Draft — AI2050 Fellows

MOTIVATION LETTER

The AI2050 Fellows program asks for research that could make AI hugely beneficial to society by 2050. Addiction is projected to remain one of the largest global disease burdens, with relapse rates above 60 percent within one year across substance classes. My Conjunctive Consolidation Threshold model, or CCT, addresses this problem directly by formalizing how reward-memory encoding can be pharmacologically prevented at the circuit level. The model couples three axes, dopaminergic reward prediction error, NMDAR-dependent long-term potentiation, and affective contrast, into a single ODE system solved with RK45 and calibrated through Bayesian MCMC using PyMC's DEMetropolisZ sampler. All fourteen free parameters were constrained by priors elicited from a systematic screen of 1,847 records in the literature. All five pre-registered hypotheses, H1 through H5, were confirmed, with posterior super-additivity of 13 to 22 percentage points across model versions. Three sole-authored preprints are currently under review at peer-reviewed journals, IART, PNPBP, and NBR, with a co-authored paper under review at Alcohol, Elsevier.

The AI2050 mission explicitly names major scientific questions and risks as hard problems. The CCT model sits at the intersection of two of them: the scientific question of how memory systems encode reward salience, and the technical challenge of building predictive dynamical models from sparse biological data. My approach treats the problem as a Bayesian inference task rather than a curve-fitting exercise. The model generates posterior distributions over pharmacological intervention parameters, which means it can predict not just whether an intervention works, but how much of each receptor system must be modulated to cross the consolidation threshold. This is a fundamentally different use of AI than pattern recognition on large datasets. It is mechanistic, interpretable, and falsifiable.

The Africa angle matters for AI2050's global mandate. Nigeria has one of the highest untreated substance use disorder burdens in West Africa, yet no computational pharmacology research group exists in the country. I am a licensed pharmacist trained at the University of Ibadan, currently enrolled in the M.Sc. Digital Health program at the Hasso Plattner Institute and the University of Potsdam. My independent research track includes the TOPOLOGIX project, which uses ESM-2 protein language model delta-embeddings combined with Morgan fingerprints and a Random Forest classifier to predict drug resistance mutations from sequence alone, achieving AUROC 0.804 on the Platinum benchmark and 0.634 on SKEMPI 2.0, while covering 100 percent of mutations compared to roughly 18 percent for structure-limited tools. This breadth, from molecular embeddings to circuit-level dynamical systems, is rare in early-career applicants.

AI2050's support would allow me to scale the CCT model from a calibrated theoretical framework to a validated clinical decision tool. The next phase requires fitting the circuit-layer parameters to real behavioral data, which is explicitly labeled as pending in my current neurocascade simulation engine. That engine couples pharmacokinetics to receptor binding to Wilson-Cowan circuit dynamics to behavioral readouts across three receptor systems, mu-opioid, D2 dopamine, and GABA-A, with 62 of 62 tests passing. The fellowship would fund the data acquisition and validation work that transforms this from a simulation platform into a predictive instrument for personalized addiction pharmacotherapy. By 2050, I intend for CCT-derived models to inform dosing protocols in clinical settings across low- and middle-income countries, where the need is greatest and computational psychiatry resources are scarcest.

RESEARCH STATEMENT

The hard problem I address is the prevention of pathological reward-memory consolidation in addiction. Current pharmacotherapy for substance use disorders targets receptor occupancy, not memory encoding. This is a fundamental mismatch. Addiction is maintained by drug-context associations that persist for decades, and no existing treatment directly interrupts the consolidation of those associations at the moment of exposure. The CCT model formalizes the conditions under which pharmacological intervention can prevent reward-memory encoding, providing a mechanistic, quantitative framework for a clinical problem that has resisted purely empirical approaches.

The model architecture is a tripartite coupled system. The first axis tracks dopaminergic reward prediction error, the second tracks NMDAR-dependent long-term potentiation in the mesolimbic pathway, and the third tracks affective contrast, the hedonic differential between drug and non-drug states. These three axes are coupled through a system of ordinary differential equations solved with RK45. The model was calibrated using Bayesian MCMC with the DEMetropolisZ sampler in PyMC, with fourteen free parameters constrained by literature-elicited priors. The prior elicitation process itself was systematic: a screen of 1,847 records from the addiction neuroscience literature, from which I extracted quantitative ranges for receptor binding affinities, synaptic plasticity time constants, and dopaminergic firing rates. This is a literature-grounded, quantitatively constrained dynamical system, not a toy model.

All five pre-registered hypotheses were confirmed. The key finding is super-additivity: combined modulation of dopaminergic and glutamatergic systems produces 13 to 22 percentage points greater consolidation prevention than the sum of individual effects. This has direct clinical implications. It suggests that combination pharmacotherapy targeting both the RPE signal and the LTP gate is not merely additive but synergistic, and that the synergy is predictable from the model's posterior distribution. The model also generates testable predictions about the timing of intervention. The consolidation threshold is crossed within a specific temporal window after exposure, and the model predicts the boundaries of that window with quantified uncertainty.

The next phase, which AI2050 support would enable, is empirical validation. The neurocascade engine, my receptor-to-behavior simulation platform, currently passes 62 of 62 unit tests across three receptor systems. The circuit-layer parameters are explicitly labeled illustrative pending fits to real behavioral data. The fellowship would fund the acquisition of human behavioral datasets, specifically cue-reactivity and drug-reinstatement paradigms, to fit those parameters. This is the critical path from a calibrated theoretical model to a validated clinical tool.

The broader research trajectory connects this work to protein-level prediction. My TOPOLOGIX project demonstrates that sequence-based representations outperform structure-based methods for drug resistance prediction, achieving AUROC 0.804 on the Platinum benchmark compared to roughly 0.70 for mCSM-lig, while covering 100 percent of mutations versus 18 percent. This matters for the CCT program because resistance mutations in opioid and dopamine receptors directly affect the binding parameters that feed into the circuit model. A unified framework that predicts resistance from sequence and predicts consolidation prevention from circuit dynamics would be a genuinely new contribution to computational pharmacology.

My methodological track record includes a pre-registered replication that settled an open question in topological data analysis. The cardiotoxicity study tested whether bipartite persistent homology predicts hERG cardiotoxicity from protein-ligand interface geometry. The result was negative: topological features achieved AUROC 0.8426 versus 0.8782 for a plain descriptor baseline. This negative result was published directly rather than reframed, which is the standard I hold for all my work. The ergofluids project followed the same discipline: synthetic-data gates passed, but the first real-data gate failed its primary pre-registered criterion, and that failure was reported as such. AI2050's emphasis on rigorous, beneficial AI aligns with this commitment to falsifiability and honest reporting.

SHORT ESSAY: INTERDISCIPLINARY APPROACH

The CCT model is interdisciplinary by construction, not by decoration. It requires simultaneous competence in three domains: neuropharmacology to specify the receptor systems and their binding kinetics, dynamical systems theory to formulate the coupled ODE structure, and Bayesian statistics to calibrate the model against sparse biological data. My training spans all three. I hold a B.Pharm from the University of Ibadan with a German equivalent grade of 1.9, and I am a licensed pharmacist in Nigeria. I have built computational models in Python using scipy, NumPy, and RK45 integrators, and I have calibrated them with PyMC's MCMC samplers. I have also worked in molecular simulation with GROMACS and AutoDock, and in protein structure prediction with AlphaFold.

This combination is unusual. Most computational neuroscientists come from physics or computer science and lack the pharmacological grounding to specify receptor-level parameters. Most pharmacologists lack the mathematical training to formulate and solve coupled dynamical systems. The CCT model exists because I can do both. The Bayesian calibration framework is particularly important because it makes the model's uncertainty explicit. Clinicians need to know not just the point estimate of an intervention's effect, but the confidence interval around it. The posterior distributions from the MCMC calibration provide exactly that.

The interdisciplinary approach also extends to my software engineering practice. I have built four independent DuckDB-based data pipelines for ingest-to-analyte corpus processing across life sciences, technology, and social science domains. I self-host local LLM serving with llama.cpp and manage production systems with Linux VPS, systemd, Caddy TLS, and automated backup and disaster recovery. This is not peripheral infrastructure. It means I can build the full stack of a clinical decision support tool, from the dynamical model to the database to the user interface, without depending on external engineering support. The psyche-twin project, a typed multi-scale knowledge graph architecture for self-modeling, demonstrates this capability. Multiple evidence streams fuse into an append-only event log, and disagreement between streams becomes an explicit graph edge rather than being averaged away. The same architecture could support patient-specific CCT model personalization, where clinical observations, self-reported craving data, and pharmacokinetic measurements are fused into a single patient model.

SHORT ESSAY: SOCIETAL IMPACT BY 2050

The World Health Organization estimates that substance use disorders affect more than 35 million people globally, with treatment gaps exceeding 80 percent in low- and middle-income countries. Nigeria, my home country, has a particularly acute problem. The National Drug Law Enforcement Agency reported a significant rise in opioid and methamphetamine use over the past decade, while psychiatric infrastructure remains severely under-resourced. By 2050, the burden will have grown unless prevention and treatment strategies become fundamentally more effective.

The CCT model offers a path to that improvement. If validated, it would enable clinicians to predict, for a given patient and a given pharmacological intervention, whether reward-memory consolidation can be prevented. This is not a distant vision. The model already exists, is calibrated, and has confirmed all five pre-registered hypotheses. The missing piece is empirical validation of the circuit-layer parameters against real behavioral data. AI2050 support would close that gap.

The Africa angle is not a token. Computational psychiatry is almost entirely absent from African research institutions. There is no computational pharmacology group in Nigeria, and no African researcher has published a dynamical systems model of addiction consolidation. By building this capability as an independent researcher based in Nigeria, I am creating a template for how African scientists can contribute to frontier AI research without leaving the continent. The AI2050 program's global scope makes this a natural fit.

The 2050 horizon also matters for the technology trajectory. Protein language models like ESM-2 are improving rapidly. My TOPOLOGIX work shows that sequence-based representations already outperform structure-based methods for resistance prediction. By 2050, these models will be substantially more powerful, and the integration of protein-level predictions with circuit-level dynamical models will be a standard methodology. The CCT framework is designed to accommodate that integration. The receptor binding parameters that feed into the circuit model can be predicted from sequence, which means the entire framework can operate from genomic data alone. This is the long-term vision: a patient's genome, a predicted receptor binding profile, a calibrated circuit model, and a personalized intervention protocol.

SHORT ESSAY: EARLY-CAREER TRAJECTORY AND CAPACITY

I am 29 years old, a licensed pharmacist, and an independent researcher with a publication record that includes three sole-authored preprints under review and a co-authored paper under review at Alcohol, Elsevier. I am enrolled in the M.Sc. Digital Health program at the Hasso Plattner Institute and the University of Potsdam, starting Winter Semester 2026/27. My research has been endorsed by Kent Berridge at the University of Michigan, Samuel Gershman at Harvard, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU. Gershman provided my arXiv endorsement.

The AI2050 Early Career Fellow track is the correct fit for this stage. I am pre-PhD, which means I have not yet formalized my research trajectory within a doctoral program. The fellowship would provide the resources and credibility to pursue a PhD with a clear research agenda already established, or to continue as an independent researcher with a validated model. The CCT model is complete enough to be published and reviewed, but early enough in its validation that the fellowship would materially change its trajectory.

My capacity to execute is demonstrated by the breadth and completion of my independent projects. The CCT model required a 1,847-record literature screen, Bayesian calibration of a fourteen-parameter model, and pre-registration of five hypotheses. The TOPOLOGIX project required building an ESM-2 embedding pipeline, Morgan fingerprint generation, and a Random Forest classifier, evaluated on two benchmarks. The neurocascade engine required implementing coupled pharmacokinetic, receptor-binding, Wilson-Cowan, and behavioral-readout layers with 62 passing tests. The ergofluids project required extending Koopman operator methods with a Mori-Zwanzig memory kernel, with pre-registered gates and honest reporting of a failed real-data gate. These are not proposals. They are completed or in-progress research artifacts.

The AI2050 program's emphasis on research excellence, novelty, feasibility, and societal impact maps directly onto this record. The novelty is the tripartite CCT framework itself, which has no precedent in the addiction neuroscience literature. The feasibility is demonstrated by the completed calibration and confirmed hypotheses. The societal impact is the potential to transform addiction pharmacotherapy, particularly in low- and middle-income countries. The interdisciplinary approach, spanning neuropharmacology, dynamical systems, and Bayesian statistics, is exactly the kind of cross-domain thinking the program seeks to support.

CHECKLIST

- [ ] Verify AI2050 Fellows application portal and current deadline on the Schmidt Sciences website
- [ ] Confirm Early Career Fellow eligibility for pre-PhD applicants; if not eligible, identify Senior Fellow or alternative track
- [ ] Prepare CV in the required format, including ORCID 0009-0001-9272-6735 and GitHub github.com/AmunRaPtah
- [ ] Obtain or confirm letters of recommendation from Kent Berridge, Samuel Gershman, Nathaniel Daw, or Marcelo Mattar
- [ ] Confirm the three preprint submissions (IART, PNPBP, NBR) are still under review and obtain submission IDs
- [ ] Confirm the co-authored Alcohol, Elsevier paper submission status
- [ ] Prepare the OSF and Zenodo links for the CCT preprints and pre-registration documents
- [ ] Verify the exact AI2050 application questions and word limits; adjust essays to match if different from this draft
- [ ] Prepare a one-page project budget for the CCT validation phase, including behavioral dataset acquisition and compute costs
- [ ] Confirm the M.Sc. Digital Health enrollment status at HPI/Potsdam and obtain proof of enrollment
- [ ] Prepare a diversity statement or personal background section if the application requires one, emphasizing the Nigeria angle
- [ ] Verify the Platinum benchmark and SKEMPI 2.0 AUROC numbers against the TOPOLOGIX repository before submission

EDITOR NOTES

- Eligibility risk: AI2050 Early Career Fellows are typically postdoctoral or pre-tenure researchers. Eniola is pre-PhD and enrolled in a master's program. This may be a mismatch. Verify the program's exact eligibility language before submitting. If ineligible, consider applying to the Senior Fellow track with a strong case, or wait until the PhD is underway.

- The AI2050 program is funded by Schmidt Sciences, not Schmidt Futures. The Forbes article URL provided is a news report, not the official program page. The official application portal and deadline must be located before submission.

- The CCT model's five confirmed hypotheses are a strong claim. Verify that the pre-registration documents on OSF and Zenodo are publicly accessible and that the hypothesis confirmation is stated in the preprints. Any reviewer will check this.

- The endorsement from Nathaniel Daw and Marcelo Mattar is listed in the profile but no specific interaction or letter is confirmed. Eniola must confirm whether these endorsements are letters, emails, or arXiv endorsements, and whether they are willing to write recommendation letters for this specific application.

- The ergofluids failed real-data gate is reported honestly, which is good, but it may raise questions about feasibility in the AI2050 review. The application should frame this as methodological discipline, not failure. The current draft does this, but the framing should be consistent across all materials.

- The budget for the CCT validation phase is not specified in the profile. Eniola must prepare a realistic budget for behavioral dataset acquisition, compute, and potentially collaboration costs with a clinical site in Nigeria or Germany.

- The M.Sc. Digital Health program at HPI/Potsdam starts Winter Semester 2026/27. If the AI2050 fellowship begins before that, there may be a scheduling conflict. Confirm the fellowship start date and the master's program schedule.

- The profile lists "OIQB" in the task instructions but no project by that name appears in the research lines. This may be a typo or an internal codename. Clarify before submission.

MOTIVATION LETTER

The AI2050 Fellows program asks for research that could make AI hugely beneficial to society by 2050. Addiction is projected to remain one of the largest global disease burdens, with relapse rates above 60 percent within one year across substance classes. My Conjunctive Consolidation Threshold model, or CCT, addresses this problem directly by formalizing how reward-memory encoding can be pharmacologically prevented at the circuit level. The model couples three axes, dopaminergic reward prediction error, NMDAR-dependent long-term potentiation, and affective contrast, into a single ODE system solved with RK45 and calibrated through Bayesian MCMC using PyMC's DEMetropolisZ sampler. All fourteen free parameters were constrained by priors elicited from a systematic screen of 1,847 records in the literature. All five pre-registered hypotheses, H1 through H5, were confirmed, with posterior super-additivity of 13 to 22 percentage points across model versions. Three sole-authored preprints are currently under review at peer-reviewed journals, IART, PNPBP, and NBR, with a co-authored paper under review at Alcohol, Elsevier.

The AI2050 mission explicitly names major scientific questions and risks as hard problems. The CCT model sits at the intersection of two of them: the scientific question of how memory systems encode reward salience, and the technical challenge of building predictive dynamical models from sparse biological data. My approach treats the problem as a Bayesian inference task rather than a curve-fitting exercise. The model generates posterior distributions over pharmacological intervention parameters, which means it can predict not just whether an intervention works, but how much of each receptor system must be modulated to cross the consolidation threshold. This is a fundamentally different use of AI than pattern recognition on large datasets. It is mechanistic, interpretable, and falsifiable.

The Africa angle matters for AI2050's global mandate. Nigeria has one of the highest untreated substance use disorder burdens in West Africa, yet no computational pharmacology research group exists in the country. I am a licensed pharmacist trained at the University of Ibadan, currently enrolled in the M.Sc. Digital Health program at the Hasso Plattner Institute and the University of Potsdam. My independent research track includes the TOPOLOGIX project, which uses ESM-2 protein language model delta-embeddings combined with Morgan fingerprints and a Random Forest classifier to predict drug resistance mutations from sequence alone, achieving AUROC 0.804 on the Platinum benchmark and 0.634 on SKEMPI 2.0, while covering 100 percent of mutations compared to roughly 18 percent for structure-limited tools. This breadth, from molecular embeddings to circuit-level dynamical systems, is rare in early-career applicants.

AI2050's support would allow me to scale the CCT model from a calibrated theoretical framework to a validated clinical decision tool. The next phase requires fitting the circuit-layer parameters to real behavioral data, which is explicitly labeled as pending in my current neurocascade simulation engine. That engine couples pharmacokinetics to receptor binding to Wilson-Cowan circuit dynamics to behavioral readouts across three receptor systems, mu-opioid, D2 dopamine, and GABA-A, with 62 of 62 tests passing. The fellowship would fund the data acquisition and validation work that transforms this from a simulation platform into a predictive instrument for personalized addiction pharmacotherapy. By 2050, I intend for CCT-derived models to inform dosing protocols in clinical settings across low- and middle-income countries, where the need is greatest and computational psychiatry resources are scarcest.

RESEARCH STATEMENT

The hard problem I address is the prevention of pathological reward-memory consolidation in addiction. Current pharmacotherapy for substance use disorders targets receptor occupancy, not memory encoding. This is a fundamental mismatch. Addiction is maintained by drug-context associations that persist for decades, and no existing treatment directly interrupts the consolidation of those associations at the moment of exposure. The CCT model formalizes the conditions under which pharmacological intervention can prevent reward-memory encoding, providing a mechanistic, quantitative framework for a clinical problem that has resisted purely empirical approaches.

The model architecture is a tripartite coupled system. The first axis tracks dopaminergic reward prediction error, the second tracks NMDAR-dependent long-term potentiation in the mesolimbic pathway, and the third tracks affective contrast, the hedonic differential between drug and non-drug states. These three axes are coupled through a system of ordinary differential equations solved with RK45. The model was calibrated using Bayesian MCMC with the DEMetropolisZ sampler in PyMC, with fourteen free parameters constrained by literature-elicited priors. The prior elicitation process itself was systematic: a screen of 1,847 records from the addiction neuroscience literature, from which I extracted quantitative ranges for receptor binding affinities, synaptic plasticity time constants, and dopaminergic firing rates. This is a literature-grounded, quantitatively constrained dynamical system, not a toy model.

All five pre-registered hypotheses were confirmed. The key finding is super-additivity: combined modulation of dopaminergic and glutamatergic systems produces 13 to 22 percentage points greater consolidation prevention than the sum of individual effects. This has direct clinical implications. It suggests that combination pharmacotherapy targeting both the RPE signal and the LTP gate is not merely additive but synergistic, and that the synergy is predictable from the model's posterior distribution. The model also generates testable predictions about the timing of intervention. The consolidation threshold is crossed within a specific temporal window after exposure, and the model predicts the boundaries of that window with quantified uncertainty.

The next phase, which AI2050 support would enable, is empirical validation. The neurocascade engine, my receptor-to-behavior simulation platform, currently passes 62 of 62 unit tests across three receptor systems. The circuit-layer parameters are explicitly labeled illustrative pending fits to real behavioral data. The fellowship would fund the acquisition of human behavioral datasets, specifically cue-reactivity and drug-reinstatement paradigms, to fit those parameters. This is the critical path from a calibrated theoretical model to a validated clinical tool.

The broader research trajectory connects this work to protein-level prediction. My TOPOLOGIX project demonstrates that sequence-based representations outperform structure-based methods for drug resistance prediction, achieving AUROC 0.804 on the Platinum benchmark compared to roughly 0.70 for mCSM-lig, while covering 100 percent of mutations versus 18 percent. This matters for the CCT program because resistance mutations in opioid and dopamine receptors directly affect the binding parameters that feed into the circuit model. A unified framework that predicts resistance from sequence and predicts consolidation prevention from circuit dynamics would be a genuinely new contribution to computational pharmacology.

My methodological track record includes a pre-registered replication that settled an open question in topological data analysis. The cardiotoxicity study tested whether bipartite persistent homology predicts hERG cardiotoxicity from protein-ligand interface geometry. The result was negative: topological features achieved AUROC 0.8426 versus 0.8782 for a plain descriptor baseline. This negative result was published directly rather than reframed, which is the standard I hold for all my work. The ergofluids project followed the same discipline: synthetic-data gates passed, but the first real-data gate failed its primary pre-registered criterion, and that failure was reported as such. AI2050's emphasis on rigorous, beneficial AI aligns with this commitment to falsifiability and honest reporting.

SHORT ESSAY: INTERDISCIPLINARY APPROACH

The CCT model is interdisciplinary by construction, not by decoration. It requires simultaneous competence in three domains: neuropharmacology to specify the receptor systems and their binding kinetics, dynamical systems theory to formulate the coupled ODE structure, and Bayesian statistics to calibrate the model against sparse biological data. My training spans all three. I hold a B.Pharm from the University of Ibadan with a German equivalent grade of 1.9, and I am a licensed pharmacist in Nigeria. I have built computational models in Python using scipy, NumPy, and RK45 integrators, and I have calibrated them with PyMC's MCMC samplers. I have also worked in molecular simulation with GROMACS and AutoDock, and in protein structure prediction with AlphaFold.

This combination is unusual. Most computational neuroscientists come from physics or computer science and lack the pharmacological grounding to specify receptor-level parameters. Most pharmacologists lack the mathematical training to formulate and solve coupled dynamical systems. The CCT model exists because I can do both. The Bayesian calibration framework is particularly important because it makes the model's uncertainty explicit. Clinicians need to know not just the point estimate of an intervention's effect, but the confidence interval around it. The posterior distributions from the MCMC calibration provide exactly that.

The interdisciplinary approach also extends to my software engineering practice. I have built four independent DuckDB-based data pipelines for ingest-to-analyte corpus processing across life sciences, technology, and social science domains. I self-host local LLM serving with llama.cpp and manage production systems with Linux VPS, systemd, Caddy TLS, and automated backup and disaster recovery. This is not peripheral infrastructure. It means I can build the full stack of a clinical decision support tool, from the dynamical model to the database to the user interface, without depending on external engineering support. The psyche-twin project, a typed multi-scale knowledge graph architecture for self-modeling, demonstrates this capability. Multiple evidence streams fuse into an append-only event log, and disagreement between streams becomes an explicit graph edge rather than being averaged away. The same architecture could support patient-specific CCT model personalization, where clinical observations, self-reported craving data, and pharmacokinetic measurements are fused into a single patient model.

SHORT ESSAY: SOCIETAL IMPACT BY 2050

The World Health Organization estimates that substance use disorders affect more than 35 million people globally, with treatment gaps exceeding 80 percent in low- and middle-income countries. Nigeria, my home country, has a particularly acute problem. The National Drug Law Enforcement Agency reported a significant rise in opioid and methamphetamine use over the past decade, while psychiatric infrastructure remains severely under-resourced. By 2050, the burden will have grown unless prevention and treatment strategies become fundamentally more effective.

The CCT model offers a path to that improvement. If validated, it would enable clinicians to predict, for a given patient and a given pharmacological intervention, whether reward-memory consolidation can be prevented. This is not a distant vision. The model already exists, is calibrated, and has confirmed all five pre-registered hypotheses. The missing piece is empirical validation of the circuit-layer parameters against real behavioral data. AI2050 support would close that gap.

The Africa angle is not a token. Computational psychiatry is almost entirely absent from African research institutions. There is no computational pharmacology group in Nigeria, and no African researcher has published a dynamical systems model of addiction consolidation. By building this capability as an independent researcher based in Nigeria, I am creating a template for how African scientists can contribute to frontier AI research without leaving the continent. The AI2050 program's global scope makes this a natural fit.

The 2050 horizon also matters for the technology trajectory. Protein language models like ESM-2 are improving rapidly. My TOPOLOGIX work shows that sequence-based representations already outperform structure-based methods for resistance prediction. By 2050, these models will be substantially more powerful, and the integration of protein-level predictions with circuit-level dynamical models will be a standard methodology. The CCT framework is designed to accommodate that integration. The receptor binding parameters that feed into the circuit model can be predicted from sequence, which means the entire framework can operate from genomic data alone. This is the long-term vision: a patient's genome, a predicted receptor binding profile, a calibrated circuit model, and a personalized intervention protocol.

SHORT ESSAY: EARLY-CAREER TRAJECTORY AND CAPACITY

I am 29 years old, a licensed pharmacist, and an independent researcher with a publication record that includes three sole-authored preprints under review and a co-authored paper under review at Alcohol, Elsevier. I am enrolled in the M.Sc. Digital Health program at the Hasso Plattner Institute and the University of Potsdam, starting Winter Semester 2026/27. My research has been endorsed by Kent Berridge at the University of Michigan, Samuel Gershman at Harvard, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU. Gershman provided my arXiv endorsement.

The AI2050 Early Career Fellow track is the correct fit for this stage. I am pre-PhD, which means I have not yet formalized my research trajectory within a doctoral program. The fellowship would provide the resources and credibility to pursue a PhD with a clear research agenda already established, or to continue as an independent researcher with a validated model. The CCT model is complete enough to be published and reviewed, but early enough in its validation that the fellowship would materially change its trajectory.

My capacity to execute is demonstrated by the breadth and completion of my independent projects. The CCT model required a 1,847-record literature screen, Bayesian calibration of a fourteen-parameter model, and pre-registration of five hypotheses. The TOPOLOGIX project required building an ESM-2 embedding pipeline, Morgan fingerprint generation, and a Random Forest classifier, evaluated on two benchmarks. The neurocascade engine required implementing coupled pharmacokinetic, receptor-binding, Wilson-Cowan, and behavioral-readout layers with 62 passing tests. The ergofluids project required extending Koopman operator methods with a Mori-Zwanzig memory kernel, with pre-registered gates and honest reporting of a failed real-data gate. These are not proposals. They are completed or in-progress research artifacts.

The AI2050 program's emphasis on research excellence, novelty, feasibility, and societal impact maps directly onto this record. The novelty is the tripartite CCT framework itself, which has no precedent in the addiction neuroscience literature. The feasibility is demonstrated by the completed calibration and confirmed hypotheses. The societal impact is the potential to transform addiction pharmacotherapy, particularly in low- and middle-income countries. The interdisciplinary approach, spanning neuropharmacology, dynamical systems, and Bayesian statistics, is exactly the kind of cross-domain thinking the program seeks to support.

CHECKLIST

- [ ] Verify AI2050 Fellows application portal and current deadline on the Schmidt Sciences website
- [ ] Confirm Early Career Fellow eligibility for pre-PhD applicants; if not eligible, identify Senior Fellow or alternative track
- [ ] Prepare CV in the required format, including ORCID 0009-0001-9272-6735 and GitHub github.com/AmunRaPtah
- [ ] Obtain or confirm letters of recommendation from Kent Berridge, Samuel Gershman, Nathaniel Daw, or Marcelo Mattar
- [ ] Confirm the three preprint submissions (IART, PNPBP, NBR) are still under review and obtain submission IDs
- [ ] Confirm the co-authored Alcohol, Elsevier paper submission status
- [ ] Prepare the OSF and Zenodo links for the CCT preprints and pre-registration documents
- [ ] Verify the exact AI2050 application questions and word limits; adjust essays to match if different from this draft
- [ ] Prepare a one-page project budget for the CCT validation phase, including behavioral dataset acquisition and compute costs
- [ ] Confirm the M.Sc. Digital Health enrollment status at HPI/Potsdam and obtain proof of enrollment
- [ ] Prepare a diversity statement or personal background section if the application requires one, emphasizing the Nigeria angle
- [ ] Verify the Platinum benchmark and SKEMPI 2.0 AUROC numbers against the TOPOLOGIX repository before submission

EDITOR NOTES

- Eligibility risk: AI2050 Early Career Fellows are typically postdoctoral or pre-tenure researchers. Eniola is pre-PhD and enrolled in a master's program. This may be a mismatch. Verify the program's exact eligibility language before submitting. If ineligible, consider applying to the Senior Fellow track with a strong case, or wait until the PhD is underway.

- The AI2050 program is funded by Schmidt Sciences, not Schmidt Futures. The Forbes article URL provided is a news report, not the official program page. The official application portal and deadline must be located before submission.

- The CCT model's five confirmed hypotheses are a strong claim. Verify that the pre-registration documents on OSF and Zenodo are publicly accessible and that the hypothesis confirmation is stated in the preprints. Any reviewer will check this.

- The endorsement from Nathaniel Daw and Marcelo Mattar is listed in the profile but no specific interaction or letter is confirmed. Eniola must confirm whether these endorsements are letters, emails, or arXiv endorsements, and whether they are willing to write recommendation letters for this specific application.

- The ergofluids failed real-data gate is reported honestly, which is good, but it may raise questions about feasibility in the AI2050 review. The application should frame this as methodological discipline, not failure. The current draft does this, but the framing should be consistent across all materials.

- The budget for the CCT validation phase is not specified in the profile. Eniola must prepare a realistic budget for behavioral dataset acquisition, compute, and potentially collaboration costs with a clinical site in Nigeria or Germany.

- The M.Sc. Digital Health program at HPI/Potsdam starts Winter Semester 2026/27. If the AI2050 fellowship begins before that, there may be a scheduling conflict. Confirm the fellowship start date and the master's program schedule.

- The profile lists "OIQB" in the task instructions but no project by that name appears in the research lines. This may be a typo or an internal codename. Clarify before submission.

AI Draft — AI2050 Fellows