Draft — DayOne Accelerator Program

MOTIVATION LETTER

The cost of drug resistance is measured in failed clinical trials and abandoned therapies. When a candidate drug enters development, the question of whether resistance will emerge, and through which mutations, is usually answered too late, after years of investment. My venture predicts drug resistance mutations from protein sequence alone, with no crystal structure required. This directly addresses a bottleneck in pharmaceutical R&D that the DayOne Accelerator Program is designed to tackle.

The technology uses ESM-2 protein language model delta-embeddings combined with ECFP4 drug fingerprints, fed into a Random Forest classifier. On the Platinum benchmark, the model achieves an AUROC of 0.804 plus or minus 0.025 under protein-grouped cross-validation across 553 mutations. This exceeds the published state of the art, mCSM-lig, which reaches approximately 0.70. More importantly, the method achieves 100 percent mutation coverage, while structure-limited tools cover only about 18 percent of mutations because they require a resolved crystal structure. For most clinically relevant proteins, no such structure exists.

The DayOne program is an equity-free, fee-free accelerator focused on healthtech and biotech startups enhancing pharmaceutical R&D. That is precisely the category this venture occupies. The program's selection criteria favor early-stage startups with less than ten million dollars in dilutive funding, which matches my pre-seed stage. The non-dilutive funding of up to CHF 50,000 would directly fund the next validation milestone: fine-tuning ESM-2 on the SKEMPI 3K mutation dataset to reach an AUROC of at least 0.70 on that benchmark. The Basel life sciences ecosystem, with its concentration of pharma companies and research institutes, offers the partnership development potential needed to move toward the Servier pilot.

The venture is currently pre-seed, with proof-of-concept validated on public benchmarks. It is not yet incorporated. The DayOne program's willingness to support early-stage, international teams makes this stage appropriate. The hybrid format, with an in-person bootcamp in Basel and travel support up to CHF 1,000, allows participation without requiring relocation.

The problem is specific, the method is validated, and the next step is clear. DayOne's focus on pharmaceutical R&D innovation and its Basel network are the right environment for this venture to advance toward its first commercial pilot.

RESEARCH STATEMENT

Drug resistance is the primary cause of treatment failure in oncology, infectious disease, and antimicrobial therapy. The standard approach to predicting resistance mutations relies on protein structures obtained through X-ray crystallography or cryo-EM. These methods are slow, expensive, and fail for the majority of clinically relevant proteins. Membrane proteins, intrinsically disordered regions, and proteins that resist crystallization are simply inaccessible. The result is that for most drugs, the resistance landscape is unknown until patients fail therapy.

My venture addresses this gap with a computational method that predicts resistance mutations from protein sequence alone. The approach uses ESM-2, a protein language model pre-trained on millions of sequences, to generate embeddings that capture evolutionary and structural information implicitly. Delta-embeddings, the difference between wild-type and mutant embeddings, are combined with ECFP4 drug fingerprints to represent the interaction between a specific drug and a specific mutation. A Random Forest classifier then predicts whether that mutation confers resistance.

The validation results are concrete. On the Platinum benchmark, which contains 553 mutations across diverse protein-drug pairs, 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 that mutations from the same protein do not appear in both training and test sets, preventing data leakage. The performance exceeds mCSM-lig, the published state of the art, which achieves approximately 0.70. On SKEMPI 2.0, a binding affinity benchmark, the model achieves 0.634, indicating room for improvement that the current roadmap addresses.

The key advantage is coverage. Structure-based tools require a resolved crystal structure for the protein of interest. In practice, this limits them to approximately 18 percent of clinically relevant mutations. My method requires only the protein sequence, which is known for essentially all clinically relevant targets. This is a difference in kind, not a marginal improvement. For a newly identified kinase mutation in a resistant tumor, a structure-based tool may have no answer at all. My method produces a prediction immediately.

The roadmap has three phases. First, fine-tune ESM-2 on the SKEMPI 3K mutation dataset, which contains approximately 3,000 experimentally measured binding affinity changes. The target is an AUROC of at least 0.70 on this benchmark, which would demonstrate generalization beyond the Platinum dataset. Second, engage with Servier in Suresnes for a pilot study on a specific resistance problem in their pipeline. Third, convert the pilot into a recurring revenue arrangement.

The venture is at pre-seed stage. The proof of concept is validated on public benchmarks, but the company is not yet incorporated. The next 12 months are focused on the SKEMPI 3K fine-tuning, the Servier pilot, and incorporation. The DayOne Accelerator Program's non-dilutive funding and Basel network directly support this timeline.

The scientific foundation is published and reproducible. The method is open to independent validation. The gap in the market is clear: pharmaceutical companies need resistance predictions early in the drug development process, and current tools fail for most targets. This venture provides a solution that is faster, broader, and more accurate than the existing alternatives.

EDITOR NOTES

- Selected the drug resistance prediction venture as the single best fit for DayOne because the program explicitly targets healthtech and biotech startups enhancing pharmaceutical R&D. The other research lines in the profile (cardiotoxicity, topology, neurocascade, ergofluids) do not match DayOne's stated pharma R&D focus as directly. The resistance prediction venture has named pharma partners (Servier, Sanofi) and a clear R&D process improvement thesis.

- Eligibility risk: the program targets startups with less than USD 10 million in dilutive funding. The venture is pre-seed with no dilutive funding, which qualifies. However, the venture is not yet incorporated. Verify whether DayOne requires incorporation at application time or only at funding time. If incorporation is required, this must be completed before the deadline.

- Fact verification needed: the AUROC of 0.804 on Platinum and 0.634 on SKEMPI 2.0 are from the applicant profile. The claim that mCSM-lig achieves approximately 0.70 should be verified against the published benchmark paper before submission. The 18 percent mutation coverage figure for structure-limited tools also needs a citable source.

- Gap to fill: the application does not specify a timeline for the Servier pilot. The applicant should insert a concrete target date for initiating the pilot conversation, as DayOne selection committees will likely ask about commercial traction and partnership status.

- The SKEMPI 3K fine-tuning target of AUROC at least 0.70 is stated in the roadmap. The applicant should be prepared to explain why this threshold is meaningful and what happens if it is not met. A contingency plan would strengthen the application.

- Travel support up to CHF 1,000 is available for the Basel bootcamp. The applicant should confirm ability to travel to Basel and include this in the budget if selected.

MOTIVATION LETTER

The cost of drug resistance is measured in failed clinical trials and abandoned therapies. When a candidate drug enters development, the question of whether resistance will emerge, and through which mutations, is usually answered too late, after years of investment. My venture predicts drug resistance mutations from protein sequence alone, with no crystal structure required. This directly addresses a bottleneck in pharmaceutical R&D that the DayOne Accelerator Program is designed to tackle.

The technology uses ESM-2 protein language model delta-embeddings combined with ECFP4 drug fingerprints, fed into a Random Forest classifier. On the Platinum benchmark, the model achieves an AUROC of 0.804 plus or minus 0.025 under protein-grouped cross-validation across 553 mutations. This exceeds the published state of the art, mCSM-lig, which reaches approximately 0.70. More importantly, the method achieves 100 percent mutation coverage, while structure-limited tools cover only about 18 percent of mutations because they require a resolved crystal structure. For most clinically relevant proteins, no such structure exists.

The DayOne program is an equity-free, fee-free accelerator focused on healthtech and biotech startups enhancing pharmaceutical R&D. That is precisely the category this venture occupies. The program's selection criteria favor early-stage startups with less than ten million dollars in dilutive funding, which matches my pre-seed stage. The non-dilutive funding of up to CHF 50,000 would directly fund the next validation milestone: fine-tuning ESM-2 on the SKEMPI 3K mutation dataset to reach an AUROC of at least 0.70 on that benchmark. The Basel life sciences ecosystem, with its concentration of pharma companies and research institutes, offers the partnership development potential needed to move toward the Servier pilot.

The venture is currently pre-seed, with proof-of-concept validated on public benchmarks. It is not yet incorporated. The DayOne program's willingness to support early-stage, international teams makes this stage appropriate. The hybrid format, with an in-person bootcamp in Basel and travel support up to CHF 1,000, allows participation without requiring relocation.

The problem is specific, the method is validated, and the next step is clear. DayOne's focus on pharmaceutical R&D innovation and its Basel network are the right environment for this venture to advance toward its first commercial pilot.

RESEARCH STATEMENT

Drug resistance is the primary cause of treatment failure in oncology, infectious disease, and antimicrobial therapy. The standard approach to predicting resistance mutations relies on protein structures obtained through X-ray crystallography or cryo-EM. These methods are slow, expensive, and fail for the majority of clinically relevant proteins. Membrane proteins, intrinsically disordered regions, and proteins that resist crystallization are simply inaccessible. The result is that for most drugs, the resistance landscape is unknown until patients fail therapy.

My venture addresses this gap with a computational method that predicts resistance mutations from protein sequence alone. The approach uses ESM-2, a protein language model pre-trained on millions of sequences, to generate embeddings that capture evolutionary and structural information implicitly. Delta-embeddings, the difference between wild-type and mutant embeddings, are combined with ECFP4 drug fingerprints to represent the interaction between a specific drug and a specific mutation. A Random Forest classifier then predicts whether that mutation confers resistance.

The validation results are concrete. On the Platinum benchmark, which contains 553 mutations across diverse protein-drug pairs, 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 that mutations from the same protein do not appear in both training and test sets, preventing data leakage. The performance exceeds mCSM-lig, the published state of the art, which achieves approximately 0.70. On SKEMPI 2.0, a binding affinity benchmark, the model achieves 0.634, indicating room for improvement that the current roadmap addresses.

The key advantage is coverage. Structure-based tools require a resolved crystal structure for the protein of interest. In practice, this limits them to approximately 18 percent of clinically relevant mutations. My method requires only the protein sequence, which is known for essentially all clinically relevant targets. This is a difference in kind, not a marginal improvement. For a newly identified kinase mutation in a resistant tumor, a structure-based tool may have no answer at all. My method produces a prediction immediately.

The roadmap has three phases. First, fine-tune ESM-2 on the SKEMPI 3K mutation dataset, which contains approximately 3,000 experimentally measured binding affinity changes. The target is an AUROC of at least 0.70 on this benchmark, which would demonstrate generalization beyond the Platinum dataset. Second, engage with Servier in Suresnes for a pilot study on a specific resistance problem in their pipeline. Third, convert the pilot into a recurring revenue arrangement.

The venture is at pre-seed stage. The proof of concept is validated on public benchmarks, but the company is not yet incorporated. The next 12 months are focused on the SKEMPI 3K fine-tuning, the Servier pilot, and incorporation. The DayOne Accelerator Program's non-dilutive funding and Basel network directly support this timeline.

The scientific foundation is published and reproducible. The method is open to independent validation. The gap in the market is clear: pharmaceutical companies need resistance predictions early in the drug development process, and current tools fail for most targets. This venture provides a solution that is faster, broader, and more accurate than the existing alternatives.

EDITOR NOTES

- Selected the drug resistance prediction venture as the single best fit for DayOne because the program explicitly targets healthtech and biotech startups enhancing pharmaceutical R&D. The other research lines in the profile (cardiotoxicity, topology, neurocascade, ergofluids) do not match DayOne's stated pharma R&D focus as directly. The resistance prediction venture has named pharma partners (Servier, Sanofi) and a clear R&D process improvement thesis.

- Eligibility risk: the program targets startups with less than USD 10 million in dilutive funding. The venture is pre-seed with no dilutive funding, which qualifies. However, the venture is not yet incorporated. Verify whether DayOne requires incorporation at application time or only at funding time. If incorporation is required, this must be completed before the deadline.

- Fact verification needed: the AUROC of 0.804 on Platinum and 0.634 on SKEMPI 2.0 are from the applicant profile. The claim that mCSM-lig achieves approximately 0.70 should be verified against the published benchmark paper before submission. The 18 percent mutation coverage figure for structure-limited tools also needs a citable source.

- Gap to fill: the application does not specify a timeline for the Servier pilot. The applicant should insert a concrete target date for initiating the pilot conversation, as DayOne selection committees will likely ask about commercial traction and partnership status.

- The SKEMPI 3K fine-tuning target of AUROC at least 0.70 is stated in the roadmap. The applicant should be prepared to explain why this threshold is meaningful and what happens if it is not met. A contingency plan would strengthen the application.

- Travel support up to CHF 1,000 is available for the Basel bootcamp. The applicant should confirm ability to travel to Basel and include this in the budget if selected.

AI Draft — DayOne Accelerator Program