The question is not “what software,” but “which layer of the stack”
A zebrafish toxicological exposure system generates two very different kinds of raw material. On one side, there are plate-level operational records: which well received which compound, at what concentration, on which day, under which exposure scenario. On the other, there are biological readouts: mortality calls, malformation scores, photomotor responses, swimming traces, and increasingly, image-derived morphometric features. Treating “data analysis tools” as a single category flattens that distinction, and it leads labs to buy a tracking package expecting it to solve a data management problem it was never designed to touch.
The practical way to think about the toolchain is in layers. Layer one captures and scores endpoints at the bench. Layer two stores the observation records with enough structure to be queryable. Layer three runs the statistics that turn observations into potency estimates, classifications, and comparisons. A lab can get away with commercial tools at one layer and open-source tools at another, and many do. Problems start when a single application is expected to cover all three without exposing where it stops.
Commercial platforms solve capture well, and that is not a small thing
ViewPoint’s Zebralab and DanioVision systems show up repeatedly in published zebrafish methods for a reason. They sit at the capture and first-pass quantification layer, and they are built around the realities of behavioral endpoints. A typical use case: larvae are loaded into a 96-well plate, placed in a Zebrabox chamber, and exposed to alternating light and dark cycles while a camera records activity. The software then returns per-well metrics such as total movement, distance traveled, and response amplitude across epochs. Studies using the Tanguay lab 120 hpf assay have relied on ViewPoint Zebralab alongside a custom photomotor response analysis tool and a separate zebrafish acquisition and analysis program, which says something useful about how these systems actually get deployed: rarely alone, usually as one component in a longer pipeline.
The strength of this class of tool is consistency at acquisition. Frame rates, illumination timing, and well-level tracking are controlled, so the operator is not making judgment calls about when a stimulus fired or which pixel belonged to which larva. The limitation is equally clear. A tracking system tells you that movement changed. It does not tell you whether the change is a specific developmental effect or a secondary consequence of general toxicity, and it does not reconcile that output with a morphology-based developmental endpoint scored by a different person on a different afternoon.
Open-source tools fill the gaps, at a cost
The open-source side of the ecosystem has expanded considerably, and the coverage is now broad enough that most labs can assemble a functional pipeline without a large software license.
FishInspector, developed at the Helmholtz Centre for Environmental Research, was built to derive unbiased quantitative assessments of morphological phenotypes from embryo images, and it is freely available through a public code repository. ZembryoAnalyser combines morphometric, motion tracking, and heart rate analysis into a single open-source platform, which is attractive for labs running cardiovascular endpoints alongside behavioral ones. ZebraFET takes a different angle entirely. It is a free desktop application built specifically around OECD Test Guideline 236, handling project setup, daily well scoring, photo management, LC50 and NOEC/LOEC calculations, validity checks, and report generation.
That last example deserves attention because it reveals a common misconception. A lab can be fully compliant with TG 236 and still have no unified data layer. ZebraFET enforces the regulatory arithmetic and the validity gates, but the observation terms a technician types into it are still the technician’s terms. The moment two labs compare results, the terminology problem returns.
Why ontologies and data pipelines are the part most labs skip
The SEAZIT program, run through NIEHS, was created precisely because cross-laboratory developmental toxicity outcomes vary when procedures and terminology are not standardized. A collaborative exercise within SEAZIT found that applying ontology terms increased consistency and decreased ambiguity, and that a common data standard should reduce heterogeneity in reported terms and improve agreement between laboratories. In the definitive interlaboratory study, three laboratories tested a blinded set of 42 chemicals across four exposure conditions, and laboratory-specific phenotype recording terms were mapped to Zebrafish Phenotype Ontology terms and grouped by anatomical region before any cross-lab comparison was attempted.
That mapping step is the unglamorous core of what a serious data analysis toolchain does. The underlying database in SEAZIT was built on PostgreSQL, with tables separated into substance information, screening data, and analyses, and the well table serving as the central connection point. Benchmark concentrations were calculated from incidence data summarized to the endpoint level, and a specificity score was derived by comparing the BMC of a non-mortality endpoint against the BMC of mortality, with the log10 difference used to separate specific developmental effects from non-specific toxicity.
The results of that interlaboratory work are a useful reality check for anyone shopping for tools. BMC values for the positive control substance varied across laboratories and exposure conditions, and the incidence rates for the vehicle control ranged from 0 to 8 percent depending on lab and scenario. No software layer eliminated that variation. What the harmonized pipeline did was make the variation visible and interpretable, which is a different and more honest goal than “standardized results.”
Where the output layers diverge, and what to actually compare
The table below puts the major tool categories side by side on the dimensions that matter when a screening campaign has to produce defensible potency calls, not just pretty videos.
| Tool layer |
Typical examples |
Primary output |
Handles ontology mapping |
Best fit |
| Behavioral tracking platform |
Zebralab, DanioVision |
Per-well activity, distance, response amplitude |
No |
Locomotor and photomotor endpoints |
| Image-based morphometrics |
FishInspector, ZembryoAnalyser |
Shape and heart rate features |
Partial or no |
Unbiased morphology quantification |
| Regulatory FET assistant |
ZebraFET |
LC50, NOEC/LOEC, validity flags |
No |
TG 236 compliance and reporting |
| Observation database with controlled vocabulary |
INTOB |
Structured phenotype and metadata records |
Yes, by design |
Long-term data management and reuse |
| Cross-lab analysis pipeline |
SEAZIT database and pipeline |
BMC, specificity scores, classifications |
Yes |
Interlaboratory comparison and harmonization |
INTOB, the Integrated Effect Database for Toxicological Observations from UFZ, is worth calling out separately because it was built around the FAIR principles of findable, accessible, interoperable, and reusable data, and it standardizes metadata and phenotypic observations through a controlled vocabulary. That is a data management answer to a data management problem, and it complements rather than competes with a tracking platform.
The deployment lesson from a real screening floor
A mid-sized contract screening group running developmental toxicity work on a few dozen compounds per month ran into a version of this problem that no single vendor demo would have surfaced. The behavioral data lived in the tracking software’s native export format. The morphology scores lived in a shared spreadsheet with free-text phenotype columns. The plate maps lived in a third place. When a sponsor asked for a concentration-response curve that combined locomotor and morphological endpoints for the same compound, an analyst spent the better part of a week reconciling well identifiers by hand, and the resulting curve was difficult to defend because the phenotype terms had been normalized post hoc rather than recorded against a controlled vocabulary from the start.
The fix was not a new imaging instrument. It was inserting a structured observation layer between the capture tools and the statistics, defining the phenotype vocabulary before the next campaign began, and requiring that every scored endpoint carry its well, plate, and run parameters in the same record. The tracking platform stayed. The spreadsheet did not.
This is where the hardware and the data layer intersect in a way that is easy to overlook when specifying a zebrafish toxicological exposure system. A workstation that logs well-level dosing parameters, imaging timestamps, and plate maps as part of its native operation removes a substantial reconciliation burden before the analysis stage even begins. Hunter Lab’s engineering work on fluid handling and plate-level control for zebrafish platforms has put it in a position to treat data capture as part of the exposure hardware rather than an afterthought bolted on by a separate vendor. The manufacturing discipline behind its zebrafish rack and workstation lines, tight tolerances on fluid paths and stable control electronics, translates into records that match what actually happened in the well, which is the only foundation a downstream analysis pipeline can build on.
The tools exist. The harder question is whether the stack is designed so that the data layer inherits the physical truth of the exposure, or whether the lab is left stitching it back together one spreadsheet at a time.