MOTIVATION LETTER The AI2050 Fellowship asks what artificial intelligence can do for humanity by 2050. My answer is specific: AI can stop addiction before it starts, and it can predict drug resistance from sequence alone, covering every mutation rather than the eighteen percent that structure-based tools can reach. These are two of the hardest problems in neuropharmacology and infectious disease, and both are tractable with the methods I have already built and validated. I am Eniola Ayodele Olutogun, an independent computational researcher based in Nigeria, trained as a pharmacist at the University of Ibadan, and enrolled in the M.Sc. Digital Health program at the Hasso Plattner Institute and University of Potsdam. My primary research line, the Conjunctive Consolidation Threshold model, is a tripartite pharmacological framework for reward-memory encoding prevention in addiction. It couples dopaminergic reward prediction error, NMDAR-dependent long-term potentiation, and affective contrast into a single dynamical system, calibrated with Bayesian MCMC against a literature screen of 1,847 records. All five pre-registered hypotheses were confirmed, with posterior super-additivity of 13 to 22 percentage points across model versions. Three sole-authored preprints are under review at peer-reviewed journals, and a co-authored paper is under review at Alcohol. The AI2050 mission emphasizes AI for health and societal resilience. Addiction is a global health burden with disproportionate impact in low- and middle-income countries, where treatment infrastructure is scarce and prevention is the only scalable intervention. My CCT model directly addresses prevention by identifying the pharmacological threshold at which reward memories fail to consolidate. The second line, TOPOLOGIX, 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 an AUROC of 0.804 on the Platinum benchmark and 0.634 on SKEMPI 2.0, outperforming structure-based baselines while covering one hundred percent of mutations. My path has been unconventional. I have worked as a clinical pharmacist, a national product manager, and a bioinformatics researcher in genomic surveillance. I have built four independent data pipelines, self-hosted local LLM serving, and production systems infrastructure. This independence is the reason my work is pre-registered, my negative results are reported directly, and my code is public on GitHub. The AI2050 Fellowship values early-career researchers with evidence of independence and promise. I have that evidence. What I need is the community, the computational resources, and the time to take CCT from a validated model to a deployable clinical decision support tool, and to extend TOPOLOGIX from benchmarks to real clinical isolates. The fellowship's global and LMIC inclusion focus is central to my application. I am Nigerian, I practice pharmacy in Nigerian contexts, and I know exactly where these tools are needed and what it takes to deploy them there. AI2050 is the programme that can turn this work into a 2050-scale contribution. RESEARCH STATEMENT The hard problem I propose to solve is the prevention of addiction through AI-driven pharmacological intervention, and the prediction of drug resistance from sequence data alone. Both problems share a common methodological core: they require integrating machine learning with dynamical systems and molecular representation learning, and both have been held back by fragmented disciplinary approaches. The Conjunctive Consolidation Threshold model is the centerpiece. Addiction is fundamentally a memory disorder. A reward memory is consolidated when three conditions coincide: a dopaminergic reward prediction error signal, NMDAR-dependent long-term potentiation in the relevant circuits, and a sufficiently strong affective contrast. My model formalizes these three axes as a coupled ODE system, solved with RK45, and calibrated with Bayesian MCMC using PyMC's DEMetropolisZ sampler. The model has 14 free parameters, with priors elicited from a systematic screen of 1,847 records from the addiction neuroscience literature. All five pre-registered hypotheses were confirmed. The key finding is super-additivity: the combined effect of targeting all three axes exceeds the sum of individual effects by 13 to 22 percentage points across model versions. This means combination therapy at sub-threshold doses may prevent reward-memory consolidation without the side-effect burden of full-dose monotherapy. The next phase is translational. I will extend CCT to a clinical decision support tool that takes patient-specific pharmacokinetic and pharmacogenetic data and outputs optimal combination dosing regimens for relapse prevention. This requires fitting the model to real behavioral data, which is the explicit next validation gate. The AI2050 Fellowship would fund the computational infrastructure, the data acquisition, and the collaboration time needed to cross that gate. The second arm is TOPOLOGIX. Drug resistance is the silent driver of treatment failure in tuberculosis, malaria, and hospital-acquired infections. Structure-based prediction tools like mCSM-lig cover only about eighteen percent of mutations because most mutations occur outside resolved crystal structures. TOPOLOGIX uses ESM-2 protein language model delta-embeddings to represent the mutation's effect on the protein sequence landscape, combined with Morgan/ECFP drug fingerprints, fed into a Random Forest classifier. On the Platinum benchmark of 553 mutations, TOPOLOGIX achieves an AUROC of 0.804 with a standard deviation of 0.025. On SKEMPI 2.0, it achieves 0.634. It covers one hundred percent of mutations. This is a category change in coverage, not a marginal improvement. The methodological thread connecting CCT and TOPOLOGIX is my commitment to pre-registration, Bayesian calibration, and honest reporting of negative results. My cardiotoxicity topology study tested whether bipartite persistent homology predicts hERG cardiotoxicity from protein-ligand interface geometry. The pre-registered, powered replication found that topological features do not beat a plain descriptor baseline: AUROC 0.8426 versus 0.8782. This settled a comparison the literature had never actually run. My interface-topology-for-resistance study found the same class of constructs carry almost no signal for drug-resistance prediction, AUROC 0.425 and 0.485 on the Platinum benchmark. These negative results are published and reported directly. They are why TOPOLOGIX pivoted to sequence representations, and they are why the field should trust my positive results. The AI2050 Fellowship's selection criteria emphasize ambition, technical novelty, feasibility, and collaborative potential. My ambition is a world where addiction is preventable and drug resistance is predictable. The technical novelty is the integration of dynamical-systems pharmacology with protein language models. The feasibility is demonstrated by my track record: five pre-registered studies, three sole-authored preprints under review, a co-authored paper under review, and a fully tested simulation engine, neurocascade, with 62 of 62 tests passing. The collaborative potential is evidenced by endorsements from Kent Berridge at Michigan, Samuel Gershman at Harvard, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU. The computational needs are real. Training and validating TOPOLOGIX on larger clinical datasets requires GPU access. Fitting CCT to behavioral data requires significant MCMC compute. The AI2050 Fellowship's provision for computational needs would be directly applied to these bottlenecks. ESSAY: GLOBAL HEALTH AND SOCIETAL RESILIENCE Addiction and antimicrobial resistance are two of the most destabilizing forces in global health. They do not respect borders, they disproportionately burden low- and middle-income countries, and they compound each other: injection drug use drives hepatitis C and HIV transmission, while antimicrobial resistance turns routine infections into fatal ones. My work addresses both, and my position as a Nigerian researcher is central to that address. Nigeria has one pharmacist per 4,000 people, a fraction of the World Health Organization's recommended ratio. There is no national addiction treatment infrastructure to speak of, and antimicrobial resistance surveillance is fragmented. This is the daily reality of clinical practice. I have worked as a clinical pharmacist in Nigerian pharmacies and as a national product manager for a healthcare company. I know what it means to prescribe without resistance data, and to treat addiction without any pharmacological prevention tool. The CCT model offers a prevention-first approach that is uniquely suited to resource-constrained settings. If combination therapy at sub-threshold doses can prevent reward-memory consolidation, it can be deployed as a preventive intervention before addiction develops, rather than as a treatment after dependence is established. This is the difference between a public health intervention and a clinical one. The AI2050 mission explicitly values AI for societal resilience. Preventing addiction is societal resilience. TOPOLOGIX offers the same logic for resistance. A sequence-based predictor that covers one hundred percent of mutations can be deployed in any setting with sequencing capacity, which is increasingly available even in low-resource contexts. It does not require crystallography infrastructure. It does not require curated structure databases. It requires a sequence and a model. That is a deployable tool. The AI2050 community includes researchers working on AI for health, AI for scientific discovery, and trustworthy AI. My work sits at the intersection of all three. The CCT model is a scientific discovery about the pharmacology of memory consolidation. TOPOLOGIX is an AI tool for health. My pre-registration practices and honest negative-result reporting are contributions to trustworthy AI. I am not asking the fellowship to fund a speculative venture. I am asking it to fund the next validation gate of work that has already passed its pre-registered hypotheses, and to bring that work into a community that can help it scale. ESSAY: INDEPENDENCE AND EARLY-CAREER TRAJECTORY My career does not follow the standard postdoc pipeline. I completed my pharmacy degree at the University of Ibadan with a CGPA of 5.1 out of 7.0, a German equivalent of 1.9, and I am licensed by the Pharmacists Council of Nigeria. I have worked as a clinical pharmacist, a research assistant in NMDA and insulin docking, and a bioinformatics researcher in antimicrobial resistance genomics. I am now enrolled in the M.Sc. Digital Health program at the Hasso Plattner Institute and University of Potsdam, starting in the Winter Semester of 2026-27. The standard pipeline would not have given me the independence to pursue a multi-domain research agenda. I built the CCT model as an independent researcher, with no institutional lab and no supervisor. I conducted the 1,847-record literature screen myself. I wrote the Bayesian calibration code myself. I pre-registered the hypotheses myself. I submitted the preprints myself. This independence is the strongest evidence of my promise as an early-career researcher. The AI2050 Fellowship's early-career track is designed for researchers within a few years of their PhD. I do not yet have a PhD, but I have a body of pre-registered, peer-reviewed-in-process research that exceeds what many doctoral theses contain. My arXiv endorsement from Samuel Gershman at Harvard, and my collaborations with Kent Berridge at Michigan, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU, demonstrate that established researchers in computational neuroscience and reinforcement learning take my work seriously. The fellowship would change my trajectory in three concrete ways. First, it would provide the computational resources to fit CCT to real behavioral data and to scale TOPOLOGIX to clinical isolate datasets. Second, it would place me in a community of AI2050 fellows working on adjacent hard problems, which is the collaboration network I currently lack as an independent researcher. Third, it would fund the time to complete the M.Sc. and transition to a PhD, with a research agenda already defined and validated. I am not applying to the AI2050 Fellowship because it is prestigious. I am applying because it is the only programme I have found that explicitly values the combination of AI for health, global and LMIC inclusion, and early-career independence. My work is ready for that combination. CHECKLIST - [ ] Verify AI2050 Fellowship application portal and current deadline on Schmidt Sciences website - [ ] Confirm eligibility for early-career track without completed PhD; contact programme staff if unclear - [ ] Obtain official transcript from University of Ibadan with CGPA and German equivalent conversion - [ ] Confirm enrollment status and matriculation date for M.Sc. Digital Health at HPI/Potsdam - [ ] Collect ORCID record and ensure all preprints are linked to ORCID iD 0009-0001-9272-6735 - [ ] Verify GitHub repository (github.com/AmunRaPtah) is public and contains CCT and TOPOLOGIX code - [ ] Update personal site (zyco.org) with current research status and links to all preprints - [ ] Confirm submission status of three sole-authored preprints (IART, PNPBP, NBR) and Alcohol co-authored paper - [ ] Draft letters of support from Kent Berridge, Samuel Gershman, Nathaniel Daw, and Marcelo Mattar - [ ] Prepare budget breakdown for computational resources (GPU access, MCMC compute, data acquisition) - [ ] Prepare two-page CV in Schmidt Sciences format if required - [ ] Confirm whether AI2050 requires a project timeline or Gantt chart; prepare if so - [ ] Verify all pre-registration documents are publicly accessible on OSF or Zenodo - [ ] Confirm neurocascade repository is public with 62/62 tests passing - [ ] Prepare statement on data availability and reproducibility practices EDITOR NOTES - Eligibility risk: The AI2050 early-career track typically expects postdocs or faculty within a few years of PhD. Eniola does not have a PhD and is enrolled in an M.Sc. This is a genuine eligibility risk. The application should lead with the research output and independence evidence, and the programme contact should be queried directly about PhD requirements before submission. - The CCT model is the correct research line to center for this programme, not TOPOLOGIX alone. The framing angle in the strategy notes is correct: CCT is the primary contribution, TOPOLOGIX is the second arm. Do not lead with the negative topology results as the main story; use them as evidence of rigor within the research statement, not as the headline. - The ergofluids project is explicitly a methods-validation exercise with a failed real-data gate. Do not present it as validated or as a venture. It is mentioned nowhere in this draft, which is correct. If the application form asks for all research activities, describe ergofluids as a pre-registered methods-validation study whose first real-data gate did not meet its primary criterion, reported directly. - The psyche-twin project is a personal knowledge-graph architecture with no external validation. It is not mentioned in this draft. If the application form requires a complete research portfolio, describe it as an exploratory personal infrastructure project, not as a research contribution. - Facts to verify: the exact total award amount ($18M total is stated in the profile but should be confirmed against the current programme cycle), the exact fellowship cohort size, whether the fellowship covers tuition or only research costs, and whether the M.Sc. enrollment at HPI/Potsdam creates any conflict with fellowship residency requirements. - The applicant must insert personal details not in the profile: specific dates of pre-registration, the exact journal names for the three preprints (IART, PNPBP, NBR are journal abbreviations that should be spelled out), the specific behavioral dataset planned for CCT fitting, and any teaching or mentoring experience that would strengthen the collaborative potential criterion. - The endorsement letters from Berridge, Gershman, Daw, and Mattar are listed as collaborators or endorsements in the profile, but it is not confirmed that they have agreed to write letters for this specific application. This must be confirmed before submission.