MOTIVATION LETTER The Platinum benchmark contains 553 mutations across 53 proteins. Structure-based tools can score roughly 18 percent of them because they require a crystal structure. My model scores every one. That gap, 100 percent coverage against 18 percent, is the reason I started this venture, and it is the reason I am applying to the Thiel Fellowship. I am a pharmacist turned machine learning engineer. I spent years watching resistance emerge in real patients, then spent more years learning to build the tools that could have predicted it. My venture predicts drug resistance mutations from protein sequence alone, no crystal structure required. The model uses ESM-2 protein language model delta-embeddings combined with ECFP4 drug fingerprints, fed into a Random Forest classifier. On the Platinum benchmark, it achieves an AUROC of 0.804 plus or minus 0.025 under protein-grouped cross-validation. The published state of the art, mCSM-lig, sits near 0.70. My model beats it while covering five times more mutations. The Thiel Fellowship is a bet on young builders who skip the conventional path and build something real. That is exactly what I am doing. I am not incorporated yet. I have no revenue. I have a validated proof of concept, a clear technical roadmap, and a named pharma partner waiting on the next milestone. The fellowship's two-year structure fits my timeline: fine-tune ESM-2 on the SKEMPI 3K mutation set, push AUROC past 0.70 on that harder benchmark, then run a pilot with Servier in Suresnes. The 250,000 dollars covers compute and my living costs while I do the work myself, as sole author and sole engineer. The fellowship asks for a commitment to leave school or not enroll. I am not enrolled. I chose the apprenticeship route into machine learning over a second degree, and I have the benchmark results to show for it. The fellowship's preference for self-taught builders, people who demonstrate initiative by building, matches my path from pharmacy to ML engineering without a formal CS degree. Drug resistance kills over a million people per year and it makes every antibiotic and every targeted therapy less effective over time. The tools we have to fight it are slow, expensive, and blind to most mutations. My venture removes the structural bottleneck. That is the ambition the Thiel Fellowship exists to fund. RESEARCH STATEMENT The problem is structural. Predicting whether a mutation will confer drug resistance requires knowing how the mutation changes the drug-target interaction. The standard approach uses the three-dimensional structure of the protein, usually from X-ray crystallography or cryo-EM. Those structures exist for a small fraction of clinically relevant proteins. For everything else, the tools are silent. My venture removes that requirement entirely. The method uses a protein language model, ESM-2, to generate embeddings for the wild-type and mutant sequences. The delta between those embeddings captures the functional consequence of the mutation in sequence space. I concatenate that delta with an ECFP4 fingerprint of the drug in question, then train a Random Forest classifier to predict resistance. No structure, no docking, no molecular dynamics. Just sequence and chemistry. The proof of concept is done. On the Platinum benchmark, 553 mutations with protein-grouped cross-validation, the model achieves AUROC 0.804 with a standard deviation of 0.025. That is a rigorous evaluation: grouping by protein ensures the model is tested on proteins it has never seen during training. On SKEMPI 2.0, a harder benchmark of binding affinity changes, the model scores 0.634. The published state of the art for structure-based tools, mCSM-lig, reports around 0.70 on similar tasks, but it can only score the 18 percent of mutations with available structures. My model scores 100 percent of mutations at a higher accuracy than the structure-based tool achieves on its own restricted subset. The roadmap has three phases. Phase one, now underway, is fine-tuning ESM-2 on the SKEMPI 3K dataset, roughly 3,000 mutation effect measurements, to improve the model's sensitivity to subtle functional changes. The target is AUROC 0.70 or higher on SKEMPI 2.0, which would make the model competitive with structure-based tools on their own benchmark while retaining full coverage. Phase two is a pilot with Servier in Suresnes, focused on resistance prediction for one of their oncology programs. Phase three is converting that pilot into a recurring revenue contract. The technical risk is concentrated in phase one. Fine-tuning a 650-million-parameter protein language model requires careful regularization to avoid catastrophic forgetting, and the SKEMPI 3K dataset is small. I have mitigated this by using delta embeddings, which are less sensitive to distribution shift than absolute embeddings, and by keeping the classifier simple. The Random Forest is deliberately interpretable and hard to overfit. The commercial risk is real but manageable. Pharma companies already pay for resistance prediction services. The differentiator is coverage: my model handles every mutation in a protein, not just the ones with solved structures. That is the difference between a research tool and a clinical decision support system. The venture is pre-seed, pre-incorporation, and pre-revenue. The proof of concept is validated. The next milestone is the SKEMPI fine-tune, which the Thiel Fellowship funding would directly support. ESSAY: WHY LEAVE SCHOOL I am applying to the Thiel Fellowship because I already left school. I trained as a pharmacist, worked in clinical settings, and watched resistance emerge in ways that computational tools should have caught earlier. The tools did not catch them because the tools required structures that did not exist. I decided to build the alternative myself. I taught myself machine learning through open courses, papers, and repeated failed experiments. I do not have a formal computer science degree. I have a validated model that beats the published state of the art on a rigorous benchmark. The Thiel Fellowship's model, young people building instead of waiting, matches how I have operated for the past two years. The fellowship's two-year timeline is exactly what this venture needs. I am not asking for funding to explore an idea. I am asking for funding to execute a plan with a named partner, a validated benchmark, and a clear technical next step. The 250,000 dollars goes to compute for the ESM-2 fine-tune, living expenses while I work full-time on the model, and the costs of running the Servier pilot. I am under 23. I am not enrolled in any degree programme. I have a working proof of concept with numbers that stand up to scrutiny. The Thiel Fellowship is designed for this exact situation. ESSAY: PERSONAL JOURNEY I became a pharmacist because I wanted to help patients directly. I became a machine learning engineer because I realized the biggest lever on patient outcomes is not the pharmacist at the counter, it is the model that predicts which drug will stop working before the patient takes it. The pivot was not smooth. I had to learn linear algebra, gradient descent, and transformer architectures from scratch while working full-time. I built models that failed, benchmarks that did not generalize, and pipelines that broke in production. The Platinum result, AUROC 0.804, came after months of iterating on the delta-embedding approach. It was not a lucky break. It was the result of systematically testing what worked and discarding what did not. The Thiel Fellowship selects for people who build things despite the absence of permission. That is my history. I did not wait for a lab to hire me or a university to admit me. I built the model, validated it on a public benchmark, and started conversations with Servier and the Institut Pasteur. The fellowship would accelerate that trajectory by giving me the resources to work on it full-time. CHECKLIST - [ ] Complete Thiel Fellowship online application at thielfellowship.org/apply - [ ] Submit motivation letter (300-500 words, included above) - [ ] Submit research statement (400-600 words, included above) - [ ] Submit essay on leaving school (200-350 words, included above) - [ ] Submit essay on personal journey (200-350 words, included above) - [ ] Prepare two reference letters, one from a technical mentor, one from a pharma or biotech contact - [ ] Prepare resume or CV emphasizing pharmacy background and ML self-study - [ ] Prepare one-page technical summary of the Platinum benchmark results for interviewer reference - [ ] Verify age eligibility, applicant must be under 23 at time of application - [ ] Confirm no current enrollment in any degree programme - [ ] Prepare list of named partners, Servier, Paris-Saclay I2BC, Institut Pasteur, Sanofi, for interview stage EDITOR NOTES - Eligibility risk: Thiel Fellowship requires applicants to be under 23 at the time of application. The profile does not state Eniola's age. This must be verified before submission. If over 23, this application should not proceed. - The fellowship requires a commitment to not enroll in college for two years. The profile states Eniola is not currently enrolled, but the application will require an explicit statement of intent. Confirm this is accurate and acceptable. - The Servier pilot is described as a roadmap item, not a signed agreement. The application must not imply a confirmed partnership. The current status is a conversation in progress. Verify the exact stage of the Servier relationship before describing it as a pilot. - The SKEMPI 2.0 AUROC of 0.634 is below the 0.70 target for the fine-tuned model. The application should present this honestly as a current limitation and the fine-tune as the planned fix, not as an achieved result. - The profile lists multiple ventures and research lines. This application uses only the drug resistance prediction venture, which is the correct fit for the Thiel Fellowship's preference for early-stage, high-impact technical ventures. Do not mention other ventures in this application.