Draft — Google Opens Applications for 2026 India AI Startup Accelerator

MOTIVATION LETTER

The 2026 India AI Startup Accelerator selects AI-first startups building specialized models for global use. TOPOLOGIX is exactly that: a sequence-based machine learning system that predicts drug-resistance mutations from protein language model embeddings and chemical fingerprints, without requiring a resolved protein structure. On the Platinum benchmark of 553 mutations, TOPOLOGIX achieves AUROC 0.804 plus or minus 0.025, and it covers 100 percent of mutations tested, versus roughly 18 percent for structure-limited tools like mCSM-lig, which scores about 0.70. That coverage gap matters for real clinical and pharmaceutical decisions, because most resistance-relevant mutations occur in proteins without crystal structures.

TOPOLOGIX emerged from a falsified hypothesis. My earlier work tested whether bipartite persistent homology of protein-ligand interfaces could predict hERG cardiotoxicity. A pre-registered, powered replication showed topological features do not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782. Applying the same topological constructs to drug-resistance prediction produced AUROC values of 0.425 and 0.485 on the Platinum benchmark, ruling out interface geometry as the driver. Those negative results redirected the project toward sequence representations, which is where TOPOLOGIX now stands. The current model uses ESM-2 delta-embeddings, Morgan/ECFP fingerprints, and a Random Forest classifier. This is a venture built on a published, reproducible negative result, not on overclaimed IP.

India is the right launch market for this model. The country manufactures roughly 20 percent of global generic pharmaceuticals by volume and hosts the largest number of US FDA-approved manufacturing plants outside the United States. Antimicrobial and antimalarial resistance surveillance is a stated public health priority under the National Action Plan on Antimicrobial Resistance. A model that predicts resistance mutations from sequence alone, without waiting for structural characterization, directly serves Indian pharma quality control, drug repurposing pipelines, and resistance surveillance programs. I am prepared to establish an Indian entity, identify local collaborators in computational biology and pharmaceutical quality assurance, and commit to the three-month Bengaluru bootcamp and October demo day.

Google's AI ecosystem is the correct infrastructure for scaling TOPOLOGIX. The model currently runs on local hardware with DuckDB-based ingest pipelines and self-hosted LLM serving. Moving training and inference to Google Cloud TPUs, integrating Gemini for automated literature-to-training-data extraction, and using Google's mentorship network to reach Indian pharma partners are concrete, near-term steps. The accelerator's focus areas include specialized AI models for global use, which is the precise category TOPOLOGIX occupies.

I am a licensed pharmacist, B.Pharm from the University of Ibadan with a German equivalent grade of 1.9, currently enrolled in the M.Sc. Digital Health program at the Hasso Plattner Institute and University of Potsdam. I have built four independent DuckDB-based ingest-to-analyze pipelines and production systems infrastructure including Linux VPS, systemd, Caddy TLS, and automated backup recovery. TOPOLOGIX is my venture, and I am seeking the accelerator's resources to scale it from a validated research model to a deployed tool for Indian pharmaceutical and public health users.

RESEARCH STATEMENT

TOPOLOGIX predicts drug-resistance mutations from protein sequence alone. The input is a protein sequence and a drug fingerprint. The model computes ESM-2 protein language model delta-embeddings between wild-type and variant sequences, concatenates those with Morgan/ECFP drug fingerprints, and feeds the combined vector to a Random Forest classifier. No crystal structure, no docking pose, no homology model is required. This design decision is the direct consequence of a measured limitation in structure-based methods: on the Platinum benchmark, structure-dependent tools cover only about 18 percent of mutations because most resistance-relevant proteins lack resolved structures. TOPOLOGIX covers 100 percent.

The measured performance on the Platinum benchmark, 553 mutations, is AUROC 0.804 plus or minus 0.025. On SKEMPI 2.0, the model achieves AUROC 0.634. For comparison, mCSM-lig, a widely used structure-based predictor, scores approximately 0.70 on the subset it can handle. TOPOLOGIX does not beat mCSM-lig on that subset by a wide margin, but it does not need the subset. It operates where structure-based tools cannot. That is the product's core value proposition.

The model's lineage includes a rigorous negative result. My pre-registered replication study on hERG cardiotoxicity tested whether bipartite persistent homology of protein-ligand interfaces predicts toxicity. It does not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782. Applying the same topological methods to drug-resistance prediction produced AUROC 0.425 and 0.485 on Platinum, effectively no signal. These results are published as preprints and are under peer review. They are not failures to be hidden; they are the evidence that motivated the sequence-based approach. TOPOLOGIX exists because interface geometry was tested and ruled out.

The current validation status is honest and specific. The model has been benchmarked on two public datasets, Platinum and SKEMPI 2.0, with pre-registered protocols. It has not yet been validated on prospective clinical or pharmaceutical samples. It has no revenue and no deployed product. What exists is a reproducible, benchmarked model with a clear deployment path.

The deployment path for India is concrete. Indian pharmaceutical manufacturers and resistance surveillance programs need to identify resistance mutations quickly and at scale. The National Action Plan on Antimicrobial Resistance requires surveillance data across human, animal, and environmental sectors. TOPOLOGIX can process sequence data from any of those sectors without structural characterization. The technical integration path is straightforward: a REST API wrapping the Random Forest model, deployed on Google Cloud Run or Cloud TPU-backed serving, with a Supabase/Postgres backend for mutation records and a DuckDB-based pipeline for ingesting new sequence data from public databases or partner submissions.

The research agenda beyond the current benchmark has three steps. First, expand training data beyond Platinum and SKEMPI 2.0 by curating resistance mutations from published literature using LLM-assisted extraction, with human verification. Second, replace the Random Forest with a gradient-boosted or small transformer model trained on the expanded dataset, targeting AUROC above 0.85 on Platinum. Third, run a prospective validation study with an Indian partner laboratory on a defined set of clinically relevant mutations, pre-registering the protocol before data collection begins.

I am the sole founder and sole developer of TOPOLOGIX. My background includes a B.Pharm from the University of Ibadan, PCN licensure as a pharmacist, and current enrollment in the M.Sc. Digital Health program at the Hasso Plattner Institute and University of Potsdam. I have built production systems infrastructure, including Linux VPS operations, systemd services, Caddy TLS, CI/CD pipelines, and automated backup and disaster recovery. I have also built four independent DuckDB-based ingest-to-analyze pipelines across life sciences, technology, and social science domains. The model code is available on GitHub, and the research preprints are on OSF and Zenodo.

ESSAY RESPONSE: WHY THIS ACCELERATOR, WHY NOW

Google's 2026 India AI Startup Accelerator targets AI-first startups building specialized models for global use, with focus areas including agentic AI and sovereign AI. TOPOLOGIX is a specialized AI model for global use, and its most urgent market is India's pharmaceutical sector. The timing is driven by three converging factors.

First, the model is benchmarked and reproducible. TOPOLOGIX has published AUROC scores on two public datasets, a pre-registered protocol, and a clear negative-result lineage that explains why the sequence-based approach is the right one. It is ready for the next stage, which is deployment-oriented validation with a partner laboratory. The accelerator's three-month structure, including the Bengaluru bootcamp and October demo day, provides the deadline and the network to make that partnership happen.

Second, India's pharmaceutical and public health infrastructure has a specific, measurable need. India produces approximately 20 percent of global generic pharmaceuticals by volume. Antimicrobial resistance is a declared national priority. Resistance surveillance requires identifying mutations in pathogens and in drug targets, and doing so fast enough to inform treatment and manufacturing decisions. Structure-based tools fail on most of these mutations because structures are unavailable. TOPOLOGIX does not require structures. That is not a marketing claim; it is a measured coverage difference, 100 percent versus 18 percent on the Platinum benchmark.

Third, Google's AI ecosystem is the right infrastructure for the next technical step. The model currently runs on local hardware. Scaling to a deployed API service, training on an expanded dataset, and integrating LLM-assisted literature curation all benefit from Google Cloud TPUs, Gemini, and the accelerator's technical mentorship. The accelerator's stated focus on specialized AI models for global use matches TOPOLOGIX's category precisely.

I am prepared to commit to the program's requirements, including establishing an Indian entity and attending the in-person bootcamp in Bengaluru. The venture is at the seed stage, which matches the accelerator's stated stage range. There is no revenue yet, and I will not claim otherwise. What exists is a validated model, a clear deployment path, and a founder with the pharmaceutical and software engineering background to execute it.

EDITOR NOTES

- Selected TOPOLOGIX as the venture for this application because it is the only research line in the profile that is a deployable AI model with benchmarked performance, which directly matches the accelerator's AI-first, specialized-model criteria. The CCT model, neurocascade, and ergofluids are research tools, not ventures, and ergofluids explicitly failed its first real-data gate. TOPOLOGIX's negative-result lineage is a strength here because the accelerator values demonstrated rigor.

- Eligibility risk: the programme is India-focused and requires an India-based startup or a plan to incorporate there. The letter states a commitment to establish an Indian entity, but the applicant must verify whether the accelerator requires incorporation before application or permits a post-selection incorporation plan. This is the single largest risk to the application.

- Facts to verify before submission: the claim that India produces approximately 20 percent of global generic pharmaceuticals by volume, and the claim about the number of US FDA-approved manufacturing plants in India. Both are commonly cited but should be checked against current sources. The mCSM-lig AUROC of approximately 0.70 and the Platinum benchmark coverage of 18 percent for structure-limited tools should be re-verified against the applicant's own benchmark runs.

- The applicant must insert personal detail not in the profile: whether they have any existing collaborators or contacts in India, whether they are willing to travel to Bengaluru for the bootcamp given their M.Sc. enrollment at HPI/Potsdam starting Winter Semester 2026/27, and whether they have a preferred Indian state or city for incorporation. The letter currently assumes willingness to travel and incorporate; the applicant must confirm this is true.

- The essay response and motivation letter both state there is no revenue and no deployed product. This is correct per the profile and must not be changed. Do not let a reviewer or the applicant inflate the venture's stage to fit the Seed to Series A criterion.

CHECKLIST

- [ ] Verify current eligibility requirements on the official Google accelerator application page, not the third-party announcement URL
- [ ] Confirm whether India incorporation is required before application or can follow selection
- [ ] Verify the 20 percent generic pharmaceutical production claim and the FDA plant count with current sources
- [ ] Re-check mCSM-lig AUROC and Platinum coverage figures against the applicant's own benchmark outputs
- [ ] Confirm applicant's availability for the Bengaluru bootcamp and October demo day given M.Sc. enrollment dates
- [ ] Identify at least one Indian collaborator or partner organization in pharma or resistance surveillance
- [ ] Decide on Indian state or city for potential incorporation
- [ ] Prepare a one-page technical appendix with TOPOLOGIX architecture diagram and benchmark tables
- [ ] Prepare a short pitch deck (10-15 slides) covering problem, negative-result lineage, model performance, deployment path, and India market rationale
- [ ] Prepare a financial or resource request outline specifying what Google Cloud credits and TPU access would be used for
- [ ] Confirm the application deadline from the official programme page
- [ ] Submit the application through the official Google accelerator portal, not the third-party URL

MOTIVATION LETTER

The 2026 India AI Startup Accelerator selects AI-first startups building specialized models for global use. TOPOLOGIX is exactly that: a sequence-based machine learning system that predicts drug-resistance mutations from protein language model embeddings and chemical fingerprints, without requiring a resolved protein structure. On the Platinum benchmark of 553 mutations, TOPOLOGIX achieves AUROC 0.804 plus or minus 0.025, and it covers 100 percent of mutations tested, versus roughly 18 percent for structure-limited tools like mCSM-lig, which scores about 0.70. That coverage gap matters for real clinical and pharmaceutical decisions, because most resistance-relevant mutations occur in proteins without crystal structures.

TOPOLOGIX emerged from a falsified hypothesis. My earlier work tested whether bipartite persistent homology of protein-ligand interfaces could predict hERG cardiotoxicity. A pre-registered, powered replication showed topological features do not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782. Applying the same topological constructs to drug-resistance prediction produced AUROC values of 0.425 and 0.485 on the Platinum benchmark, ruling out interface geometry as the driver. Those negative results redirected the project toward sequence representations, which is where TOPOLOGIX now stands. The current model uses ESM-2 delta-embeddings, Morgan/ECFP fingerprints, and a Random Forest classifier. This is a venture built on a published, reproducible negative result, not on overclaimed IP.

India is the right launch market for this model. The country manufactures roughly 20 percent of global generic pharmaceuticals by volume and hosts the largest number of US FDA-approved manufacturing plants outside the United States. Antimicrobial and antimalarial resistance surveillance is a stated public health priority under the National Action Plan on Antimicrobial Resistance. A model that predicts resistance mutations from sequence alone, without waiting for structural characterization, directly serves Indian pharma quality control, drug repurposing pipelines, and resistance surveillance programs. I am prepared to establish an Indian entity, identify local collaborators in computational biology and pharmaceutical quality assurance, and commit to the three-month Bengaluru bootcamp and October demo day.

Google's AI ecosystem is the correct infrastructure for scaling TOPOLOGIX. The model currently runs on local hardware with DuckDB-based ingest pipelines and self-hosted LLM serving. Moving training and inference to Google Cloud TPUs, integrating Gemini for automated literature-to-training-data extraction, and using Google's mentorship network to reach Indian pharma partners are concrete, near-term steps. The accelerator's focus areas include specialized AI models for global use, which is the precise category TOPOLOGIX occupies.

I am a licensed pharmacist, B.Pharm from the University of Ibadan with a German equivalent grade of 1.9, currently enrolled in the M.Sc. Digital Health program at the Hasso Plattner Institute and University of Potsdam. I have built four independent DuckDB-based ingest-to-analyze pipelines and production systems infrastructure including Linux VPS, systemd, Caddy TLS, and automated backup recovery. TOPOLOGIX is my venture, and I am seeking the accelerator's resources to scale it from a validated research model to a deployed tool for Indian pharmaceutical and public health users.

RESEARCH STATEMENT

TOPOLOGIX predicts drug-resistance mutations from protein sequence alone. The input is a protein sequence and a drug fingerprint. The model computes ESM-2 protein language model delta-embeddings between wild-type and variant sequences, concatenates those with Morgan/ECFP drug fingerprints, and feeds the combined vector to a Random Forest classifier. No crystal structure, no docking pose, no homology model is required. This design decision is the direct consequence of a measured limitation in structure-based methods: on the Platinum benchmark, structure-dependent tools cover only about 18 percent of mutations because most resistance-relevant proteins lack resolved structures. TOPOLOGIX covers 100 percent.

The measured performance on the Platinum benchmark, 553 mutations, is AUROC 0.804 plus or minus 0.025. On SKEMPI 2.0, the model achieves AUROC 0.634. For comparison, mCSM-lig, a widely used structure-based predictor, scores approximately 0.70 on the subset it can handle. TOPOLOGIX does not beat mCSM-lig on that subset by a wide margin, but it does not need the subset. It operates where structure-based tools cannot. That is the product's core value proposition.

The model's lineage includes a rigorous negative result. My pre-registered replication study on hERG cardiotoxicity tested whether bipartite persistent homology of protein-ligand interfaces predicts toxicity. It does not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782. Applying the same topological methods to drug-resistance prediction produced AUROC 0.425 and 0.485 on Platinum, effectively no signal. These results are published as preprints and are under peer review. They are not failures to be hidden; they are the evidence that motivated the sequence-based approach. TOPOLOGIX exists because interface geometry was tested and ruled out.

The current validation status is honest and specific. The model has been benchmarked on two public datasets, Platinum and SKEMPI 2.0, with pre-registered protocols. It has not yet been validated on prospective clinical or pharmaceutical samples. It has no revenue and no deployed product. What exists is a reproducible, benchmarked model with a clear deployment path.

The deployment path for India is concrete. Indian pharmaceutical manufacturers and resistance surveillance programs need to identify resistance mutations quickly and at scale. The National Action Plan on Antimicrobial Resistance requires surveillance data across human, animal, and environmental sectors. TOPOLOGIX can process sequence data from any of those sectors without structural characterization. The technical integration path is straightforward: a REST API wrapping the Random Forest model, deployed on Google Cloud Run or Cloud TPU-backed serving, with a Supabase/Postgres backend for mutation records and a DuckDB-based pipeline for ingesting new sequence data from public databases or partner submissions.

The research agenda beyond the current benchmark has three steps. First, expand training data beyond Platinum and SKEMPI 2.0 by curating resistance mutations from published literature using LLM-assisted extraction, with human verification. Second, replace the Random Forest with a gradient-boosted or small transformer model trained on the expanded dataset, targeting AUROC above 0.85 on Platinum. Third, run a prospective validation study with an Indian partner laboratory on a defined set of clinically relevant mutations, pre-registering the protocol before data collection begins.

I am the sole founder and sole developer of TOPOLOGIX. My background includes a B.Pharm from the University of Ibadan, PCN licensure as a pharmacist, and current enrollment in the M.Sc. Digital Health program at the Hasso Plattner Institute and University of Potsdam. I have built production systems infrastructure, including Linux VPS operations, systemd services, Caddy TLS, CI/CD pipelines, and automated backup and disaster recovery. I have also built four independent DuckDB-based ingest-to-analyze pipelines across life sciences, technology, and social science domains. The model code is available on GitHub, and the research preprints are on OSF and Zenodo.

ESSAY RESPONSE: WHY THIS ACCELERATOR, WHY NOW

Google's 2026 India AI Startup Accelerator targets AI-first startups building specialized models for global use, with focus areas including agentic AI and sovereign AI. TOPOLOGIX is a specialized AI model for global use, and its most urgent market is India's pharmaceutical sector. The timing is driven by three converging factors.

First, the model is benchmarked and reproducible. TOPOLOGIX has published AUROC scores on two public datasets, a pre-registered protocol, and a clear negative-result lineage that explains why the sequence-based approach is the right one. It is ready for the next stage, which is deployment-oriented validation with a partner laboratory. The accelerator's three-month structure, including the Bengaluru bootcamp and October demo day, provides the deadline and the network to make that partnership happen.

Second, India's pharmaceutical and public health infrastructure has a specific, measurable need. India produces approximately 20 percent of global generic pharmaceuticals by volume. Antimicrobial resistance is a declared national priority. Resistance surveillance requires identifying mutations in pathogens and in drug targets, and doing so fast enough to inform treatment and manufacturing decisions. Structure-based tools fail on most of these mutations because structures are unavailable. TOPOLOGIX does not require structures. That is not a marketing claim; it is a measured coverage difference, 100 percent versus 18 percent on the Platinum benchmark.

Third, Google's AI ecosystem is the right infrastructure for the next technical step. The model currently runs on local hardware. Scaling to a deployed API service, training on an expanded dataset, and integrating LLM-assisted literature curation all benefit from Google Cloud TPUs, Gemini, and the accelerator's technical mentorship. The accelerator's stated focus on specialized AI models for global use matches TOPOLOGIX's category precisely.

I am prepared to commit to the program's requirements, including establishing an Indian entity and attending the in-person bootcamp in Bengaluru. The venture is at the seed stage, which matches the accelerator's stated stage range. There is no revenue yet, and I will not claim otherwise. What exists is a validated model, a clear deployment path, and a founder with the pharmaceutical and software engineering background to execute it.

EDITOR NOTES

- Selected TOPOLOGIX as the venture for this application because it is the only research line in the profile that is a deployable AI model with benchmarked performance, which directly matches the accelerator's AI-first, specialized-model criteria. The CCT model, neurocascade, and ergofluids are research tools, not ventures, and ergofluids explicitly failed its first real-data gate. TOPOLOGIX's negative-result lineage is a strength here because the accelerator values demonstrated rigor.

- Eligibility risk: the programme is India-focused and requires an India-based startup or a plan to incorporate there. The letter states a commitment to establish an Indian entity, but the applicant must verify whether the accelerator requires incorporation before application or permits a post-selection incorporation plan. This is the single largest risk to the application.

- Facts to verify before submission: the claim that India produces approximately 20 percent of global generic pharmaceuticals by volume, and the claim about the number of US FDA-approved manufacturing plants in India. Both are commonly cited but should be checked against current sources. The mCSM-lig AUROC of approximately 0.70 and the Platinum benchmark coverage of 18 percent for structure-limited tools should be re-verified against the applicant's own benchmark runs.

- The applicant must insert personal detail not in the profile: whether they have any existing collaborators or contacts in India, whether they are willing to travel to Bengaluru for the bootcamp given their M.Sc. enrollment at HPI/Potsdam starting Winter Semester 2026/27, and whether they have a preferred Indian state or city for incorporation. The letter currently assumes willingness to travel and incorporate; the applicant must confirm this is true.

- The essay response and motivation letter both state there is no revenue and no deployed product. This is correct per the profile and must not be changed. Do not let a reviewer or the applicant inflate the venture's stage to fit the Seed to Series A criterion.

CHECKLIST

- [ ] Verify current eligibility requirements on the official Google accelerator application page, not the third-party announcement URL
- [ ] Confirm whether India incorporation is required before application or can follow selection
- [ ] Verify the 20 percent generic pharmaceutical production claim and the FDA plant count with current sources
- [ ] Re-check mCSM-lig AUROC and Platinum coverage figures against the applicant's own benchmark outputs
- [ ] Confirm applicant's availability for the Bengaluru bootcamp and October demo day given M.Sc. enrollment dates
- [ ] Identify at least one Indian collaborator or partner organization in pharma or resistance surveillance
- [ ] Decide on Indian state or city for potential incorporation
- [ ] Prepare a one-page technical appendix with TOPOLOGIX architecture diagram and benchmark tables
- [ ] Prepare a short pitch deck (10-15 slides) covering problem, negative-result lineage, model performance, deployment path, and India market rationale
- [ ] Prepare a financial or resource request outline specifying what Google Cloud credits and TPU access would be used for
- [ ] Confirm the application deadline from the official programme page
- [ ] Submit the application through the official Google accelerator portal, not the third-party URL

AI Draft — Google Opens Applications for 2026 India AI Startup Accelerator