MOTIVATION LETTER The Seed STEM Fellows Program at ByteDance funds researchers who use AI to accelerate scientific discovery. My research does exactly that, and I can point to specific results. I am an independent computational researcher with a B.Pharm from the University of Ibadan, currently enrolled in the M.Sc. Digital Health program at Hasso Plattner Institute / University of Potsdam. My primary research line, TOPOLOGIX, predicts drug-resistance mutations from protein sequence alone using ESM-2 protein language model delta-embeddings combined with Morgan/ECFP drug fingerprints and a Random Forest classifier. On the Platinum benchmark of 553 mutations, TOPOLOGIX achieves AUROC 0.804 plus or minus 0.025. On SKEMPI 2.0, it achieves 0.634. These numbers beat structure-based baselines such as mCSM-lig at approximately 0.70 while covering 100 percent of mutations, compared to roughly 18 percent for structure-limited tools. This is a direct demonstration of AI accelerating a bottleneck in drug development: knowing which mutations will confer resistance before they emerge in the clinic. The program asks for candidates who can assess AI's value and limitations in their field. I have done both. My cardiotoxicity topology study tested whether bipartite persistent homology predicts hERG cardiotoxicity from protein-ligand interface geometry. I pre-registered the analysis, ran a powered replication, and found that topological features do not beat a plain descriptor baseline: AUROC 0.8426 versus 0.8782. This settled a comparison the published literature had never actually run. I reported the negative result directly. That is the same rigor I apply to TOPOLOGIX, where I know exactly which parts of the pipeline carry signal and which do not. I use AI as a daily productivity tool across my research. My Bayesian calibration pipeline runs PyMC DEMetropolisZ with 14 free parameters for my CCT model of reward-memory encoding in addiction. I maintain four independent DuckDB-based ingest-to-analyze pipelines across life sciences, tech/AI/security, and social science domains. I self-host local LLM serving with llama.cpp and on-demand model swapping. This is hands-on engineering, not theoretical familiarity. I understand the on-site requirement in Beijing's Haidian District. My M.Sc. program at HPI/Potsdam is structured with flexible research components, and I am prepared to discuss how I would structure my collaboration to meet the residency requirement. I am applying as a Scientific Advisor, not a Ph.D. intern, because my independent research record, including three sole-authored preprints under review at peer-reviewed journals and a co-authored paper at Alcohol (Elsevier), demonstrates equivalent research expertise. I am available to discuss my fit with the program team at your convenience. RESEARCH STATEMENT TOPOLOGIX: Predicting Drug-Resistance Mutations from Sequence Alone with Protein Language Models The problem. Drug resistance is a primary cause of treatment failure across oncology, infectious disease, and psychiatry. Predicting which mutations will confer resistance, before they appear in a patient, would let clinicians choose therapies with longer useful lives and let drug developers design compounds that evade known escape routes. The current standard tools, such as mCSM-lig, require a protein structure. That requirement excludes roughly 82 percent of clinically relevant mutations because no high-resolution structure exists for the relevant protein-ligand complex. A sequence-based predictor would cover the full mutation space. The method. TOPOLOGIX uses ESM-2 protein language model delta-embeddings. For each mutation, I compute the change in the ESM-2 embedding between the wild-type and mutant sequence. I concatenate this delta-embedding with Morgan/ECFP drug fingerprints and train a Random Forest classifier. The pipeline is implemented in Python with scikit-learn, RDKit, and the HuggingFace transformers library. The full code is on GitHub. The results. On the Platinum benchmark of 553 mutations, TOPOLOGIX achieves AUROC 0.804 plus or minus 0.025. On SKEMPI 2.0, it achieves 0.634. Structure-based baselines such as mCSM-lig score approximately 0.70 on the subsets they can cover. TOPOLOGIX covers 100 percent of mutations while beating that baseline on the full set. This is the first result I know of that demonstrates sequence-only prediction outperforming structure-based prediction for this task at this coverage level. The rigor behind the result. I did not arrive at TOPOLOGIX by chasing a positive result. My prior study tested whether interface topology, specifically bipartite persistent homology with an opposition-distance metric, could predict drug resistance. On the Platinum benchmark, the topological features achieved AUROC 0.425 and 0.485, essentially no signal. I published that negative result and pivoted to sequence representations. That falsification is why TOPOLOGIX exists. The program's selection criteria ask for candidates who can assess AI's limitations. I have a documented record of doing exactly that, including my hERG cardiotoxicity replication study where topological features lost to a plain descriptor baseline (AUROC 0.8426 versus 0.8782) in a pre-registered, powered analysis. Why this fits Seed STEM. The program seeks researchers with deep STEM expertise, strong AI/ML skills, and a demonstrated interest in accelerating scientific discovery. TOPOLOGIX is a working system that accelerates a specific scientific bottleneck. It is a running pipeline with benchmark numbers, not a proposal. I also bring the broader computational toolkit the program values: Bayesian MCMC calibration with PyMC, ODE modeling with RK45, topological data analysis with Ripser and GUDHI, molecular dynamics with GROMACS, and HPC workflow management with Nextflow and SLURM. Next steps. I am extending TOPOLOGIX in three directions. First, I am testing whether ESM-3 or other newer protein language models improve the delta-embedding signal. Second, I am expanding the benchmark set beyond Platinum and SKEMPI 2.0 to include resistance data from clinical cohorts. Third, I am integrating the predictor into a full drug-development pipeline that includes ADMET screening and circuit-level pharmacology simulation from my neurocascade engine. The on-site collaboration at ByteDance's Beijing campus would let me work directly with the Seed STEM team on scaling the ESM-2 inference pipeline and exploring whether ByteDance's infrastructure can accelerate the embedding computation across the full human proteome. ESSAY: KEY CHALLENGES IN MY FIELD AND HOW AI ADDRESSES THEM Two challenges dominate computational drug resistance prediction. The first is structural coverage. Most resistance mutations occur in proteins without high-resolution structures, so structure-based predictors simply cannot score them. The second is signal extraction. Resistance is a systems-level phenotype driven by subtle changes in binding affinity, allostery, and conformational dynamics, and it is not obvious which representation of a mutation carries the predictive signal. AI addresses both. Protein language models such as ESM-2 learn representations from sequence alone, so they cover the full mutation space without requiring a structure. My TOPOLOGIX pipeline demonstrates this: 100 percent coverage on the Platinum benchmark versus roughly 18 percent for structure-limited tools. For signal extraction, the delta-embedding approach captures the directional change in the learned representation space, which encodes the functional consequence of the mutation in a way that hand-crafted features do not. The AUROC of 0.804 on Platinum shows the signal is there. But AI also has limits, and I have documented them. My persistent homology studies for both hERG cardiotoxicity and drug resistance showed that topological features of protein-ligand interfaces carry almost no predictive signal for these tasks. The AUROC values of 0.425 and 0.485 on Platinum were worse than random for one task. I reported these results honestly. This matters for the Seed STEM program because the program explicitly asks for candidates who can assess AI's value and limitations. I have a track record of doing both in the same research line. CHECKLIST - [ ] Confirm Seed STEM Fellows Program application deadline on https://seed.bytedance.com/en/seedstem - [ ] Verify whether the program requires a formal affiliation letter from HPI/Potsdam for the Scientific Advisor track - [ ] Prepare CV with TOPOLOGIX preprint and the Alcohol (Elsevier) co-authored paper as the two representative publications - [ ] Draft the brief statement of 1-2 key challenges in the field and how AI is used, using the ESSAY section above as the base - [ ] Confirm on-site collaboration schedule in Beijing (Haidian District) and coordinate with M.Sc. Digital Health program structure at HPI/Potsdam - [ ] Prepare PDF copies of TOPOLOGIX preprint and hERG topology replication study as supporting evidence - [ ] Verify ORCID (0009-0001-9272-6735) and GitHub (github.com/AmunRaPtah) links are active and public - [ ] Confirm whether letters of recommendation are required; if so, request from Kent Berridge, Samuel Gershman, or Nathaniel Daw - [ ] Prepare a one-page research summary of TOPOLOGIX with benchmark tables for quick review - [ ] Submit application and save confirmation page EDITOR NOTES - Eligibility risk: The program's Scientific Advisor track requires affiliation with an academic institution or senior industry expertise. Eniola is enrolled in an M.Sc. program, not a Ph.D., and is an independent researcher. The M.Sc. enrollment at HPI/Potsdam may satisfy the academic affiliation requirement, but this must be verified with the program directly before submitting. - On-site requirement: The program requires on-site collaboration in Beijing. Eniola's M.Sc. program is in Potsdam, Germany. The motivation letter states he is prepared to discuss structuring the collaboration, but the actual feasibility of extended residency in Beijing during the M.Sc. program must be confirmed with HPI/Potsdam before any commitment. - Facts to verify: The AUROC values for TOPOLOGIX (0.804 on Platinum, 0.634 on SKEMPI 2.0) and the mCSM-lig baseline (approximately 0.70) come from the applicant profile and should be cross-checked against the actual preprint before submission. The claim that TOPOLOGIX is the first sequence-only predictor to beat structure-based baselines at this coverage level should be softened or verified against recent literature, as this is a strong claim that reviewers may challenge. - Missing personal detail: The application materials do not include Eniola's specific motivation for wanting to work with ByteDance specifically, beyond the program's mission. He should insert one or two sentences about why ByteDance's infrastructure or research environment specifically appeals to him, drawn from his own knowledge of the company, before submitting. - The CCT model, neurocascade, ergofluids, and psyche-twin research lines were deliberately excluded from this application. They do not match the Seed STEM program's focus on AI-accelerated scientific discovery as directly as TOPOLOGIX does. Do not add them to the application materials; they would dilute the focus.