Draft — USAID Development Innovation Funding

MOTIVATION LETTER

The World Health Organization estimates that antimicrobial resistance will cause 10 million deaths per year by 2050, with sub-Saharan Africa bearing a disproportionate share of that burden. In Nigeria, where pharmacists dispense antibiotics without routine susceptibility testing, resistant infections are already undermining treatment for malaria, tuberculosis, and bacterial pneumonia. My venture addresses this gap directly: a computational platform that predicts drug resistance mutations from protein sequence alone, without requiring crystal structures that do not exist for most clinically relevant proteins.

I am Eniola Olutogun, a pharmacist and machine learning engineer. My proof-of-concept model uses ESM-2 protein language model delta-embeddings combined with ECFP4 drug fingerprints, classified by a Random Forest. On the Platinum benchmark, it achieves an AUROC of 0.804 plus or minus 0.025 under protein-grouped cross-validation across 553 mutations. This exceeds the published performance of mCSM-lig, which reports an AUROC of approximately 0.70. Critically, my approach achieves 100 percent mutation coverage, while structure-limited tools cover only about 18 percent of known mutations. On SKEMPI 2.0, the model reaches 0.634 AUROC, which reflects the harder generalization task of unseen protein families.

USAID's Development Innovation Funding targets digital health solutions that work in low-income settings. My platform is designed for exactly that environment. It runs on protein sequence data that can be obtained from portable sequencing devices, not on expensive structural biology infrastructure. The output is a ranked list of likely resistance mutations for a given drug, which a clinician or pharmacist can use to select alternative therapies. This is a decision-support tool, not a replacement for clinical judgment, and it is built to function where laboratory capacity is limited.

The venture is at pre-seed stage with validated proof-of-concept. I have not yet incorporated, and I have no revenue. What I have is a model that outperforms published state-of-the-art tools on benchmark data, a clear roadmap to improve it, and a named partnership pipeline that includes Servier in Suresnes, the Paris-Saclay Institute of Integrative Biology, and Sanofi in Gentilly. The next technical milestone is fine-tuning ESM-2 on the SKEMPI 3K mutation set to reach an AUROC of at least 0.70 on that harder benchmark, followed by a pilot with Servier.

USAID's selection criteria emphasize alignment with local needs, measurable impact, and cost-effectiveness. Nigeria has a national action plan for antimicrobial resistance but lacks the computational tools to implement it. My pharmacist background means I understand the dispensing workflow where this tool would sit. The model's 100 percent mutation coverage matters because it does not fail on the proteins that matter most in low-resource settings. This is a deployment pathway for resistance surveillance and treatment optimization in a country that needs both urgently.

RESEARCH STATEMENT

The technical problem is straightforward: given a protein sequence and a drug, predict which mutations will confer resistance. Existing tools require crystal structures, which exist for only a fraction of clinically relevant proteins. My approach removes that requirement entirely.

The model architecture combines three components. First, ESM-2, a protein language model pretrained on 65 million protein sequences, generates embeddings for the wild-type and mutant sequences. Second, the delta between those embeddings captures the biophysical effect of the mutation. Third, ECFP4 fingerprints encode the drug's chemical structure. A Random Forest classifier integrates these features and outputs a resistance probability.

Validation results are as follows. On the Platinum benchmark, which contains 553 mutations across 47 proteins, the model achieves an AUROC of 0.804 with a standard deviation of 0.025 under protein-grouped cross-validation. This grouping is critical: it ensures the model is tested on proteins it has never seen during training, which is the realistic deployment scenario. The published state-of-the-art tool, mCSM-lig, reports an AUROC of approximately 0.70 on similar tasks. My model also achieves 100 percent mutation coverage, meaning it can score every mutation in the benchmark, whereas structure-limited tools can only handle the 18 percent of mutations for which crystal structures are available.

On SKEMPI 2.0, a more challenging benchmark of binding affinity changes across diverse protein families, the model achieves 0.634 AUROC. This lower score reflects the difficulty of generalizing to unseen protein families, and it defines the next technical milestone. I will fine-tune ESM-2 on the SKEMPI 3K mutation set, which contains approximately 3,000 additional mutations, to improve generalization. The target is an AUROC of at least 0.70 on SKEMPI 2.0 after fine-tuning.

The roadmap has three phases. Phase one, over the next six months, is the SKEMPI 3K fine-tuning and benchmark evaluation. Phase two is a pilot with Servier in Suresnes, where the model will be tested on their internal resistance datasets. Phase three is the transition to a revenue-generating software-as-a-service product for pharmaceutical companies and clinical laboratories.

The relevance to USAID's mission is direct. Antimicrobial resistance is a development problem, not just a biomedical one. In Nigeria, the absence of routine susceptibility testing means that resistance emerges silently and spreads. My model can be deployed with sequence data alone, which makes it compatible with the portable sequencing infrastructure that is increasingly available in low-income settings. The output, a ranked list of likely resistance mutations, can inform treatment guidelines and surveillance programs.

The scientific novelty is the combination of protein language model embeddings with drug fingerprints for resistance prediction without structural data. The practical novelty is the 100 percent mutation coverage, which means the tool does not fail on the proteins that matter most in resource-limited settings. The model is validated, reproducible, and ready for the next stage of development.

SHORT ANSWER ESSAY: INNOVATION AND NOVELTY

Existing resistance prediction tools, such as mCSM-lig, require a crystal structure of the protein-drug complex. This requirement excludes the majority of clinically relevant mutations. My model eliminates the structural requirement by using ESM-2 protein language model embeddings, which encode evolutionary and biophysical information from sequence alone. The delta between wild-type and mutant embeddings captures the mutation's effect, and ECFP4 drug fingerprints integrate the chemical context. This architecture achieves 0.804 AUROC on the Platinum benchmark, outperforming mCSM-lig's 0.70, while covering 100 percent of mutations versus 18 percent for structure-limited tools. The novelty is a different input modality that expands the usable design space for resistance prediction.

SHORT ANSWER ESSAY: FEASIBILITY AND TECHNICAL SOUNDNESS

The model is already built and validated. The Random Forest classifier is computationally lightweight, trainable on a single GPU, and does not require the inference infrastructure that larger deep learning models demand. This is a deliberate design choice for deployment in low-resource settings. The validation protocol uses protein-grouped cross-validation, which prevents data leakage and provides an honest estimate of generalization performance. The next milestone, fine-tuning ESM-2 on SKEMPI 3K, is well-defined and achievable with existing computational resources. The partnership pipeline, including Servier and Sanofi, provides access to industrial validation data. The technical risk is concentrated in the SKEMPI 2.0 generalization gap, which the fine-tuning roadmap directly addresses.

SHORT ANSWER ESSAY: SCALABILITY AND SUSTAINABILITY

The platform scales along two axes. Technically, the model can be retrained on new protein families and new drugs as data becomes available, and the inference pipeline runs on standard hardware. Commercially, the target customers are pharmaceutical companies that need resistance profiling during drug development and clinical laboratories that need surveillance tools. The revenue model is software-as-a-service, with tiered pricing for academic, clinical, and industrial users. In Nigeria, the tool can be deployed through the existing pharmacy network, where my pharmacist background provides direct insight into the workflow. The long-term sustainability depends on the Servier pilot succeeding and converting into a paid contract, which is the stated roadmap.

SHORT ANSWER ESSAY: TEAM CAPACITY

I am a pharmacist with a license to practice and a machine learning engineer with hands-on experience building and validating the model described in this application. I am the sole author of the codebase and the sole designer of the validation protocol. This means there is no gap between the technical claims and the person responsible for them. The partnership pipeline includes Servier in Suresnes, the Paris-Saclay Institute of Integrative Biology, and Sanofi in Gentilly, which provides access to domain expertise and validation data. What I lack is a co-founder and a formal corporate structure; the venture is not yet incorporated. This is a known gap, and the first use of any grant funding will be to incorporate and hire a second engineer.

SHORT ANSWER ESSAY: MEASURABLE IMPACT

The impact metrics are defined at three levels. Technical impact: the model's AUROC on the Platinum benchmark is 0.804, and the target after SKEMPI 3K fine-tuning is 0.70 or higher on SKEMPI 2.0. Clinical impact: in a Nigerian pilot, the tool will be evaluated on its ability to identify resistance mutations in clinical isolates, with the target being at least 90 percent sensitivity compared to phenotypic susceptibility testing. Development impact: the tool will be deployed in at least one Nigerian clinical laboratory within 18 months of funding, and the resistance surveillance data will be shared with the national antimicrobial resistance action plan. Each metric is measurable, time-bound, and directly tied to USAID's development objectives.

SHORT ANSWER ESSAY: COST-EFFECTIVENESS

The model's computational footprint is small. Training the Random Forest classifier takes hours on a single GPU, and inference is near-instantaneous on standard hardware. This means the marginal cost per prediction is negligible, which is the opposite of structure-based methods that require expensive crystallography or cryo-electron microscopy. The grant request is for personnel, cloud compute, and the Nigerian pilot. The alternative, deploying phenotypic susceptibility testing infrastructure across Nigerian clinics, would require capital expenditure on laboratory equipment and trained staff. My platform provides a software-only complement that can be deployed on existing laptops and portable sequencers. The cost per patient outcome improved is substantially lower than the laboratory-based alternative.

CHECKLIST

- [ ] Confirm USAID Development Innovation Funding eligibility for non-U.S. applicants and pre-incorporation entities
- [ ] Verify the current AUROC figures against the latest Platinum benchmark and SKEMPI 2.0 results
- [ ] Obtain a letter of support or intent from Servier for the pilot phase
- [ ] Obtain a letter of support from a Nigerian clinical laboratory or pharmacy network for the deployment pilot
- [ ] Prepare a one-page budget breakdown for the grant amount requested
- [ ] Prepare a timeline with milestones for the SKEMPI 3K fine-tuning and Servier pilot
- [ ] Confirm the legal structure for receiving grant funds as a pre-incorporation founder
- [ ] Draft a data-sharing agreement template for the Nigerian clinical pilot
- [ ] Verify the contact details and submission portal for USAID Development Innovation Funding
- [ ] Prepare a one-page CV or resume for Eniola Olutogun

EDITOR NOTES

- Eligibility risk: USAID grants typically require a registered entity or a fiscal sponsor. The venture is not yet incorporated, so the applicant must either incorporate before submission or identify a fiscal sponsor. This must be resolved before the application is submitted.
- Verification needed: The Platinum benchmark AUROC of 0.804 and the SKEMPI 2.0 score of 0.634 must be re-run and documented with the exact code version and random seed before submission. The applicant should be prepared to share the code repository and validation scripts.
- Gap to fill: The application references a Nigerian clinical laboratory or pharmacy network for the deployment pilot, but no named partner is listed in the profile. The applicant must identify and secure a letter of support from a specific Nigerian institution before submission.
- Gap to fill: The budget breakdown is not included in the profile. The applicant must prepare a detailed budget that covers personnel, cloud compute, and the Nigerian pilot, and ensure it aligns with USAID's cost-effectiveness criteria.
- Framing note: The motivation letter leads with the antimicrobial resistance burden in Nigeria, which is the correct angle for USAID. The research statement is more technical and should be positioned as the evidence base for the development claim. The short answer essays map directly to USAID's stated selection criteria, which is intentional.

MOTIVATION LETTER

The World Health Organization estimates that antimicrobial resistance will cause 10 million deaths per year by 2050, with sub-Saharan Africa bearing a disproportionate share of that burden. In Nigeria, where pharmacists dispense antibiotics without routine susceptibility testing, resistant infections are already undermining treatment for malaria, tuberculosis, and bacterial pneumonia. My venture addresses this gap directly: a computational platform that predicts drug resistance mutations from protein sequence alone, without requiring crystal structures that do not exist for most clinically relevant proteins.

I am Eniola Olutogun, a pharmacist and machine learning engineer. My proof-of-concept model uses ESM-2 protein language model delta-embeddings combined with ECFP4 drug fingerprints, classified by a Random Forest. On the Platinum benchmark, it achieves an AUROC of 0.804 plus or minus 0.025 under protein-grouped cross-validation across 553 mutations. This exceeds the published performance of mCSM-lig, which reports an AUROC of approximately 0.70. Critically, my approach achieves 100 percent mutation coverage, while structure-limited tools cover only about 18 percent of known mutations. On SKEMPI 2.0, the model reaches 0.634 AUROC, which reflects the harder generalization task of unseen protein families.

USAID's Development Innovation Funding targets digital health solutions that work in low-income settings. My platform is designed for exactly that environment. It runs on protein sequence data that can be obtained from portable sequencing devices, not on expensive structural biology infrastructure. The output is a ranked list of likely resistance mutations for a given drug, which a clinician or pharmacist can use to select alternative therapies. This is a decision-support tool, not a replacement for clinical judgment, and it is built to function where laboratory capacity is limited.

The venture is at pre-seed stage with validated proof-of-concept. I have not yet incorporated, and I have no revenue. What I have is a model that outperforms published state-of-the-art tools on benchmark data, a clear roadmap to improve it, and a named partnership pipeline that includes Servier in Suresnes, the Paris-Saclay Institute of Integrative Biology, and Sanofi in Gentilly. The next technical milestone is fine-tuning ESM-2 on the SKEMPI 3K mutation set to reach an AUROC of at least 0.70 on that harder benchmark, followed by a pilot with Servier.

USAID's selection criteria emphasize alignment with local needs, measurable impact, and cost-effectiveness. Nigeria has a national action plan for antimicrobial resistance but lacks the computational tools to implement it. My pharmacist background means I understand the dispensing workflow where this tool would sit. The model's 100 percent mutation coverage matters because it does not fail on the proteins that matter most in low-resource settings. This is a deployment pathway for resistance surveillance and treatment optimization in a country that needs both urgently.

RESEARCH STATEMENT

The technical problem is straightforward: given a protein sequence and a drug, predict which mutations will confer resistance. Existing tools require crystal structures, which exist for only a fraction of clinically relevant proteins. My approach removes that requirement entirely.

The model architecture combines three components. First, ESM-2, a protein language model pretrained on 65 million protein sequences, generates embeddings for the wild-type and mutant sequences. Second, the delta between those embeddings captures the biophysical effect of the mutation. Third, ECFP4 fingerprints encode the drug's chemical structure. A Random Forest classifier integrates these features and outputs a resistance probability.

Validation results are as follows. On the Platinum benchmark, which contains 553 mutations across 47 proteins, the model achieves an AUROC of 0.804 with a standard deviation of 0.025 under protein-grouped cross-validation. This grouping is critical: it ensures the model is tested on proteins it has never seen during training, which is the realistic deployment scenario. The published state-of-the-art tool, mCSM-lig, reports an AUROC of approximately 0.70 on similar tasks. My model also achieves 100 percent mutation coverage, meaning it can score every mutation in the benchmark, whereas structure-limited tools can only handle the 18 percent of mutations for which crystal structures are available.

On SKEMPI 2.0, a more challenging benchmark of binding affinity changes across diverse protein families, the model achieves 0.634 AUROC. This lower score reflects the difficulty of generalizing to unseen protein families, and it defines the next technical milestone. I will fine-tune ESM-2 on the SKEMPI 3K mutation set, which contains approximately 3,000 additional mutations, to improve generalization. The target is an AUROC of at least 0.70 on SKEMPI 2.0 after fine-tuning.

The roadmap has three phases. Phase one, over the next six months, is the SKEMPI 3K fine-tuning and benchmark evaluation. Phase two is a pilot with Servier in Suresnes, where the model will be tested on their internal resistance datasets. Phase three is the transition to a revenue-generating software-as-a-service product for pharmaceutical companies and clinical laboratories.

The relevance to USAID's mission is direct. Antimicrobial resistance is a development problem, not just a biomedical one. In Nigeria, the absence of routine susceptibility testing means that resistance emerges silently and spreads. My model can be deployed with sequence data alone, which makes it compatible with the portable sequencing infrastructure that is increasingly available in low-income settings. The output, a ranked list of likely resistance mutations, can inform treatment guidelines and surveillance programs.

The scientific novelty is the combination of protein language model embeddings with drug fingerprints for resistance prediction without structural data. The practical novelty is the 100 percent mutation coverage, which means the tool does not fail on the proteins that matter most in resource-limited settings. The model is validated, reproducible, and ready for the next stage of development.

SHORT ANSWER ESSAY: INNOVATION AND NOVELTY

Existing resistance prediction tools, such as mCSM-lig, require a crystal structure of the protein-drug complex. This requirement excludes the majority of clinically relevant mutations. My model eliminates the structural requirement by using ESM-2 protein language model embeddings, which encode evolutionary and biophysical information from sequence alone. The delta between wild-type and mutant embeddings captures the mutation's effect, and ECFP4 drug fingerprints integrate the chemical context. This architecture achieves 0.804 AUROC on the Platinum benchmark, outperforming mCSM-lig's 0.70, while covering 100 percent of mutations versus 18 percent for structure-limited tools. The novelty is a different input modality that expands the usable design space for resistance prediction.

SHORT ANSWER ESSAY: FEASIBILITY AND TECHNICAL SOUNDNESS

The model is already built and validated. The Random Forest classifier is computationally lightweight, trainable on a single GPU, and does not require the inference infrastructure that larger deep learning models demand. This is a deliberate design choice for deployment in low-resource settings. The validation protocol uses protein-grouped cross-validation, which prevents data leakage and provides an honest estimate of generalization performance. The next milestone, fine-tuning ESM-2 on SKEMPI 3K, is well-defined and achievable with existing computational resources. The partnership pipeline, including Servier and Sanofi, provides access to industrial validation data. The technical risk is concentrated in the SKEMPI 2.0 generalization gap, which the fine-tuning roadmap directly addresses.

SHORT ANSWER ESSAY: SCALABILITY AND SUSTAINABILITY

The platform scales along two axes. Technically, the model can be retrained on new protein families and new drugs as data becomes available, and the inference pipeline runs on standard hardware. Commercially, the target customers are pharmaceutical companies that need resistance profiling during drug development and clinical laboratories that need surveillance tools. The revenue model is software-as-a-service, with tiered pricing for academic, clinical, and industrial users. In Nigeria, the tool can be deployed through the existing pharmacy network, where my pharmacist background provides direct insight into the workflow. The long-term sustainability depends on the Servier pilot succeeding and converting into a paid contract, which is the stated roadmap.

SHORT ANSWER ESSAY: TEAM CAPACITY

I am a pharmacist with a license to practice and a machine learning engineer with hands-on experience building and validating the model described in this application. I am the sole author of the codebase and the sole designer of the validation protocol. This means there is no gap between the technical claims and the person responsible for them. The partnership pipeline includes Servier in Suresnes, the Paris-Saclay Institute of Integrative Biology, and Sanofi in Gentilly, which provides access to domain expertise and validation data. What I lack is a co-founder and a formal corporate structure; the venture is not yet incorporated. This is a known gap, and the first use of any grant funding will be to incorporate and hire a second engineer.

SHORT ANSWER ESSAY: MEASURABLE IMPACT

The impact metrics are defined at three levels. Technical impact: the model's AUROC on the Platinum benchmark is 0.804, and the target after SKEMPI 3K fine-tuning is 0.70 or higher on SKEMPI 2.0. Clinical impact: in a Nigerian pilot, the tool will be evaluated on its ability to identify resistance mutations in clinical isolates, with the target being at least 90 percent sensitivity compared to phenotypic susceptibility testing. Development impact: the tool will be deployed in at least one Nigerian clinical laboratory within 18 months of funding, and the resistance surveillance data will be shared with the national antimicrobial resistance action plan. Each metric is measurable, time-bound, and directly tied to USAID's development objectives.

SHORT ANSWER ESSAY: COST-EFFECTIVENESS

The model's computational footprint is small. Training the Random Forest classifier takes hours on a single GPU, and inference is near-instantaneous on standard hardware. This means the marginal cost per prediction is negligible, which is the opposite of structure-based methods that require expensive crystallography or cryo-electron microscopy. The grant request is for personnel, cloud compute, and the Nigerian pilot. The alternative, deploying phenotypic susceptibility testing infrastructure across Nigerian clinics, would require capital expenditure on laboratory equipment and trained staff. My platform provides a software-only complement that can be deployed on existing laptops and portable sequencers. The cost per patient outcome improved is substantially lower than the laboratory-based alternative.

CHECKLIST

- [ ] Confirm USAID Development Innovation Funding eligibility for non-U.S. applicants and pre-incorporation entities
- [ ] Verify the current AUROC figures against the latest Platinum benchmark and SKEMPI 2.0 results
- [ ] Obtain a letter of support or intent from Servier for the pilot phase
- [ ] Obtain a letter of support from a Nigerian clinical laboratory or pharmacy network for the deployment pilot
- [ ] Prepare a one-page budget breakdown for the grant amount requested
- [ ] Prepare a timeline with milestones for the SKEMPI 3K fine-tuning and Servier pilot
- [ ] Confirm the legal structure for receiving grant funds as a pre-incorporation founder
- [ ] Draft a data-sharing agreement template for the Nigerian clinical pilot
- [ ] Verify the contact details and submission portal for USAID Development Innovation Funding
- [ ] Prepare a one-page CV or resume for Eniola Olutogun

EDITOR NOTES

- Eligibility risk: USAID grants typically require a registered entity or a fiscal sponsor. The venture is not yet incorporated, so the applicant must either incorporate before submission or identify a fiscal sponsor. This must be resolved before the application is submitted.
- Verification needed: The Platinum benchmark AUROC of 0.804 and the SKEMPI 2.0 score of 0.634 must be re-run and documented with the exact code version and random seed before submission. The applicant should be prepared to share the code repository and validation scripts.
- Gap to fill: The application references a Nigerian clinical laboratory or pharmacy network for the deployment pilot, but no named partner is listed in the profile. The applicant must identify and secure a letter of support from a specific Nigerian institution before submission.
- Gap to fill: The budget breakdown is not included in the profile. The applicant must prepare a detailed budget that covers personnel, cloud compute, and the Nigerian pilot, and ensure it aligns with USAID's cost-effectiveness criteria.
- Framing note: The motivation letter leads with the antimicrobial resistance burden in Nigeria, which is the correct angle for USAID. The research statement is more technical and should be positioned as the evidence base for the development claim. The short answer essays map directly to USAID's stated selection criteria, which is intentional.

AI Draft — USAID Development Innovation Funding