Draft — French Tech Grant

MOTIVATION LETTER

The French Tech Grant funds research-led projects with commercial potential, and the venture meets that definition with a validated proof-of-concept and a named industrial pilot. The venture predicts drug resistance mutations from protein sequence alone, using ESM-2 protein language model delta-embeddings combined with ECFP4 drug fingerprints and a Random Forest classifier. No crystal structure is required, which gives the technology 100% mutation coverage on the Platinum benchmark, compared to roughly 18% for structure-limited tools. The classifier achieves an AUROC of 0.804 plus or minus 0.025 on Platinum under protein-grouped cross-validation, and 0.634 on SKEMPI 2.0. Published state of the art, mCSM-lig, reports an AUROC near 0.70 on comparable tasks, so the venture already exceeds that baseline while covering five times more mutations.

The venture is pre-seed, proof-of-concept validated, and not yet incorporated. The roadmap is specific: fine-tune ESM-2 on the SKEMPI 3K mutation set to reach an AUROC of at least 0.70 on that benchmark, then run a pilot with Servier in Suresnes. The pilot targets a concrete ARR figure, which is the commercial milestone that turns this research into revenue. Named partnerships with Paris-Saclay (I2BC), Institut Pasteur, and Sanofi in Gentilly provide the scientific and industrial network needed to execute that roadmap in France.

The founder background is pharmacist-turned-ML-engineer, which means the venture is built by someone who has dispensed drugs and then learned to build the models that predict how those drugs fail. That combination is rare in computational drug discovery, where most teams are strong in one domain and weak in the other. The venture needs both to be credible with Servier and with the French research ecosystem.

The French Tech Grant is the right instrument because it explicitly prioritizes deeptech health innovation and research commercialization. The venture is a hard technical barrier in a regulated industry, with a clear path from academic benchmark to industrial pilot, not a mobile app or a marketplace. The grant amount sought is within the programme range, and the pre-seed stage fits the programme's expectation that the money reduces technical risk and produces measurable proof.

The application logic is simple: the model works on public benchmarks, the next step is fine-tuning on a larger curated dataset, and the success metric is a specific AUROC threshold that unlocks a paid pilot with a named pharmaceutical partner. Every euro requested goes to compute, dataset curation, and the pilot engineering work. The budget ties directly to those milestones.

The venture aligns with French public priorities in health, deeptech, and industrial technology. The technology is developed and will be commercialized in France, with French research partners and a French pharmaceutical pilot. That is the definition of research commercialization the French Tech Grant exists to fund.

RESEARCH STATEMENT

The problem is drug resistance. Every antibiotic, antiviral, and oncology therapy eventually faces mutations that reduce or eliminate its efficacy. Predicting which mutations will cause resistance, before they emerge in the clinic, would let drug developers design molecules that are harder to evade. The current tools require a crystal structure of the target protein, which is unavailable for most clinically relevant proteins. The venture removes that requirement.

The method uses ESM-2, a protein language model trained on millions of sequences, to generate delta-embeddings: the difference between the embedding of the wild-type protein and the embedding of a mutated variant. That delta captures the biophysical effect of the mutation without needing a structure. The delta-embedding is concatenated with an ECFP4 fingerprint of the drug molecule, and the combined vector is classified by a Random Forest as resistant or not resistant.

The validation is public and rigorous. On the Platinum benchmark, 553 mutations with protein-grouped cross-validation, the classifier achieves an AUROC of 0.804 plus or minus 0.025. On SKEMPI 2.0, the AUROC is 0.634. The coverage advantage is structural: because the method needs only sequence, it can score every mutation in a protein, whereas structure-limited tools can only score mutations in proteins with a solved crystal structure. That is 100% coverage versus roughly 18%.

The next technical milestone is fine-tuning ESM-2 on the SKEMPI 3K mutation set. The current model uses a frozen ESM-2 backbone; fine-tuning the language model on binding-affinity data should improve the delta-embeddings specifically for resistance prediction. The target is an AUROC of at least 0.70 on SKEMPI 2.0, which would bring the sequence-only method to parity with structure-based tools while retaining the coverage advantage. That is the threshold that makes the tool commercially credible.

The commercial path runs through Servier in Suresnes. The pilot is designed around a specific use case in their oncology pipeline: given a candidate drug and a target protein, predict the resistance mutations most likely to emerge in a resistance screen. The pilot has a defined deliverable, a ranked list of mutations with confidence scores, and a success metric, agreement with their in-house resistance screening data. The pilot converts to ARR if the predictions reduce the number of mutants they need to generate experimentally.

The scientific partnerships are in place. Paris-Saclay (I2BC) provides structural biology expertise for validating predictions on proteins where structures do exist. Institut Pasteur provides clinical microbiology context for antimicrobial resistance applications. Sanofi in Gentilly is a second potential pilot site, which de-risks the commercial path if the Servier pilot timeline slips.

The intellectual property position is a hard technical barrier: the specific combination of ESM-2 delta-embeddings with drug fingerprints for resistance prediction, validated on Platinum, is not described in the published literature. The barrier is not the individual components, which are public, but the demonstrated performance of the combination, which is the venture's asset.

The budget request covers three items: GPU compute for fine-tuning ESM-2 on SKEMPI 3K, dataset curation and cleaning for the fine-tuning set, and engineering time for the Servier pilot integration. Each item maps to a milestone in the roadmap. The grant reduces the technical risk that the fine-tuned model fails to reach the AUROC threshold, which is the gating event for the pilot.

ESSAY: WHY THIS PROJECT AND WHY NOW

Drug resistance is a moving target, and the current prediction tools are structurally limited. They require crystal structures, which exist for a small fraction of clinically relevant proteins. That means the tools that should predict resistance are blind for most of the proteins where resistance actually matters. The venture solves that with a sequence-only method that covers every mutation in every protein with a sequence.

The timing is driven by two factors. First, ESM-2 and protein language models reached sufficient quality in the last two years that delta-embeddings carry real biophysical signal. Second, the Platinum benchmark provides a public, protein-grouped evaluation that lets a small team prove performance without access to proprietary pharma data. The venture has already done that: AUROC 0.804 on Platinum, beating published state of the art while covering five times more mutations.

The next step is fine-tuning on SKEMPI 3K. That is a defined task with a defined success metric, AUROC at least 0.70 on SKEMPI 2.0. Reaching that threshold unlocks the Servier pilot, which is the first revenue event. The grant funds exactly that step. The project is ready now because the proof-of-concept is done, the benchmark is public, the partner is named, and the roadmap is specific.

ESSAY: COMMERCIAL POTENTIAL AND MARKET FIT

The market is drug resistance prediction in oncology, antivirals, and antimicrobial resistance. The customer is a pharma company running a resistance screen during lead optimization. Today, that screen is experimental: generate mutants, test them, and hope the panel covers the relevant resistance space. The venture replaces part of that screen with a ranked prediction list, so the company tests fewer mutants and finds the resistant ones faster.

The pricing model is a pilot fee followed by an annual subscription for the prediction platform. The Servier pilot is the first test of that model. The ARR target is tied to the pilot outcome: if the predictions match Servier's in-house resistance data, the subscription renews and expands to other programs.

The competitive landscape is structure-based tools like mCSM-lig, which cover only proteins with solved structures. The venture's coverage advantage, 100% versus 18%, is the wedge. The performance advantage, AUROC 0.804 versus 0.70, is the proof. The combination is the commercial argument: better coverage and better accuracy on the mutations that matter.

The venture is pre-seed and not yet incorporated, which is appropriate for this stage. The grant funds the technical milestone that converts the proof-of-concept into a pilot-ready product. The French Tech Grant is the right instrument because it funds exactly this transition, from validated research to commercial engagement, in a French deeptech health context.

CHECKLIST

- [ ] Confirm current French Tech Grant application deadline on the programme website
- [ ] Verify the venture's eligibility for the Founder track as a non-incorporated pre-seed entity
- [ ] Prepare the motivation letter as the primary application document (300-500 words)
- [ ] Prepare the research statement as the technical annex (400-600 words)
- [ ] Prepare the two essay responses (200-350 words each) for the application form
- [ ] Compile the Platinum benchmark results (AUROC 0.804 plus or minus 0.025, 553 mutations, protein-grouped CV) as a one-page technical appendix
- [ ] Compile the SKEMPI 2.0 result (AUROC 0.634) and the SKEMPI 3K fine-tuning plan as a second appendix
- [ ] Obtain a letter of intent or email confirmation from Servier (Suresnes) confirming pilot interest
- [ ] Obtain a letter of support from Paris-Saclay (I2BC) or Institut Pasteur confirming the partnership
- [ ] Prepare a one-page budget table: GPU compute, dataset curation, pilot engineering time, mapped to the fine-tuning and pilot milestones
- [ ] Prepare a one-page roadmap timeline: fine-tuning completion, AUROC threshold, Servier pilot start, ARR target
- [ ] Verify the founder's pharmacist license and ML engineering credentials are documented in the CV
- [ ] Confirm the venture name and legal form to be used in the application (pre-incorporation)
- [ ] Check whether the French Tech Grant requires a French bank account or French entity for disbursement

EDITOR NOTES

- Eligibility risk: the venture is not yet incorporated. The French Tech Grant may require a French legal entity for disbursement. Confirm before submitting, and if required, plan the incorporation timeline around the grant decision date.
- The Servier pilot is named but not confirmed in writing. The application must include a letter of intent or at minimum an email from Servier expressing interest. Without that, the commercial roadmap is a claim, not a commitment.
- The SKEMPI 3K fine-tuning target of AUROC at least 0.70 is a projection, not a result. The application must present it as a milestone to be achieved with grant funding, not as an existing capability.
- The founder's pharmacist background is a differentiator but needs a concrete narrative: where the pharmacy experience meets the ML engineering, and why that combination is necessary for this specific problem. Insert a specific example from the founder's pharmacy or ML work that motivated the venture.
- The budget must be itemized and tied to the milestones. The grant reviewers will check whether the requested amount matches the stated work. Do not submit a lump-sum request.

MOTIVATION LETTER

The French Tech Grant funds research-led projects with commercial potential, and the venture meets that definition with a validated proof-of-concept and a named industrial pilot. The venture predicts drug resistance mutations from protein sequence alone, using ESM-2 protein language model delta-embeddings combined with ECFP4 drug fingerprints and a Random Forest classifier. No crystal structure is required, which gives the technology 100% mutation coverage on the Platinum benchmark, compared to roughly 18% for structure-limited tools. The classifier achieves an AUROC of 0.804 plus or minus 0.025 on Platinum under protein-grouped cross-validation, and 0.634 on SKEMPI 2.0. Published state of the art, mCSM-lig, reports an AUROC near 0.70 on comparable tasks, so the venture already exceeds that baseline while covering five times more mutations.

The venture is pre-seed, proof-of-concept validated, and not yet incorporated. The roadmap is specific: fine-tune ESM-2 on the SKEMPI 3K mutation set to reach an AUROC of at least 0.70 on that benchmark, then run a pilot with Servier in Suresnes. The pilot targets a concrete ARR figure, which is the commercial milestone that turns this research into revenue. Named partnerships with Paris-Saclay (I2BC), Institut Pasteur, and Sanofi in Gentilly provide the scientific and industrial network needed to execute that roadmap in France.

The founder background is pharmacist-turned-ML-engineer, which means the venture is built by someone who has dispensed drugs and then learned to build the models that predict how those drugs fail. That combination is rare in computational drug discovery, where most teams are strong in one domain and weak in the other. The venture needs both to be credible with Servier and with the French research ecosystem.

The French Tech Grant is the right instrument because it explicitly prioritizes deeptech health innovation and research commercialization. The venture is a hard technical barrier in a regulated industry, with a clear path from academic benchmark to industrial pilot, not a mobile app or a marketplace. The grant amount sought is within the programme range, and the pre-seed stage fits the programme's expectation that the money reduces technical risk and produces measurable proof.

The application logic is simple: the model works on public benchmarks, the next step is fine-tuning on a larger curated dataset, and the success metric is a specific AUROC threshold that unlocks a paid pilot with a named pharmaceutical partner. Every euro requested goes to compute, dataset curation, and the pilot engineering work. The budget ties directly to those milestones.

The venture aligns with French public priorities in health, deeptech, and industrial technology. The technology is developed and will be commercialized in France, with French research partners and a French pharmaceutical pilot. That is the definition of research commercialization the French Tech Grant exists to fund.

RESEARCH STATEMENT

The problem is drug resistance. Every antibiotic, antiviral, and oncology therapy eventually faces mutations that reduce or eliminate its efficacy. Predicting which mutations will cause resistance, before they emerge in the clinic, would let drug developers design molecules that are harder to evade. The current tools require a crystal structure of the target protein, which is unavailable for most clinically relevant proteins. The venture removes that requirement.

The method uses ESM-2, a protein language model trained on millions of sequences, to generate delta-embeddings: the difference between the embedding of the wild-type protein and the embedding of a mutated variant. That delta captures the biophysical effect of the mutation without needing a structure. The delta-embedding is concatenated with an ECFP4 fingerprint of the drug molecule, and the combined vector is classified by a Random Forest as resistant or not resistant.

The validation is public and rigorous. On the Platinum benchmark, 553 mutations with protein-grouped cross-validation, the classifier achieves an AUROC of 0.804 plus or minus 0.025. On SKEMPI 2.0, the AUROC is 0.634. The coverage advantage is structural: because the method needs only sequence, it can score every mutation in a protein, whereas structure-limited tools can only score mutations in proteins with a solved crystal structure. That is 100% coverage versus roughly 18%.

The next technical milestone is fine-tuning ESM-2 on the SKEMPI 3K mutation set. The current model uses a frozen ESM-2 backbone; fine-tuning the language model on binding-affinity data should improve the delta-embeddings specifically for resistance prediction. The target is an AUROC of at least 0.70 on SKEMPI 2.0, which would bring the sequence-only method to parity with structure-based tools while retaining the coverage advantage. That is the threshold that makes the tool commercially credible.

The commercial path runs through Servier in Suresnes. The pilot is designed around a specific use case in their oncology pipeline: given a candidate drug and a target protein, predict the resistance mutations most likely to emerge in a resistance screen. The pilot has a defined deliverable, a ranked list of mutations with confidence scores, and a success metric, agreement with their in-house resistance screening data. The pilot converts to ARR if the predictions reduce the number of mutants they need to generate experimentally.

The scientific partnerships are in place. Paris-Saclay (I2BC) provides structural biology expertise for validating predictions on proteins where structures do exist. Institut Pasteur provides clinical microbiology context for antimicrobial resistance applications. Sanofi in Gentilly is a second potential pilot site, which de-risks the commercial path if the Servier pilot timeline slips.

The intellectual property position is a hard technical barrier: the specific combination of ESM-2 delta-embeddings with drug fingerprints for resistance prediction, validated on Platinum, is not described in the published literature. The barrier is not the individual components, which are public, but the demonstrated performance of the combination, which is the venture's asset.

The budget request covers three items: GPU compute for fine-tuning ESM-2 on SKEMPI 3K, dataset curation and cleaning for the fine-tuning set, and engineering time for the Servier pilot integration. Each item maps to a milestone in the roadmap. The grant reduces the technical risk that the fine-tuned model fails to reach the AUROC threshold, which is the gating event for the pilot.

ESSAY: WHY THIS PROJECT AND WHY NOW

Drug resistance is a moving target, and the current prediction tools are structurally limited. They require crystal structures, which exist for a small fraction of clinically relevant proteins. That means the tools that should predict resistance are blind for most of the proteins where resistance actually matters. The venture solves that with a sequence-only method that covers every mutation in every protein with a sequence.

The timing is driven by two factors. First, ESM-2 and protein language models reached sufficient quality in the last two years that delta-embeddings carry real biophysical signal. Second, the Platinum benchmark provides a public, protein-grouped evaluation that lets a small team prove performance without access to proprietary pharma data. The venture has already done that: AUROC 0.804 on Platinum, beating published state of the art while covering five times more mutations.

The next step is fine-tuning on SKEMPI 3K. That is a defined task with a defined success metric, AUROC at least 0.70 on SKEMPI 2.0. Reaching that threshold unlocks the Servier pilot, which is the first revenue event. The grant funds exactly that step. The project is ready now because the proof-of-concept is done, the benchmark is public, the partner is named, and the roadmap is specific.

ESSAY: COMMERCIAL POTENTIAL AND MARKET FIT

The market is drug resistance prediction in oncology, antivirals, and antimicrobial resistance. The customer is a pharma company running a resistance screen during lead optimization. Today, that screen is experimental: generate mutants, test them, and hope the panel covers the relevant resistance space. The venture replaces part of that screen with a ranked prediction list, so the company tests fewer mutants and finds the resistant ones faster.

The pricing model is a pilot fee followed by an annual subscription for the prediction platform. The Servier pilot is the first test of that model. The ARR target is tied to the pilot outcome: if the predictions match Servier's in-house resistance data, the subscription renews and expands to other programs.

The competitive landscape is structure-based tools like mCSM-lig, which cover only proteins with solved structures. The venture's coverage advantage, 100% versus 18%, is the wedge. The performance advantage, AUROC 0.804 versus 0.70, is the proof. The combination is the commercial argument: better coverage and better accuracy on the mutations that matter.

The venture is pre-seed and not yet incorporated, which is appropriate for this stage. The grant funds the technical milestone that converts the proof-of-concept into a pilot-ready product. The French Tech Grant is the right instrument because it funds exactly this transition, from validated research to commercial engagement, in a French deeptech health context.

CHECKLIST

- [ ] Confirm current French Tech Grant application deadline on the programme website
- [ ] Verify the venture's eligibility for the Founder track as a non-incorporated pre-seed entity
- [ ] Prepare the motivation letter as the primary application document (300-500 words)
- [ ] Prepare the research statement as the technical annex (400-600 words)
- [ ] Prepare the two essay responses (200-350 words each) for the application form
- [ ] Compile the Platinum benchmark results (AUROC 0.804 plus or minus 0.025, 553 mutations, protein-grouped CV) as a one-page technical appendix
- [ ] Compile the SKEMPI 2.0 result (AUROC 0.634) and the SKEMPI 3K fine-tuning plan as a second appendix
- [ ] Obtain a letter of intent or email confirmation from Servier (Suresnes) confirming pilot interest
- [ ] Obtain a letter of support from Paris-Saclay (I2BC) or Institut Pasteur confirming the partnership
- [ ] Prepare a one-page budget table: GPU compute, dataset curation, pilot engineering time, mapped to the fine-tuning and pilot milestones
- [ ] Prepare a one-page roadmap timeline: fine-tuning completion, AUROC threshold, Servier pilot start, ARR target
- [ ] Verify the founder's pharmacist license and ML engineering credentials are documented in the CV
- [ ] Confirm the venture name and legal form to be used in the application (pre-incorporation)
- [ ] Check whether the French Tech Grant requires a French bank account or French entity for disbursement

EDITOR NOTES

- Eligibility risk: the venture is not yet incorporated. The French Tech Grant may require a French legal entity for disbursement. Confirm before submitting, and if required, plan the incorporation timeline around the grant decision date.
- The Servier pilot is named but not confirmed in writing. The application must include a letter of intent or at minimum an email from Servier expressing interest. Without that, the commercial roadmap is a claim, not a commitment.
- The SKEMPI 3K fine-tuning target of AUROC at least 0.70 is a projection, not a result. The application must present it as a milestone to be achieved with grant funding, not as an existing capability.
- The founder's pharmacist background is a differentiator but needs a concrete narrative: where the pharmacy experience meets the ML engineering, and why that combination is necessary for this specific problem. Insert a specific example from the founder's pharmacy or ML work that motivated the venture.
- The budget must be itemized and tied to the milestones. The grant reviewers will check whether the requested amount matches the stated work. Do not submit a lump-sum request.

AI Draft — French Tech Grant