Plain-language explainers on the methods that determine whether a candidate is worth advancing. Each note is written around one decision question: what must be true, with evidence, before a candidate can move deeper in the pipeline.
Dish-scale assays matter when the outputs can be used to change a specific translational decision.
May 20, 2026
Why “in a dish” does not mean “simple”
“Clinical trials in a dish” is easy to over-interpret as “small.” In practice, the model stack can include three-dimensional architecture, controlled perfusion, co-culture, and repeated readouts at multiple time points.
The key test is not visual realism, but decision utility. A translational model is useful only when a result changes what a team does next—advance a candidate, revise a strategy, or stop spending on a weak path.
For a candidate-facing assay, we expect at minimum:
A defined endpoint tied to a human-relevant mechanism, not just a “positive-looking” assay output.
A protocol that names inclusion criteria, handling thresholds, and what variability is acceptable.
A decision rule that maps the result to a specific next action in the candidate workflow.
A documented limitation that explains where the model can still fail to predict clinical performance.
That structure is what makes a dish experiment translationally meaningful rather than visually impressive.
May 12, 2026
AI candidates need biological filters
AI systems can propose many biologically plausible compounds and mechanisms, but they do not solve the translational bottleneck on their own. The bottleneck is triage: turning hundreds of plausible ideas into a small set that merits expensive follow-up.
A biological filter is a deliberate set of preclinical experiments that each have a known purpose in that triage. It is not a “gate” for its own sake, and it should never be a black box. The point is to protect expensive animal and clinical work from weak hypotheses while preserving room for high-uncertainty innovation.
Filter input: AI-generated candidates or assay leads.
Filter outputs: a ranked shortlist and an explicit pass/fail rationale.
Filter record: all methods, endpoints, and assumptions preserved for audit and peer review.
Clinical trials in a dish can be one layer in this filter stack when they are benchmarked, documented, and interpreted against a predeclared decision rule.
May 6, 2026
Translation-ready assay design: what must be decided first
A useful assay is designed backwards from the translational question. If the decision question is “Should this molecule move into nonclinical development?” then the assay must answer exactly that question in measurable terms.
Teams often start with a biology-first model and only later ask how it informs decisions. That inversion creates data that is elegant but difficult to use. A translation-ready protocol starts with three pre-commitments:
Which decision it will influence (go, no-go, or redesign).
Which model feature is the minimal sufficient evidence.
What uncertainty the team can tolerate before escalating.
The result is easier to defend: if the assay says “advance,” it is because the decision gate was met under pre-specified criteria; if it says “do not advance,” the evidence trail shows what failed and why.
Candidate triage depends on pre-specified gates, not informal intuition.
April 27, 2026
What to include when publishing methods notes
For the translational public, the methods section is not optional boilerplate. It is the part of the paper that allows another team to understand the practical credibility of an assay.
Our notes format is built around three layers: model rationale, execution reproducibility, and interpretation constraints. If any layer is missing, a result can be misread as a general claim.
Model rationale: why this system over alternatives.
Execution protocol: where variability enters and how it is controlled.
Interpretation frame: what the result can and cannot predict.
This may feel strict, but strictness is a form of generosity in translational science: it prevents teams from over-trusting a signal before enough context has been established.
