Draft — Single-Cell Biology Data Insights

MOTIVATION LETTER

A single-cell transcriptome from the nucleus accumbens of a rodent after cocaine self-administration contains roughly 50,000 cells and 20 million unique molecular identifiers. The field has standard pipelines for clustering and differential expression, but the step from those clusters to a mechanistic, testable model of how cell-type-specific gene expression changes produce circuit-level and behavioral phenotypes remains largely manual and ad-hoc. My research over the past three years has focused on building exactly this kind of bridge: linking molecular events to circuit dynamics to behavior using Bayesian-calibrated ordinary differential equation models, protein-language-model embeddings, and scalable data infrastructure.

I am a 29-year-old independent computational researcher from Nigeria, currently enrolled in the M.Sc. Digital Health programme at the Hasso Plattner Institute in Germany. My primary research lines include the Conjunctive Consolidation Threshold model, a tripartite pharmacological framework for reward-memory encoding in addiction that confirmed all five pre-registered hypotheses using Bayesian MCMC calibration on 14 free parameters with literature-elicited priors from an 1,847-record screen. I also built neurocascade, a receptor-to-behavior brain-circuit simulation engine that couples pharmacokinetics to receptor-binding to Wilson-Cowan circuit dynamics, with 62 of 62 tests passing. On the protein ML side, I developed TOPOLOGIX, which uses ESM-2 delta-embeddings plus Morgan fingerprints and a Random Forest classifier to predict drug-resistance mutations from sequence alone, achieving an AUROC of 0.804 on the Platinum benchmark while covering 100 percent of mutations versus roughly 18 percent for structure-limited tools.

For the Chan Zuckerberg Initiative Single-Cell Biology Data Insights programme, I propose to build a computational tool that takes single-nucleus RNA-seq data from reward-related brain regions and outputs cell-type-specific, parameterized ODE models of the dopamine and NMDA receptor systems that drive addiction-related plasticity. The tool will use my existing DuckDB-based ingest-to-analyze pipeline architecture to handle datasets of up to one million cells, apply Bayesian calibration (PyMC, DEMetropolisZ) to fit receptor-expression-level parameters to each cell cluster, and generate a ranked list of drug targets per cell type. This directly addresses the programme goal of gaining insights into health and disease from single-cell data by converting static expression profiles into dynamic, testable predictions about circuit-level pharmacology. All code will be released open-source under a BSD-3 license on GitHub, with full documentation and a tutorial notebook using publicly available data from the Gene Expression Omnibus. My track record of pre-registered, fully-reproducible research and my commitment to open science align with CZI mission.

RESEARCH STATEMENT

The central problem this project addresses is the gap between single-cell transcriptomic data and mechanistic, dynamical models of disease. Current single-cell analysis pipelines excel at identifying cell types and differentially expressed genes, but they do not tell you how a 1.7-fold increase in Drd2 expression in D2-medium spiny neurons alters the firing rate of those neurons, or how that firing-rate change propagates through a Wilson-Cowan circuit to produce a measurable behavioral phenotype like cue-induced reinstatement of drug seeking. My proposed tool, which I call sc2ODE, will close this gap.

The technical approach has three layers. First, a data-ingestion layer built on DuckDB and Nextflow that reads 10x Genomics HDF5 or h5ad files, performs standard quality control and normalization, and clusters cells using a Leiden algorithm. Second, a model-calibration layer that takes the mean expression level of each receptor gene per cluster and uses those values to set the receptor-density parameters in a pre-built ODE model of the mesolimbic dopamine circuit. The ODE model is a simplified version of my neurocascade engine: three coupled Wilson-Cowan populations representing ventral tegmental area, nucleus accumbens, and prefrontal cortex, with dopamine D1 and D2 receptor densities, NMDA receptor conductance, and GABA-A receptor conductance as free parameters. Bayesian calibration using PyMC DEMetropolisZ will fit these parameters to match published electrophysiological firing-rate ranges from the Allen Brain Atlas and the literature. Third, an output layer that generates a ranked list of drug targets per cell type, where the ranking is based on the sensitivity of the circuit-level firing rate to a 50 percent reduction in each receptor parameter, computed via local sensitivity analysis.

The innovation is that this tool converts a static, descriptive dataset into a dynamical, predictive model. A researcher studying addiction will be able to take their snRNA-seq data from the nucleus accumbens of a cocaine-exposed mouse, run it through sc2ODE, and get a prediction that D2-MSNs in the shell region are 3.2 times more sensitive to a D2 antagonist than D2-MSNs in the core, a prediction that can be tested with optogenetics or chemogenetics. The tool will be validated on three publicly available datasets: a mouse cocaine self-administration dataset from GEO, a human postmortem nucleus accumbens dataset from the Lieber Institute, and a rat alcohol self-administration dataset from the NIAAA. For each dataset, I will compare the model-predicted drug-target rankings against published pharmacological and behavioral data.

The broader impact is that this framework generalizes beyond addiction. The same pipeline architecture, with a different ODE model swapped in, can be applied to any brain region and any disease where receptor-expression data and circuit-level dynamics are relevant: Parkinson disease, schizophrenia, chronic pain. The open-source release will include a plugin interface so that other researchers can contribute their own ODE models. The project timeline is 12 months: months 1-3 for pipeline development and testing on synthetic data, months 4-6 for calibration on the three validation datasets, months 7-9 for sensitivity analysis and target ranking, months 10-12 for documentation, tutorial writing, and community dissemination via a workshop at the Society for Neuroscience annual meeting.

SHORT ESSAY: RELEVANCE TO CZI MISSION AND OPEN SCIENCE

The Chan Zuckerberg Initiative mission centers on open science and equity. My project directly serves both pillars. On open science, every line of code will be released under a BSD-3 license on GitHub, with a DOI-assigned Zenodo archive. The DuckDB-based pipeline architecture is designed to run on a single laptop with 16 GB of RAM, removing the need for institutional HPC access. The tutorial notebook will use a publicly available dataset from GEO, so any researcher anywhere can reproduce the entire analysis from raw data to final model predictions without paying for software or data. On equity, I am a Nigerian researcher who has built my entire computational toolkit as an independent researcher without a PhD or a faculty position. I know firsthand that the most expensive barrier to entry in computational biology is not the hardware but the proprietary software and the undocumented, fragile pipelines. My tool will be dependency-light: Python, DuckDB, PyMC, and scipy. No Docker container required, no cloud credits needed. The documentation will be written in plain English with worked examples, and I will host a monthly office-hours session via video call for any researcher in Africa or other LMIC regions who wants to use the tool. This is not a side project; it is the core of how I do research.

SHORT ESSAY: TECHNICAL FEASIBILITY AND TRACK RECORD

I have already built and validated the three technical components that sc2ODE requires. The DuckDB-based ingest-to-analyze pipeline architecture is running in production for four independent domains: life sciences, tech/AI security, social science, and a corpus of 1,847 addiction-neuroscience papers that I screened for the CCT model priors. The neurocascade ODE engine has 62 of 62 tests passing, with Bayesian-calibrated parameters for mu-opioid, D2 dopamine, and GABA-A receptor systems. The TOPOLOGIX protein ML pipeline achieved an AUROC of 0.804 on the Platinum benchmark using ESM-2 embeddings and a Random Forest classifier, demonstrating my ability to combine sequence representations with classical ML in a reproducible, open-source framework. The risk is that single-cell data has unique noise properties that my existing pipelines have not been tested against. I mitigate this by using the three validation datasets as gates: if the tool fails to produce biologically plausible parameter fits on the first dataset, I will report that result directly rather than reframing it, exactly as I did with the ergofluids project where the first real-data gate did not meet its pre-registered criterion. I have a pre-registration template ready on OSF and will register the full analysis plan before touching any validation data.

BUDGET NARRATIVE

The requested funds will support 12 months of full-time research. Personnel: 50,000 USD for a 12-month stipend at the independent-researcher rate consistent with NIH NRSA F32 guidelines. Equipment: 2,500 USD for a dedicated GPU workstation (NVIDIA RTX 4090, 64 GB RAM) to run ESM-2 embedding generation and Bayesian MCMC chains in parallel. Software and data: 500 USD for cloud compute credits on AWS for large-scale validation runs. Travel and dissemination: 2,000 USD for registration and travel to the Society for Neuroscience annual meeting to present results and host a workshop. Publication fees: 1,000 USD for open-access publication in a journal like PLOS Computational Biology. Overhead: 0 USD, as I am an independent researcher with no institutional overhead. Total: 56,000 USD.

CHECKLIST

- [ ] Complete CZI online application form at https://chanzuckerberg.com/science/science-funding/
- [ ] Upload motivation letter as PDF, 500 words maximum
- [ ] Upload research statement as PDF, 600 words maximum
- [ ] Upload short essay on CZI mission and open science as PDF, 350 words maximum
- [ ] Upload short essay on technical feasibility and track record as PDF, 350 words maximum
- [ ] Upload budget narrative as PDF, 300 words maximum
- [ ] Upload CV including ORCID, GitHub, publications, and preprints
- [ ] Upload two letters of recommendation: one from Kent Berridge (Michigan), one from Samuel Gershman (Harvard)
- [ ] Verify eligibility: confirm that independent researchers without a PhD are eligible for this track
- [ ] Verify deadline on programme website; no deadline was captured in the profile

EDITOR NOTES

- Eligibility risk: The CZI programme page should be checked to confirm that independent researchers without a PhD or institutional affiliation are eligible to apply. The profile lists Eniola as an independent researcher with a B.Pharm and enrollment in an M.Sc. programme, which may not meet a PhD-required criterion.
- Verification needed: The three validation datasets (mouse cocaine self-administration from GEO, human postmortem nucleus accumbens from Lieber Institute, rat alcohol self-administration from NIAAA) need specific GEO accession numbers or database identifiers inserted before submission.
- Gap: The budget narrative assumes a 12-month stipend at 50,000 USD, but the CZI programme may have a fixed award amount or a maximum per year. The applicant should check the programme page for budget limits and adjust the narrative accordingly.
- Gap: The applicant should confirm that Kent Berridge and Samuel Gershman are willing to provide letters of recommendation, and that those letters can be submitted by the deadline.
- Fact check: The profile states the CCT model confirmed all five pre-registered hypotheses, but the preprint links on OSF/Zenodo should be verified to ensure they are publicly accessible and that the pre-registrations are time-stamped before the analysis was run.

MOTIVATION LETTER

A single-cell transcriptome from the nucleus accumbens of a rodent after cocaine self-administration contains roughly 50,000 cells and 20 million unique molecular identifiers. The field has standard pipelines for clustering and differential expression, but the step from those clusters to a mechanistic, testable model of how cell-type-specific gene expression changes produce circuit-level and behavioral phenotypes remains largely manual and ad-hoc. My research over the past three years has focused on building exactly this kind of bridge: linking molecular events to circuit dynamics to behavior using Bayesian-calibrated ordinary differential equation models, protein-language-model embeddings, and scalable data infrastructure.

I am a 29-year-old independent computational researcher from Nigeria, currently enrolled in the M.Sc. Digital Health programme at the Hasso Plattner Institute in Germany. My primary research lines include the Conjunctive Consolidation Threshold model, a tripartite pharmacological framework for reward-memory encoding in addiction that confirmed all five pre-registered hypotheses using Bayesian MCMC calibration on 14 free parameters with literature-elicited priors from an 1,847-record screen. I also built neurocascade, a receptor-to-behavior brain-circuit simulation engine that couples pharmacokinetics to receptor-binding to Wilson-Cowan circuit dynamics, with 62 of 62 tests passing. On the protein ML side, I developed TOPOLOGIX, which uses ESM-2 delta-embeddings plus Morgan fingerprints and a Random Forest classifier to predict drug-resistance mutations from sequence alone, achieving an AUROC of 0.804 on the Platinum benchmark while covering 100 percent of mutations versus roughly 18 percent for structure-limited tools.

For the Chan Zuckerberg Initiative Single-Cell Biology Data Insights programme, I propose to build a computational tool that takes single-nucleus RNA-seq data from reward-related brain regions and outputs cell-type-specific, parameterized ODE models of the dopamine and NMDA receptor systems that drive addiction-related plasticity. The tool will use my existing DuckDB-based ingest-to-analyze pipeline architecture to handle datasets of up to one million cells, apply Bayesian calibration (PyMC, DEMetropolisZ) to fit receptor-expression-level parameters to each cell cluster, and generate a ranked list of drug targets per cell type. This directly addresses the programme goal of gaining insights into health and disease from single-cell data by converting static expression profiles into dynamic, testable predictions about circuit-level pharmacology. All code will be released open-source under a BSD-3 license on GitHub, with full documentation and a tutorial notebook using publicly available data from the Gene Expression Omnibus. My track record of pre-registered, fully-reproducible research and my commitment to open science align with CZI mission.

RESEARCH STATEMENT

The central problem this project addresses is the gap between single-cell transcriptomic data and mechanistic, dynamical models of disease. Current single-cell analysis pipelines excel at identifying cell types and differentially expressed genes, but they do not tell you how a 1.7-fold increase in Drd2 expression in D2-medium spiny neurons alters the firing rate of those neurons, or how that firing-rate change propagates through a Wilson-Cowan circuit to produce a measurable behavioral phenotype like cue-induced reinstatement of drug seeking. My proposed tool, which I call sc2ODE, will close this gap.

The technical approach has three layers. First, a data-ingestion layer built on DuckDB and Nextflow that reads 10x Genomics HDF5 or h5ad files, performs standard quality control and normalization, and clusters cells using a Leiden algorithm. Second, a model-calibration layer that takes the mean expression level of each receptor gene per cluster and uses those values to set the receptor-density parameters in a pre-built ODE model of the mesolimbic dopamine circuit. The ODE model is a simplified version of my neurocascade engine: three coupled Wilson-Cowan populations representing ventral tegmental area, nucleus accumbens, and prefrontal cortex, with dopamine D1 and D2 receptor densities, NMDA receptor conductance, and GABA-A receptor conductance as free parameters. Bayesian calibration using PyMC DEMetropolisZ will fit these parameters to match published electrophysiological firing-rate ranges from the Allen Brain Atlas and the literature. Third, an output layer that generates a ranked list of drug targets per cell type, where the ranking is based on the sensitivity of the circuit-level firing rate to a 50 percent reduction in each receptor parameter, computed via local sensitivity analysis.

The innovation is that this tool converts a static, descriptive dataset into a dynamical, predictive model. A researcher studying addiction will be able to take their snRNA-seq data from the nucleus accumbens of a cocaine-exposed mouse, run it through sc2ODE, and get a prediction that D2-MSNs in the shell region are 3.2 times more sensitive to a D2 antagonist than D2-MSNs in the core, a prediction that can be tested with optogenetics or chemogenetics. The tool will be validated on three publicly available datasets: a mouse cocaine self-administration dataset from GEO, a human postmortem nucleus accumbens dataset from the Lieber Institute, and a rat alcohol self-administration dataset from the NIAAA. For each dataset, I will compare the model-predicted drug-target rankings against published pharmacological and behavioral data.

The broader impact is that this framework generalizes beyond addiction. The same pipeline architecture, with a different ODE model swapped in, can be applied to any brain region and any disease where receptor-expression data and circuit-level dynamics are relevant: Parkinson disease, schizophrenia, chronic pain. The open-source release will include a plugin interface so that other researchers can contribute their own ODE models. The project timeline is 12 months: months 1-3 for pipeline development and testing on synthetic data, months 4-6 for calibration on the three validation datasets, months 7-9 for sensitivity analysis and target ranking, months 10-12 for documentation, tutorial writing, and community dissemination via a workshop at the Society for Neuroscience annual meeting.

SHORT ESSAY: RELEVANCE TO CZI MISSION AND OPEN SCIENCE

The Chan Zuckerberg Initiative mission centers on open science and equity. My project directly serves both pillars. On open science, every line of code will be released under a BSD-3 license on GitHub, with a DOI-assigned Zenodo archive. The DuckDB-based pipeline architecture is designed to run on a single laptop with 16 GB of RAM, removing the need for institutional HPC access. The tutorial notebook will use a publicly available dataset from GEO, so any researcher anywhere can reproduce the entire analysis from raw data to final model predictions without paying for software or data. On equity, I am a Nigerian researcher who has built my entire computational toolkit as an independent researcher without a PhD or a faculty position. I know firsthand that the most expensive barrier to entry in computational biology is not the hardware but the proprietary software and the undocumented, fragile pipelines. My tool will be dependency-light: Python, DuckDB, PyMC, and scipy. No Docker container required, no cloud credits needed. The documentation will be written in plain English with worked examples, and I will host a monthly office-hours session via video call for any researcher in Africa or other LMIC regions who wants to use the tool. This is not a side project; it is the core of how I do research.

SHORT ESSAY: TECHNICAL FEASIBILITY AND TRACK RECORD

I have already built and validated the three technical components that sc2ODE requires. The DuckDB-based ingest-to-analyze pipeline architecture is running in production for four independent domains: life sciences, tech/AI security, social science, and a corpus of 1,847 addiction-neuroscience papers that I screened for the CCT model priors. The neurocascade ODE engine has 62 of 62 tests passing, with Bayesian-calibrated parameters for mu-opioid, D2 dopamine, and GABA-A receptor systems. The TOPOLOGIX protein ML pipeline achieved an AUROC of 0.804 on the Platinum benchmark using ESM-2 embeddings and a Random Forest classifier, demonstrating my ability to combine sequence representations with classical ML in a reproducible, open-source framework. The risk is that single-cell data has unique noise properties that my existing pipelines have not been tested against. I mitigate this by using the three validation datasets as gates: if the tool fails to produce biologically plausible parameter fits on the first dataset, I will report that result directly rather than reframing it, exactly as I did with the ergofluids project where the first real-data gate did not meet its pre-registered criterion. I have a pre-registration template ready on OSF and will register the full analysis plan before touching any validation data.

BUDGET NARRATIVE

The requested funds will support 12 months of full-time research. Personnel: 50,000 USD for a 12-month stipend at the independent-researcher rate consistent with NIH NRSA F32 guidelines. Equipment: 2,500 USD for a dedicated GPU workstation (NVIDIA RTX 4090, 64 GB RAM) to run ESM-2 embedding generation and Bayesian MCMC chains in parallel. Software and data: 500 USD for cloud compute credits on AWS for large-scale validation runs. Travel and dissemination: 2,000 USD for registration and travel to the Society for Neuroscience annual meeting to present results and host a workshop. Publication fees: 1,000 USD for open-access publication in a journal like PLOS Computational Biology. Overhead: 0 USD, as I am an independent researcher with no institutional overhead. Total: 56,000 USD.

CHECKLIST

- [ ] Complete CZI online application form at https://chanzuckerberg.com/science/science-funding/
- [ ] Upload motivation letter as PDF, 500 words maximum
- [ ] Upload research statement as PDF, 600 words maximum
- [ ] Upload short essay on CZI mission and open science as PDF, 350 words maximum
- [ ] Upload short essay on technical feasibility and track record as PDF, 350 words maximum
- [ ] Upload budget narrative as PDF, 300 words maximum
- [ ] Upload CV including ORCID, GitHub, publications, and preprints
- [ ] Upload two letters of recommendation: one from Kent Berridge (Michigan), one from Samuel Gershman (Harvard)
- [ ] Verify eligibility: confirm that independent researchers without a PhD are eligible for this track
- [ ] Verify deadline on programme website; no deadline was captured in the profile

EDITOR NOTES

- Eligibility risk: The CZI programme page should be checked to confirm that independent researchers without a PhD or institutional affiliation are eligible to apply. The profile lists Eniola as an independent researcher with a B.Pharm and enrollment in an M.Sc. programme, which may not meet a PhD-required criterion.
- Verification needed: The three validation datasets (mouse cocaine self-administration from GEO, human postmortem nucleus accumbens from Lieber Institute, rat alcohol self-administration from NIAAA) need specific GEO accession numbers or database identifiers inserted before submission.
- Gap: The budget narrative assumes a 12-month stipend at 50,000 USD, but the CZI programme may have a fixed award amount or a maximum per year. The applicant should check the programme page for budget limits and adjust the narrative accordingly.
- Gap: The applicant should confirm that Kent Berridge and Samuel Gershman are willing to provide letters of recommendation, and that those letters can be submitted by the deadline.
- Fact check: The profile states the CCT model confirmed all five pre-registered hypotheses, but the preprint links on OSF/Zenodo should be verified to ensure they are publicly accessible and that the pre-registrations are time-stamped before the analysis was run.

AI Draft — Single-Cell Biology Data Insights