Blog
Articles on statistics, data engineering, and AI applied to pharmaceutical, biotech, and cosmetic research. In French and English.
Single-cell RNA sequencing workflows break in predictable ways when you scale beyond one site, one batch, or one analyst.
Most early-stage biotech teams treat data infrastructure as a Phase 2 problem — until it costs three times as much to fix.
Bench and pilot KPIs look similar on paper but were measured under different definitions — a data infrastructure problem, not a scientific one.
Medical device teams discover FDA 21 CFR Part 11 requirements too late. Here is the six-question checklist before your first patient is enrolled.
In AI-driven drug discovery, the model is rarely the limiting factor. The experimental data pipeline is.
Most biotech R&D analyses live in one person's head. When that person leaves, the analysis leaves with them.
ICH E6(R2) requires that clinical trial data be traceable, reproducible, and attributable. A PowerPoint deck with manually pasted numbers satisfies none of those requirements. Most R&D teams do not realize their internal
ADC programs generate a data complexity that is qualitatively different from classical biologics: DAR distribution, plasma stability, conjugation site heterogeneity, multi-analyte PK, and tumor efficacy all need to be an
OoC platforms generate multi-parameter data (MEA recordings, live imaging, viability assays, cytokine panels) that most teams store in disconnected spreadsheets. This article explains why that creates a reproducibility a
ctDNA, liquid biopsy, ddPCR. Molecular diagnostic teams generate sequencing volumes their analytical infrastructure cannot keep up with. Most of the bottleneck is not the biology. It is the absence of a reproducible pipe
Most R&D data pipelines start as ad hoc scripts and accumulate technical debt until they become unmaintainable. Building reproducibility in from day one is not harder than building without it, it just requires different
Most preclinical data delays before an IND or CTA submission are not caused by the biology. They are caused by how the data was structured - or wasn't. Three recurring patterns account for the majority of these delays, a
AlphaFold solved a well-defined prediction problem. Most of drug discovery is not a well-defined prediction problem. The gap between "AI predicted a protein structure" and "AI discovered a drug" is where most AI drug dis
Cortical Labs' CL1 uses living cortical neurons on a silicon chip to perform computation. Brian2 simulates spiking neural networks in software. On MNIST, both achieve similar accuracy. The comparison reveals what biologi
For early-phase exploratory work, not pivotal, not submission-bound, a well-built Shiny application can do most of what a CRO statistician does. Not because Shiny is a statistics engine, but because early-phase analysis
Most pharma R&D teams do not need another data scientist who can train models. They need someone who understands GxP constraints, can translate between statisticians and engineers, and delivers work that is audit-ready o
A static report answers the questions you anticipated. An interactive tool answers the questions you could not anticipate. For R&D teams who need to explore data, not just read summaries, the difference is the difference
Most CDISC compliance issues are not caused by missing knowledge of the standard. They are caused by decisions made before anyone looks at the CDISC guide, in the data collection design, the database structure, and the m
SAS is the traditional tool for CDISC submissions. It is also expensive, requires specialist skills, and is increasingly unnecessary. R and Python can produce FDA-acceptable SDTM and ADaM datasets, if you know how to do
Every time an analyst copies a p-value from SAS output into a Word table, a small error becomes possible. Across a drug development program, these small errors compound into a significant liability. Automated reporting e
GxP compliance does not require enterprise infrastructure. With S3, dbt, renv, signed Git tags, and Quarto, a 5-person biotech team can build an audit-ready analytical stack in two weeks. This article walks through each
Non-compartmental analysis done manually in Excel is a liability. A reproducible R pipeline using PKNCA reduces errors, produces CDISC PC domain output automatically, and runs in minutes instead of days.
Pharmacokinetics is not black magic. One-compartment models, half-life, and AUC can be understood visually without a single equation. This article builds intuition for PK/PD concepts that every drug development professio
A statistical analysis plan that exists only as a Word document is half a plan. When the SAP and the analysis code are the same document, via Quarto, you eliminate the drift between what you said you would do and what yo
Quasi-complete separation occurs when a predictor perfectly or near-perfectly predicts an outcome category. The model will run, but the estimates are unreliable, often inflated to extreme values with huge standard errors
The proportional odds assumption underlies every cumulative logit model. Most analysts never test it. When it fails, which is common, the reported odds ratios are meaningless averages. This article explains the assumptio
Kaplan-Meier describes survival. Cox regression explains it. They answer different questions and should be used together, not chosen between. Understanding when each is appropriate, and when Cox assumptions fail, is fund
Repeated-measures ANOVA was designed for balanced, complete data from identical subjects. Cosmetic and clinical panel studies violate all three assumptions. Linear mixed models handle the real world correctly, and they a
EU Regulation 655/2013 sets six common criteria for cosmetic claims. Criterion 2 (truthfulness) and criterion 5 (fairness) impose specific statistical requirements that most cosmetics companies get wrong. This article ex
A cosmetic efficacy trial produces numbers. A marketing claim is a sentence. The gap between them is filled by decisions, about which endpoints to report, how to round, which subgroups to highlight. A transparent data pi
No articles match your search.