Writing (English) | Daiki Kumakurahttps://daikikumakura.github.io/feed-en.xml2026-10-28T13:00:00+09:00Daiki Kumakurahttps://daikikumakura.github.io/Gemcitabine infusion duration: plasma AUC and intracellular exposurehttps://daikikumakura.github.io/writing/gemcitabine-infusion-en.html2026-10-28T13:00:00+09:002026-10-28T13:00:00+09:00A fixed-effect plasma/PBMC subsystem from a published clinical model is implemented independently. Equal doses give equal plasma AUC, while PBMC AUC differences are determined by a categorical formation-clearance assumption. Constant-formation and tail scenarios clarify limits; no tumor, clinical benefit, safety or regimen recommendation is inferred.How does AlphaProtein Novo select enzymes for experimental testing?https://daikikumakura.github.io/writing/alphaprotein-novo-en.html2026-10-28T09:00:00+09:002026-10-28T09:00:00+09:00A source-based examination of catalytic motifs, structural metrics, sequence ensembles and denominators in enzyme design. Reported counts are recalculated; no model inference or enzyme experiments were performed.When shrinking totals hide growing lesions: an audit of public CT datahttps://daikikumakura.github.io/writing/organ-lesion-response-en.html2026-10-27T18:00:00+09:002026-10-28T00:00:00+09:00Two-timepoint lesion volumes from 276 patients reveal heterogeneous changes hidden by aggregation. This analysis examines measurement selection and within-patient organ contrasts, not causal drug effects or RECIST progression.How can AI predict cellular responses to drugs that have not been profiled?https://daikikumakura.github.io/writing/map-unprofiled-drugs-en.html2026-10-27T09:00:00+09:002026-10-28T00:00:00+09:00A descriptive audit of published MAP scores for 16 held-out drugs and four knowledge-pretraining stages, distinguishing unprofiled compounds from unseen structural or mechanistic groups. No model training or inference was performed.Unobserved mediators do not disappear from a networkhttps://daikikumakura.github.io/writing/hidden-mediator-networks-en.html2026-10-26T18:00:00+09:002026-10-28T00:00:00+09:00A Japanese essay on the difference between edges inferred from observed variables and the molecular interactions behind them. It links the mathematics of eliminating a mediator, a reproducible synthetic example, chemical-mediated microbial interactions, single-cell gene regulatory networks and mechanism assessment in drug discovery. It makes no claim about any real organism or drug.Before comparing TROP2 ADC exposure: why a reported AUC changedhttps://daikikumakura.github.io/writing/trop2-payload-exposure-en.html2026-10-26T13:00:00+09:002026-10-28T00:00:00+09:00An audit of sacituzumab govitecan and datopotamab deruxtecan distinguishes measured and derived analytes, label versions, exposure windows and response endpoints. The public summaries do not establish a shared payload target.Capivasertib rest-day placement at the same weekly dose: a public PopPK simulationhttps://daikikumakura.github.io/writing/capivasertib-rest-schedule-en.html2026-10-26T09:00:00+09:002026-10-28T00:00:00+09:00Three equal-dose weekly schedules are compared with an independent implementation of published PK equations. Weekly AUC is nearly unchanged, while minima and cutoff-dependent low-concentration times differ. Conditional sensitivity simulations do not establish clinical equivalence or predict efficacy, toxicity, or recovery.Repotrectinib escalation timing: exploring lead-in regimens with a public population PK modelhttps://daikikumakura.github.io/writing/repotrectinib-escalation-en.html2026-10-25T18:00:00+09:002026-10-28T00:00:00+09:00An independent Python implementation of published structural equations compares 7-, 14- and 21-day lead-ins. Conditional paired simulations and sensitivity analyses separate cumulative exposure, peak-window effects and late troughs. No patient refitting, NONMEM equivalence or efficacy/safety prediction is claimed.What changes when 100 assay labels are added to Boltz-2 candidate selection?https://daikikumakura.github.io/writing/boltz2-assay-finetuning-en.html2026-10-25T13:00:00+09:002026-10-28T00:00:00+09:00An independent aggregation of released predictions for eight MF-PCBA targets examines affinity-head adaptation, top-100 hit recovery, rescoring budgets and a nearest-active comparison. No model inference, retraining or prospective experiments were performed; scaffold annotations are author supplied.When a model fails to beat the baseline, what should we measure next?https://daikikumakura.github.io/writing/model-data-information-en.html2026-10-25T09:00:00+09:002026-10-28T00:00:00+09:00A Japanese essay on separating limitations of data, models and evaluation in perturbation prediction. Original analytical examples illustrate sampling-time identifiability and signal dilution; recent studies inform a practical measurement audit. This is not a reanalysis of empirical cell data.Does a model that predicts well understand the mechanism? Separating prediction from mechanistic inferencehttps://daikikumakura.github.io/writing/prediction-vs-mechanism-en.html2026-10-24T09:00:00+09:002026-10-28T00:00:00+09:00A Japanese essay distinguishing observational prediction, prediction under intervention and mechanistic claims, drawing on recent perturbation-prediction benchmarks and ICH M15. A reproducible synthetic indirect-response example shows that parameter sets compatible with the same data agree on a slowly varying regimen but diverge in the unobserved early hours, and that early sampling resolves the practical non-identifiability. It makes no claim about any real drug.Does separating PSA decline and regrowth improve early prediction?https://daikikumakura.github.io/writing/psa-early-observation.html2026-10-23T09:00:00+09:002026-10-28T00:00:00+09:00A public clinical PSA dataset used to compare landmark prediction, uncertainty calibration and decline/regrowth identifiability against simple benchmarks.Is another dose comparison informative? A benefit–risk analysis of ifinatamab deruxtecanhttps://daikikumakura.github.io/writing/b7h3-adc-dose-information.html2026-10-22T09:00:00+09:002026-10-28T00:00:00+09:00Public randomized B7-H3 ADC dose-cohort counts used to quantify conditional information value, decision uncertainty and sensitivity to unreported joint outcomes and follow-up.What can aggregate randomized data establish about treatment duration and benefit–risk?https://daikikumakura.github.io/writing/treatment-duration-benefit-risk.html2026-10-21T09:00:00+09:002026-10-28T00:00:00+09:00SCOT trial efficacy and neuropathy compared with endpoint-specific denominators, uncertainty, missingness bounds and limits of joint benefit–risk inference.Do fixed doses of anti-PD-1 antibodies keep exposure in range at extreme body weights?https://daikikumakura.github.io/writing/fixed-vs-weight-dosing-en.html2026-10-20T09:00:00+09:002026-10-28T00:00:00+09:00Re-implemented population PK models for pembrolizumab and nivolumab were applied to a public adult weight distribution. Fixed doses let trough exposure fall below the trial-tested range most often in heavy patients, weight-based doses in the first interval of light patients; the verdict on extended intervals depends on how the range is defined. No dose is recommended.RAS inhibitor dosing: from public PK/PD data to target-coverage scenarioshttps://daikikumakura.github.io/writing/ras-inhibitor-target-coverage.html2026-10-19T09:00:00+09:002026-10-28T00:00:00+09:00Public blood–tumor–DUSP6 reconstruction and hypothetical QD/BID comparisons with target turnover, variability and exposure costs.What changes in paired tumor immune-expression measurements during nivolumab?https://daikikumakura.github.io/writing/on-treatment-immune-change.html2026-10-18T09:00:00+09:002026-10-28T00:00:00+09:00Fixed immune scores in 43 public RNA pairs, with paired change, normalization sensitivity and patient-separated response-association diagnostics.How much of an existing drug's dose comparison can a new drug borrow?https://daikikumakura.github.io/writing/borrowing-prior-dose-selection-en.html2026-10-17T09:00:00+09:002026-10-28T00:00:00+09:00Existing-drug dose comparisons inform priors for hypothetical new drugs. Under the defined utility and deterioration tolerance, weight 0.5 was acceptable in the main condition and 0.25 across all sensitivity conditions; these are not general borrowing rules.When patients stop being measured: how much does dropout change a tumor-shrinkage estimate?https://daikikumakura.github.io/writing/tumor-dropout-sensitivity.html2026-10-16T09:00:00+09:002026-10-28T00:00:00+09:00Most tumor-scan dropout was death. A mixed model removed most of the survivor bias in 48-week shrinkage; a joint model and delta adjustment showed the 24-week arm difference was robust but the 48-week difference was fragile.Public T-cell engager evidence and the limits of regimen designhttps://daikikumakura.github.io/writing/tcell-engager-decision-framework.html2026-10-15T09:00:00+09:002026-10-28T00:00:00+09:00Public PopPK and efficacy reconstruction, Bayesian CRS sensitivity, and hypothetical measurement value and sampling design.How far does a dose-comparison trial support dose selection? Arm-level data for sotorasib 960 mg and 240 mghttps://daikikumakura.github.io/writing/sotorasib-dose-comparison-en.html2026-10-14T09:00:00+09:002026-10-28T00:00:00+09:00Published arm-level data from a randomized dose comparison are used to show the decision regions as the weights on benefit and harm change. The difference in severe toxicity is robust, but differences in response rate and OS are uncertain, and the conclusion depends on how harm is defined. No clinical dose is recommended.Do two immune-biology scores improve response prediction across urothelial cancer cohorts?https://daikikumakura.github.io/writing/immune-outcome-validation.html2026-10-13T09:00:00+09:002026-10-27T00:00:00+09:00Fixed Teff and stromal features evaluated across public urothelial cancer cohorts, with discrimination, calibration, missingness and transport limits.Can maximal degradation alone set the dosing interval of a targeted protein degrader?https://daikikumakura.github.io/writing/er-protac-pkpd-en.html2026-10-12T09:00:00+09:002026-10-27T00:00:00+09:00A review of public ER-PROTAC data and a minimal model of how assumptions about site-of-action exposure and target recovery change comparisons of dosing frequency and PD metrics. The model is not fitted to the actual drug, and no clinical dose is proposed.What did the Virtual Cell predict? Target genes and evaluation shortcutshttps://daikikumakura.github.io/writing/virtual-cell-target-evaluation-en.html2026-10-11T09:00:00+09:002026-10-27T00:00:00+09:00A Japanese review of perturbation forecasting, target-gene masking and recent benchmarks, with a reproducible synthetic example distinguishing evaluation shortcuts from downstream prediction. The example does not establish real-data performance or the superiority of pairwise masking.How complex does a population PK model need to be for exposure comparison? Measured nimotuzumab PKhttps://daikikumakura.github.io/writing/nimotuzumab-poppk-en.html2026-10-10T09:00:00+09:002026-10-27T00:00:00+09:00Three population PK structures estimated from individual PK data, with internal validation, convergence and exposure differences across dosing conditions used to examine the limits of model selection.Does a tumor-kinetic model add anything to a single early tumor change?https://daikikumakura.github.io/writing/tumor-kinetics-landmark.html2026-10-09T09:00:00+09:002026-10-27T00:00:00+09:00At 16, 24 and 32 weeks, a mixed-effects tumor-kinetic estimate added no held-out survival discrimination beyond the observed one-point change, and carrying the last value forward predicted the next measurement best.How far can public summary PK data evaluate a change in the tarlatamab dosing interval?https://daikikumakura.github.io/writing/tarlatamab-summary-pk-en.html2026-10-08T09:00:00+09:002026-10-27T00:00:00+09:00A two-compartment model estimated from published aggregate PK values, comparing Q2W, Q3W and Q4W exposure in light of identifiability and uncertainty.Where are the bottlenecks in AI-enabled drug discovery?https://daikikumakura.github.io/writing/ai-drug-discovery-bottlenecks-en.html2026-10-07T09:00:00+09:002026-10-28T00:00:00+09:00A Japanese essay reviewing experimental workflows, industry disclosures and clinical progress. It distinguishes experimental throughput from evidence useful for development decisions; revenue growth and individual clinical advances do not establish an AI-driven supply shortage or a causal improvement in clinical success.Can matching average exposure alone set the tarlatamab dosing interval?https://daikikumakura.github.io/writing/tcell-engager-regimen-design-en.html2026-10-05T09:00:00+09:002026-10-27T00:00:00+09:00Between-subject variability simulated from a published population PK model, with efficacy exposure–response, period-specific CRS and maintenance exposure examined separately. Matching average exposure does not preserve peaks and troughs, and the analysis shows the limits of deciding from public information.When does individual information improve PK/PD prediction?https://daikikumakura.github.io/writing/warfarin-pkpd.html2026-10-04T00:00:00+09:002026-10-27T00:00:00+09:00A reanalysis of warfarin PK/PCA data, revised with an absorption lag and a Hill exponent: effect delay, held-out prediction and the limits of extra PK.Do early tumor measurements improve prediction of subsequent survival?https://daikikumakura.github.io/writing/tumor-dynamics-survival.html2026-10-04T00:00:00+09:002026-10-27T00:00:00+09:00A patient-separated landmark analysis of public colorectal cancer data, including calibration, censoring, uncertainty and temporal-integrity sensitivity.Reimplementing a published mosunetuzumab population PK model: step-up dosing exposure and the limits of extrapolating to receptor occupancy and CRShttps://daikikumakura.github.io/writing/mosunetuzumab-public-pk-en.html2026-10-02T00:00:00+09:002026-10-28T00:00:00+09:00Mosunetuzumab PK recalculated from the published paper and its supplement. A paired analysis of 2,000 virtual patients examines the effects of dose and interval and how far exposure can be translated to CD20 occupancy and CRS. Figures, Python code and CSV files included.