Biotech platforms have to manage organizational complexity across teams, functions, and projects, alongside the scale and variety of their data.
2
Metadata is often harder to standardize than the underlying data because it records how experiments were run and gives the results their context.
3
Machine learning teams should start with inputs, outputs, and users before building the model in the middle of a system.
Summary
Jesse Johnson describes how his path from academic mathematics led through software engineering, healthcare data, drug discovery, and biotech research. He explains why building useful systems requires starting with the outside of a project, including its users, inputs, outputs, and operating constraints, before developing the central model. His diabetes prediction example shows how a highly accurate model can fail when doctors have no practical way to use its predictions or when the source-of-truth labels are unreliable. The conversation then turns to biotech data platforms. Jesse argues that metadata, such as sample preparation and experiment conditions, is often more difficult to manage than the instrument data itself. At Dewpoint Therapeutics, he uses cross-functional teams and shared mental models to align wet-lab scientists, data scientists, and data engineers. He also places MLOps early in the build-versus-buy cycle, with many companies still maintaining internal tools while commercial products mature.
Jesse moved from topology into data science because real problems mattered more than abstract methods
Jesse Johnson began as an academic mathematician studying topology. His work in topological data analysis brought him toward data science and machine learning, but he found that many practical difficulties were on the data engineering side. Living in the Boston area also made life sciences a natural direction. He describes biotech as a useful balance between clearly scoped work and the chance to think about science and research. His later roles took him from Google to Verily Life Sciences, Sanofi, Cellarity, and Dewpoint Therapeutics.
Software projects should begin with the boring interfaces around the central idea
At Google, Jesse learned that knowing how to code was different from knowing how to build a software project. He initially expected to be more prepared than he was and took months to become productive. One lesson was to begin with the outside of a project: identify the inputs, outputs, users, and systems it must connect to. Developers often start with the most interesting algorithm or model, then discover that the surrounding interfaces do not fit. Jesse says working from the outside in reveals those problems before the central work consumes too much time.
A good model can fail when its data and workflow do not match how people work
Jesse uses diabetes prediction as an example. A team might train a model to predict which patients will receive a diagnosis in six months, then use that prediction to encourage earlier intervention. In practice, doctors usually think about patients during or just before a visit, so a daily dashboard may not fit their workflow. The data may arrive after the visit, and there may be several competing ways to define whether a patient has diabetes. A model can therefore have high accuracy and still be unhelpful because its labels, timing, and delivery mechanism are wrong.
Career changes often require patience, luck, and repeated conversations
Jesse says his move from Google to Verily happened partly because a new team was forming in Cambridge and he happened to have the right conversation. He later followed colleagues from Verily to Sanofi and from Sanofi to Cellarity before joining Dewpoint Therapeutics. He is careful not to present this as a universal career formula. His practical advice is to look for opportunities, discuss desired changes with managers, and accept that transitions take time. The same applies when someone wants to alter their responsibilities without changing jobs.
The build-versus-buy cycle leaves companies with new products beside older internal systems
Jesse describes a recurring pattern in data engineering and MLOps. When a useful product does not exist, many internal teams build similar systems independently. Later, a project is open sourced or becomes a commercial product. Teams then compare the new option with their internal system, which may be more advanced because it started earlier, even though the external product could free engineers from maintenance. Data catalogs and data lakes are further along this cycle, while MLOps was still early in it. Legacy systems and managed services can remain side by side for a long time.
Metadata carries the experimental context that raw instrument data cannot provide
Jesse explains that a biotech instrument can produce consistent data files, while the metadata around those files varies across scientists. A biologist may track sample contents, preparation steps, and timing in an Excel sheet, then store the analysis in an electronic lab notebook. When several experiments are combined, column names may differ in capitalization or use different synonyms. The raw files may be easy to collect, but interpreting them requires knowing how each sample was prepared. Jesse calls metadata an organizational and operational problem because teams must agree on shared conventions.
Cross-functional ownership helps teams agree on data conventions
At Dewpoint Therapeutics, Jesse focuses on organizational design as much as technical tooling. Scientists and data teams may all recognize inconsistent spreadsheets as a problem, yet no individual feels able to tell everyone else what to change. Jesse creates cross-functional groups that include the relevant stakeholders, asks a smaller group to propose a solution, and brings that proposal back for review. This gives the problem an owner and creates a shared process. He connects the approach to shared mental models, since people from wet-lab, data science, and data engineering backgrounds often define the problem differently.
Data science can bridge wet-lab work and data engineering
Jesse describes separate responsibilities at Dewpoint. The machine learning group develops neural networks and tools for image analysis. The data science group works closely with wet-lab teams, translates their needs, and acts as a communication layer. The data engineering team builds the underlying data capabilities. This arrangement addresses the gap between highly technical engineering problems and the experimental questions that biologists care about. Jesse says people need time together to align their assumptions about the situation, the available tools, and what counts as an acceptable solution.
"Machine learning as we all know tends to be more effective if you can have a lot of data, create a black box model, and you can get more accurate results. But that doesn't exactly align up with the science."Jesse Johnson48:49
Who should watch
You are building data or machine learning systems for biotech and need to understand why experiment context is difficult to standardize.
Your model works in development, but the people expected to use it have no matching workflow or reliable source-of-truth labels.
You are deciding whether to keep an internal platform or adopt a newer managed product while legacy systems remain in place.
