For toxicology and safety assessment teams
Morphology search for toxicology archives. Precedent answers the pathologist’s oldest question: has this been seen before? It indexes your slide archive as it stands, with no labelling, and brings search, triage and treatment-effect analysis into one web viewer.
Modules
Three modules on one archive index
Every prior case of a lesion, in seconds.
- Select any region on a slide and get ranked look-alike regions from the whole archive
- Filter by species, strain, organ, compound, dose, timepoint and control group
- Automatic historical-control matching, with full study provenance
- Compare matches side by side in WsInsight
- No labelling: the archive is indexed as it stands
Read the abnormal slides first.
- Slide-level triage, normal or needs review, before the study is opened
- A heatmap of abnormal regions on every flagged slide
- A prioritised reading queue across the whole study
- One model across organs, with no per-finding classifiers to build
- Runs alongside the read, and the pathologist confirms every call
See what the compound is doing to tissue.
- Measures every tissue phenotype across dose groups and timepoints
- Flags dose-related changes against concurrent and historical controls
- Links phenotypes to study readouts: clinical chemistry, haematology and organ weights
- A study-level summary for the pathologist and toxicologist
The problem
Slide archives cannot be searched by what the tissue looks like
Historical controls are matched on metadata
Species, strain, age and study design indicate that two control groups should be comparable. They do not confirm that the tissue is, and verifying it by hand across years of studies is not practical.
Cross-study recall depends on individuals
Whether a finding has been seen before in another compound’s studies depends on which pathologist is asked. The archive itself cannot answer the question.
Only the metadata is searchable
Study number, accession, dose group and diagnosis code do not find the case that resembles the one under review, so assembling a matched set remains a manual task.
How it works
From stored slides to a searchable index
- 01
Index what you already hold
Whole-slide images across your studies are tiled and encoded. No re-scanning, no annotation step, and no need to decide up front what you will later want to search for.
- 02
Build the phenotype vocabulary
Tiles are clustered into histomorphological phenotypes, giving every slide a token composition, and giving the tissue-language model a slide-level representation that keeps the spatial arrangement of those tokens intact.
- 03
Search, screen and measure
Query with a region, triage a study before it is read, or measure phenotypes across dose groups, all in the WsInsight viewer. Every result is a real case with real tiles that a pathologist can open and judge.
Why this is different
Matching at case level, not patch level
Tile-level similarity search returns another patch with the same texture, which is rarely the question being asked. Because our tissue-language model reads a slide as a composition of phenotype tokens with their spatial relationships preserved, a query can match on how morphology is arranged across the whole slide. The result is a similar case rather than a similar tile.
The deliverable
What you get
You get: three modules (Search, Screen and Signal) in the WsInsight web viewer, deployed on your own hardware or hosted in the TileBio cloud.
Questions
Frequently asked
- Is Precedent a diagnostic or GLP tool?
- No. Precedent is decision support and research software, not a medical device. It surfaces prior cases, flags slides for review and summarises treatment effects, and the pathologist reads and signs every finding.
- Do you need labelled lesion data to start?
- No. The model is self-supervised: it learns tissue phenotypes from unlabelled whole-slide images. Your existing scores are useful for validating and interpreting what it finds, but they are not required to train it.
- Which species and tissues are supported?
- The approach is not species-specific, because it is not trained against a fixed lesion vocabulary. Coverage in practice depends on what is in your slide set, which is one of the things a pilot establishes.
- Does it work across studies run years apart on different scanners?
- Yes. The model learns representations that are robust to stain and scanner variation, which is what allows a multi-year, multi-site archive to be searched as one collection rather than study by study.
- Do we need to re-scan or annotate the archive first?
- No. Indexing runs on the whole-slide images you already hold, with no annotation step. The model’s representations are robust to scanner and stain variation, which is what lets a multi-year, multi-scanner archive be searched as a single collection.
- What can we search with?
- A whole slide, a region you draw on one, or a phenotype from the learned vocabulary. The result is a ranked list of matching cases, each traceable back to the tiles that made it match.
- Where does our data have to sit?
- Precedent deploys on your own hardware in Docker, or is hosted in the TileBio cloud with an API. Slide archives usually carry governance constraints that determine the deployment model, and we confirm early whether we can work within yours.
- How do we start?
- Most teams start with a safety finding investigation: a fixed-fee study on a defined question using your own slides. Precedent is also available as an annual licence covering one or more modules.