MOTIVATION LETTER Servier's Paris-Saclay R&D center houses the infrastructure that can close the loop between computational prediction and experimental validation. My venture predicts drug resistance mutations from protein sequence alone, using an ESM-2 protein language model with delta-embeddings and ECFP4 drug fingerprints classified by a Random Forest. On the Platinum benchmark of 553 mutations, the model achieves AUROC 0.804 plus or minus 0.025 under protein-grouped cross-validation. This covers 100 percent of mutations in the test set, compared to roughly 18 percent for structure-dependent tools. The published SOTA, mCSM-lig, reports AUROC approximately 0.70. SPartners targets biotech startups aligned with Servier's therapeutic focus in oncology and immunology. Drug resistance is a direct clinical bottleneck in both areas. My model requires no crystal structure, which means it can score mutations for any protein target where sequence data exists. The current proof-of-concept is validated and pre-seed. I am a pharmacist turned machine learning engineer and the sole founder. The venture is not yet incorporated. The incubator's equipped labs at Paris-Saclay offer a path to experimental validation. I have an existing relationship with Servier in Suresnes and with Paris-Saclay's I2BC and Institut Pasteur. SPartners would provide the wet-lab access needed to test predicted resistance mutations against actual drug compounds, generating data that strengthens the model and builds credibility with pharma partners. The Golden Ticket option for applications before December 1st is a clear incentive to move quickly. I am applying to SPartners because the fit is specific. Servier is a named partner. The geography is Ile-de-France. The incubator's resources match the venture's next step: bridging computational prediction with experimental confirmation. The programme's evaluation criteria include scientific innovation, clinical impact potential, team expertise, and feasibility of the development plan. My background in both pharmacy and machine learning, combined with the validated benchmark results, addresses each criterion directly. SHORT ESSAY: TECHNICAL INNOVATION AND CLINICAL IMPACT Drug resistance mutations emerge during treatment and render therapies ineffective. Current computational tools require a crystal structure of the protein-drug complex to make predictions. For the majority of clinically relevant mutations, no structure exists. My venture solves this by using ESM-2 protein language model delta-embeddings, which capture the biophysical effect of a mutation from sequence alone. The model combines these embeddings with ECFP4 drug fingerprints and classifies resistance versus sensitivity using a Random Forest. The Platinum benchmark is the standard for resistance prediction. My model scores AUROC 0.804, beating the published SOTA of 0.70. On SKEMPI 2.0, a binding affinity benchmark, the model achieves 0.634. The next development step is fine-tuning ESM-2 on the full SKEMPI 3K mutation set, targeting AUROC 0.70 or higher on that harder benchmark. This would match or exceed structure-based tools while maintaining 100 percent mutation coverage. Clinical impact is direct. In oncology, resistance to kinase inhibitors and targeted therapies limits patient survival. In infectious disease, antimicrobial resistance is a global health crisis. A tool that predicts resistance before it emerges clinically allows drug developers to design around escape mutations and clinicians to select therapies less likely to fail. Servier's oncology pipeline is a natural first application. SHORT ESSAY: DEVELOPMENT PLAN AND INCUBATOR FIT The venture's roadmap has three phases. Phase one, duration six months: fine-tune ESM-2 on the SKEMPI 3K mutation set. Target AUROC 0.70 or higher. Phase two, months seven through twelve: run a pilot with Servier's computational biology group in Suresnes, testing the model on internal resistance datasets. Phase three, months thirteen through eighteen: establish a subscription-based software-as-a-service model targeting pharma R&D teams, targeting annual recurring revenue. SPartners fits each phase. The incubator's compute infrastructure supports the model fine-tuning in phase one. The equipped labs allow experimental validation of predicted mutations in phase two, generating data that Servier's team will trust. The incubator's network within Servier and Paris-Saclay accelerates the pilot partnership. The Golden Ticket, if secured, provides free one-year residency, reducing burn rate during the pre-revenue period. I have already initiated contact with SEMIA and Quest for Health. WILCO One BioTech is scheduled for October 2026. IncubAlliance and AI House applications are submitted. The EIC Accelerator and BPI i-Lab are future targets. SPartners is the near-term priority because it directly connects to Servier, the named partner, and provides the physical infrastructure that a computational biotech venture needs to generate experimental evidence. RESEARCH STATEMENT Drug resistance is a moving target. A mutation that confers resistance to one compound may be sensitive to a second. Predicting which mutations will arise and which drugs will fail requires a model that generalizes across protein families and drug chemotypes. My approach uses protein language model embeddings to represent the mutation's effect on the protein, and drug fingerprints to represent the compound. The classifier learns the interaction between these two representations. The model architecture is straightforward. ESM-2 generates per-residue embeddings for the wild-type and mutant sequences. The delta embedding, the vector difference between the two, captures the mutation's biophysical signature. ECFP4 fingerprints encode the drug's molecular structure. These two feature vectors are concatenated and fed into a Random Forest classifier. No crystal structure, no docking, no molecular dynamics. The model runs on any protein-drug pair where sequence and SMILES are available. Benchmark results support the approach. On Platinum, 553 mutations across diverse protein targets, the model achieves AUROC 0.804. Protein-grouped cross-validation ensures that the model is not memorizing protein-specific patterns. On SKEMPI 2.0, a more challenging benchmark of binding affinity mutations, the model scores 0.634. The gap between these two numbers defines the next technical milestone: improve SKEMPI performance to 0.70 or higher through fine-tuning. The fine-tuning strategy uses the full SKEMPI 3K mutation set, which includes both binding affinity and resistance data. I will freeze the ESM-2 backbone and train a lightweight adapter layer, preserving the pretrained representations while adapting to the resistance prediction task. This approach requires modest compute and is feasible within the incubator's resources. Experimental validation is the critical next step. Predicted resistance mutations must be tested in cell-based assays. SPartners' labs at Paris-Saclay provide this capability. I will design a validation panel of 20 to 30 mutations predicted by the model, test them against the corresponding drugs, and compare the results to the model's confidence scores. A concordance rate above 80 percent would de-risk the technology for pharma partnerships. The long-term research goal is a platform that predicts resistance for any protein target and any drug, updated continuously as new resistance data becomes available. This platform would serve as a decision support tool for drug discovery teams, flagging potential resistance liabilities early in the development cycle. CHECKLIST - [ ] Motivation letter, 300-500 words, submitted as part of the application form - [ ] Short essay on technical innovation and clinical impact, 200-350 words - [ ] Short essay on development plan and incubator fit, 200-350 words - [ ] Research statement, 400-600 words - [ ] CV or resume for Eniola Olutogun - [ ] Proof-of-concept results summary (AUROC 0.804 on Platinum, 0.634 on SKEMPI 2.0) - [ ] Letter of support or expression of interest from Servier contact, if available - [ ] Application submitted before December 1st to qualify for Golden Ticket - [ ] Company incorporation documents, if incorporation is completed before submission - [ ] Any additional documents specified on the SPartners application portal EDITOR NOTES - Verify the exact word or character limits for each essay on the SPartners application portal. The limits stated here are fallback values. - Confirm that the Golden Ticket offer is still active and that the December 1st date is correct for the current call cycle. - Eniola should obtain a written expression of interest from the Servier contact in Suresnes before submitting. This strengthens the application significantly. - The venture is not yet incorporated. Confirm whether SPartners requires incorporation at the time of application or allows pre-incorporation applications. - Verify the SKEMPI 3K mutation set size and confirm that the fine-tuning target of AUROC 0.70 is realistic given current performance of 0.634.