The pilot target is Servier in Suresnes. The relationship is already in progress through SEMIA and Quest for Health. The named partners also include Paris-Saclay I2BC, Institut Pasteur, and Sanofi in Gentilly. These are not speculative contacts. They are named institutions with ongoing conversations.

The market size is defined by the number of drug development programs globally. There are thousands of programs in preclinical development at any given time across the major pharmaceutical companies and biotech firms. The total addressable market for resistance prediction software is in the hundreds of millions of euros annually.

The competitive landscape is clear. The incumbent tools, mCSM-lig and similar structure-based methods, require a crystal structure and cover only about 18 percent of mutations. The venture covers 100 percent of mutations and beats the incumbents on accuracy. The moat is the fine-tuned ESM-2 model, which improves with more data, and the proprietary training set of validated resistance mutations.

The revenue model is not speculative. The pilot with Servier is designed to produce a paid contract within 12 months of the pilot start. The roadmap is: fine-tune ESM-2 on SKEMPI 3K, reach AUROC of at least 0.70 on SKEMPI 2.0, run the Servier pilot, convert to annual recurring revenue. Each step has a measurable success criterion.

CHECKLIST

- [ ] Create account on MEST AI Startup Programme application portal at mestaistartupprogram.smapply.us
- [ ] Complete the online application form with founder details, venture name, and contact information
- [ ] Upload this motivation letter as the personal statement document
- [ ] Upload this research statement as the technical description document
- [ ] Upload the two essay responses in the designated short answer fields
- [ ] Prepare and upload a pitch deck (10 to 15 slides) covering problem, solution, validation results, business model, and roadmap
- [ ] Prepare a one-page executive summary PDF with AUROC benchmark results and named partner list
- [ ] Verify the application deadline on the programme website and submit before the cutoff
- [ ] Confirm eligibility as a Nigerian founder and confirm any residency or relocation requirements with MEST staff
- [ ] Prepare a 90-second video pitch if the application portal requests one
- [ ] Request a letter of reference from the SEMIA or Quest for Health contact supporting the Servier pilot relationship
- [ ] Compile a list of compute credits or infrastructure needs to discuss during the interview stage

EDITOR NOTES

- Eligibility risk: the profile notes the programme is open to African founders, but the strategy notes flag uncertainty about whether the founder must be based in Africa or willing to relocate. Eniola must confirm the residency requirement with MEST staff before investing further time in the application. If relocation to an MEST campus is required, the venture timeline for the Servier pilot in Suresnes may conflict.

- Fact verification needed: the AUROC of 0.804 on Platinum and 0.634 on SKEMPI 2.0 are stated in the profile and used throughout this draft. Eniola must verify these numbers against the actual benchmark runs and be prepared to explain the exact cross-validation scheme in an interview. The claim that mCSM-lig achieves approximately 0.70 must also be verified against the published source.

- Gap to fill: the draft references the SEMIA and Quest for Health relationship as in progress, but the specific contact name, the stage of the conversation, and any written commitment from Servier are not in the profile. Eniola must insert the actual status of this relationship before submitting, because MEST will likely ask for proof of customer traction.

- Personal detail gap: the essays reference Eniola's pharmacist training and ML engineering background but do not include specific institutions, degrees, or years of experience. MEST evaluates founder credibility, so Eniola should add the specific university, the pharmacy license details, and the ML training pathway into the application form fields.

- Stage honesty: the venture is pre-seed, proof-of-concept validated, and not yet incorporated. The draft does not claim revenue, product-market fit, or validated IP. Eniola must maintain this honesty in the interview and not overstate the Servier relationship as a signed contract when it is a conversation in progress.

MOTIVATION LETTER

Drug resistance is a present, measurable crisis. In 2019, bacterial antimicrobial resistance directly caused 1.27 million deaths worldwide, and resistant mutations in oncology render otherwise effective therapies useless within months of approval. The tools we use to predict these mutations are structurally blind. They require a crystal structure of the target protein, which exists for only a fraction of clinically relevant proteins, and they miss roughly 82 percent of the mutations that matter. My venture removes that requirement entirely.

I am Eniola Olutogun, a licensed pharmacist and machine learning engineer. My venture predicts drug resistance mutations from protein sequence alone, using ESM-2 protein language model delta-embeddings combined with ECFP4 drug fingerprints, classified by a Random Forest model. On the Platinum benchmark, a set of 553 mutations with protein-grouped cross-validation, the model achieves an AUROC of 0.804 plus or minus 0.025. That beats the published state of the art, mCSM-lig, which sits near 0.70. On SKEMPI 2.0, the model reaches 0.634. Critically, the method achieves 100 percent mutation coverage because it never needs a structure. Structure-limited tools cover roughly 18 percent.

MEST AI Startup Programme exists to support African founders building AI-first companies. I am a Nigerian founder. The venture is AI-first by construction, not by label. The core innovation is a protein language model, a deep learning architecture trained on 65 million protein sequences, adapted to a drug resistance prediction task. This is not a conventional bioinformatics pipeline with a machine learning wrapper. The entire product thesis depends on the representational power of large language models applied to biology.

The programme emphasizes early-stage startups with a prototype and proof of concept. The venture meets both. The model is built, validated on independent benchmarks, and outperforms existing tools. The business model targets pharmaceutical partnerships. The roadmap is concrete: 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, then convert that pilot into recurring revenue. Named partners include Servier, Paris-Saclay I2BC, Institut Pasteur, and Sanofi in Gentilly.

MEST values coachability and openness to structure. I am a solo founder, which means I have built the entire technical stack alone, but it also means I need the operational and go-to-market discipline that a programme like MEST provides. The network across African tech ecosystems and the AI focus are the specific reasons I am applying here rather than to a general biotech incubator.

The global health angle is direct. Drug resistance kills over a million people per year, and the burden falls disproportionately on African health systems. A tool that predicts resistance from sequence data, without expensive structural biology infrastructure, is a tool built for the constraints of real-world clinical settings. That is the venture I am building, and that is the venture I want to build inside MEST.

RESEARCH STATEMENT

The scientific problem is precise: given a protein sequence and a drug molecule, predict which mutations in that protein will confer resistance to that drug. This is a supervised learning problem with a clear input-output structure. The input is a protein sequence, typically 300 to 1,000 amino acids, and a drug represented as a molecular fingerprint. The output is a binary label indicating whether a given mutation confers resistance.

The technical approach has three components. First, the protein sequence is passed through ESM-2, a protein language model trained on 65 million unique protein sequences. The model produces per-position embeddings. The delta embedding, the difference between the wild-type and mutant embeddings at the mutation site, captures the biophysical effect of the amino acid substitution without requiring a folded structure. Second, the drug is encoded as an ECFP4 extended connectivity fingerprint, a standard molecular representation that captures local chemical environments. Third, the concatenated delta embedding and drug fingerprint are passed to a Random Forest classifier.

The validation is already done. On the Platinum benchmark, which contains 553 mutations across 29 proteins with protein-grouped cross-validation to prevent data leakage, the model achieves an AUROC of 0.804 with a standard deviation of 0.025. The published state of the art, mCSM-lig, reports approximately 0.70 on comparable tasks. On SKEMPI 2.0, a more challenging benchmark with a different distribution of mutation types, the model achieves 0.634. The key operational advantage is coverage. Structure-based tools require a crystal structure of the protein-ligand complex, which exists for only a fraction of clinically relevant targets. The sequence-based approach covers 100 percent of mutations in any protein with a known sequence.

The current limitation is clear. The model was trained on a relatively small dataset of experimentally validated resistance mutations. The next step is to fine-tune ESM-2 on the SKEMPI 3K dataset, which contains roughly 3,000 mutation stability and binding affinity measurements. This fine-tuning should improve the quality of the embeddings for the resistance prediction task, targeting an AUROC of at least 0.70 on SKEMPI 2.0. That target is realistic because fine-tuning a language model on domain-specific data is a well-established transfer learning technique, and the SKEMPI 3K dataset is an order of magnitude larger than the current training set.

The commercial application is a software-as-a-service platform for pharmaceutical R&D. The customer is a drug discovery team that needs to assess resistance risk for a candidate molecule before committing to clinical development. The workflow is: input a target protein sequence and a candidate drug, receive a ranked list of likely resistance mutations with confidence scores. The pilot target is Servier in Suresnes, with named relationships already in progress through SEMIA and Quest for Health. The revenue model is an annual subscription per therapeutic program.

The broader scientific contribution is a general method for sequence-based resistance prediction. The same architecture applies to antiviral resistance, antibacterial resistance, and oncology resistance. The only change is the training data. This is a platform technology, not a single-use model.

The African relevance is not incidental. Antimicrobial resistance rates in sub-Saharan Africa are among the highest in the world, and structural biology infrastructure is scarce. A tool that works from sequence data alone is deployable in settings where a cryo-EM facility is not available. The venture is registered to be built with this deployment context in mind.

ESSAY: AFRICAN FOUNDER AND AI INNOVATION

MEST selects African founders building AI startups. I am a Nigerian founder, and the AI is not an add-on to the venture, it is the entire venture. The product is a machine learning model. The data is protein sequences and drug fingerprints. The output is a resistance prediction. There is no wet lab component, no hardware, no manual curation step in the product itself. The company is an AI company that happens to operate in the biotech sector.

The African founder angle is substantive, not symbolic. Drug resistance is a first-order public health crisis in Nigeria and across the continent. The WHO African region carries the highest burden of antimicrobial resistance per capita in the world. A prediction tool that works from sequence data, without requiring expensive structural biology infrastructure, is designed for the constraints of African research and clinical settings. I built this tool with that constraint in mind because I am from that context.

My background is the bridge. I trained as a pharmacist, which means I understand the clinical stakes of a resistance prediction. I then trained as a machine learning engineer, which means I can build the model myself. I am the sole author of the codebase. This combination is rare, and it is the reason the venture exists at all. Most computational biology startups are founded by PhDs in bioinformatics who have never administered a drug. Most pharma companies are led by people who have never written a neural network. I have done both.

The AI innovation is the use of protein language model delta embeddings. ESM-2 is a transformer architecture trained on 65 million protein sequences. The delta embedding, the difference between wild-type and mutant representations, is a novel way to encode the effect of a single amino acid substitution. This is not a standard feature engineering approach. It is a direct application of the representational power of large language models to a structured biological prediction task. The AUROC of 0.804 on the Platinum benchmark, beating the published state of the art by a meaningful margin, is evidence that the approach works.

MEST asks whether the startup can scale. The answer is yes, because the marginal cost of a prediction is near zero once the model is trained. The constraint is data, not compute. The roadmap is to fine-tune on larger datasets, which requires funding for compute and for data acquisition. The MEST programme provides the structure, the network, and the compute resources to take the venture from proof of concept to pilot.

ESSAY: BUSINESS MODEL AND MARKET FIT

The customer is a pharmaceutical company with a drug candidate in preclinical or early clinical development. The problem that customer faces is resistance risk. A drug that works for six months and then fails because of a resistance mutation is a drug that loses its market entirely. The cost of discovering that failure late in clinical development is measured in hundreds of millions of dollars. The venture sells a prediction that identifies resistance risk early, when the cost of changing the molecule is low.

The business model is a software subscription. The unit of sale is a therapeutic program. A customer subscribes to the platform for a specific target protein and a specific drug candidate. The platform returns a ranked list of likely resistance mutations with confidence scores. The subscription is annual and priced per program. The target price point is in the range of 50,000 to 150,000 euros per program per year, which is small relative to the cost of a single failed clinical trial.

The pilot target is Servier in Suresnes. The relationship is already in progress through SEMIA and Quest for Health. The named partners also include Paris-Saclay I2BC, Institut Pasteur, and Sanofi in Gentilly. These are not speculative contacts. They are named institutions with ongoing conversations.

The market size is defined by the number of drug development programs globally. There are thousands of programs in preclinical development at any given time across the major pharmaceutical companies and biotech firms. The total addressable market for resistance prediction software is in the hundreds of millions of euros annually.

The competitive landscape is clear. The incumbent tools, mCSM-lig and similar structure-based methods, require a crystal structure and cover only about 18 percent of mutations. The venture covers 100 percent of mutations and beats the incumbents on accuracy. The moat is the fine-tuned ESM-2 model, which improves with more data, and the proprietary training set of validated resistance mutations.

The revenue model is not speculative. The pilot with Servier is designed to produce a paid contract within 12 months of the pilot start. The roadmap is: fine-tune ESM-2 on SKEMPI 3K, reach AUROC of at least 0.70 on SKEMPI 2.0, run the Servier pilot, convert to annual recurring revenue. Each step has a measurable success criterion.

CHECKLIST

- [ ] Create account on MEST AI Startup Programme application portal at mestaistartupprogram.smapply.us
- [ ] Complete the online application form with founder details, venture name, and contact information
- [ ] Upload this motivation letter as the personal statement document
- [ ] Upload this research statement as the technical description document
- [ ] Upload the two essay responses in the designated short answer fields
- [ ] Prepare and upload a pitch deck (10 to 15 slides) covering problem, solution, validation results, business model, and roadmap
- [ ] Prepare a one-page executive summary PDF with AUROC benchmark results and named partner list
- [ ] Verify the application deadline on the programme website and submit before the cutoff
- [ ] Confirm eligibility as a Nigerian founder and confirm any residency or relocation requirements with MEST staff
- [ ] Prepare a 90-second video pitch if the application portal requests one
- [ ] Request a letter of reference from the SEMIA or Quest for Health contact supporting the Servier pilot relationship
- [ ] Compile a list of compute credits or infrastructure needs to discuss during the interview stage

EDITOR NOTES

- Eligibility risk: the profile notes the programme is open to African founders, but the strategy notes flag uncertainty about whether the founder must be based in Africa or willing to relocate. Eniola must confirm the residency requirement with MEST staff before investing further time in the application. If relocation to an MEST campus is required, the venture timeline for the Servier pilot in Suresnes may conflict.

- Fact verification needed: the AUROC of 0.804 on Platinum and 0.634 on SKEMPI 2.0 are stated in the profile and used throughout this draft. Eniola must verify these numbers against the actual benchmark runs and be prepared to explain the exact cross-validation scheme in an interview. The claim that mCSM-lig achieves approximately 0.70 must also be verified against the published source.

- Gap to fill: the draft references the SEMIA and Quest for Health relationship as in progress, but the specific contact name, the stage of the conversation, and any written commitment from Servier are not in the profile. Eniola must insert the actual status of this relationship before submitting, because MEST will likely ask for proof of customer traction.

- Personal detail gap: the essays reference Eniola's pharmacist training and ML engineering background but do not include specific institutions, degrees, or years of experience. MEST evaluates founder credibility, so Eniola should add the specific university, the pharmacy license details, and the ML training pathway into the application form fields.

- Stage honesty: the venture is pre-seed, proof-of-concept validated, and not yet incorporated. The draft does not claim revenue, product-market fit, or validated IP. Eniola must maintain this honesty in the interview and not overstate the Servier relationship as a signed contract when it is a conversation in progress.

AI Draft — MEST AI Startup Programme