CATOCATO

Research Tools

Scientific computing in isolated sandboxes with Python, R, persistent notebooks, durable jobs, and versioned outputs.

Ask in scientific terms. CATO can select tools, but you remain responsible for checking the method, package versions, assumptions, reference data, generated code, and interpretation.

Execution Environments

Durable background job

Best for longer or multi-file runs. It persists independently of the browser and reports progress and artifacts.

Persistent notebook kernel

Best for iterative Python or R exploration where variables and loaded data should remain available between cells.

One-shot analysis

Best for bounded tasks whose code, output, and files can be completed in a single coordinated run.

Execution is isolated from the application services. Project identity and authorized object references are checked when source files move into a sandbox and when produced artifacts return to the project. Available compute, networking, storage, and run duration are governed by current production controls.

Python and R

CATO supports both Python and R workflows. It can generate scripts or notebooks, run code, capture console output, and persist tables, figures, serialized results, and environment information. R packages used by common biomedical and statistical workflows are available alongside the Python stack.

Specify the language in your request: write “Run this analysis in Python” or “Run this analysis in R” when the language matters. Include any required package, version, output format, or reproducibility constraint in the same request.

  • Specify a language when package choice or reproducibility requires it.
  • Ask CATO to check whether a package and version are present before designing a long workflow around it.
  • Record random seeds and deterministic settings where applicable.
  • Save intermediate analysis-ready data rather than relying only on in-memory notebook state.
  • Export the notebook, scripts, dependency information, and outputs needed to reproduce a result.

Scientific Capability Areas

Core data science and machine learning

pandas, NumPy, SciPy, statsmodels, scikit-learn, gradient boosting, model diagnostics, visualization, and reproducible reporting.

Genomics and single-cell

Sequence and variant formats, BioPython, pysam, scanpy, anndata, differential expression, clustering, dimensionality reduction, and pathway workflows.

Proteomics and mass spectrometry

mzML/MGF handling, pyOpenMS, pyteomics, quality control, normalization, peptide/protein summaries, and downstream statistics.

Medical imaging and signals

DICOM and NIfTI handling, pydicom, nibabel, SimpleITK, image metadata, segmentation-oriented workflows, and signal processing.

Molecular and chemical analysis

RDKit, structural files, molecular descriptors, similarity, substructure workflows, sequence/structure utilities, and interactive 3D views.

Causal, survival, and longitudinal analysis

Survival models, time-to-event diagnostics, mixed and longitudinal models, causal diagrams, effect estimation, sensitivity analysis, and time series.

Networks, pathways, and ontologies

Graph analytics, enrichment, gene sets, pathway databases, community structure, centrality, and biological identifier lookups.

Literature and external research services

18-source scholarly search plus specialized clinical-trial and patent connectors, metadata resolution, open-access retrieval, and evidence synthesis.

This is a capability map, not a guarantee that every package or external service is available in every environment. Licensing, provider availability, plan controls, and production configuration can affect a tool.

A Reliable Analysis Workflow

  1. Define the estimand or output. Name the population, variables, units, and decision the result will inform.
  2. Inspect before transforming. Profile schemas, missingness, duplicates, ranges, and coding conventions.
  3. Approve a plan. Use Plan mode for exclusions, transformations, models, multiplicity, or evidence criteria.
  4. Run with traceability. Save code, parameters, environment, logs, and machine-readable results.
  5. Diagnose. Check assumptions, convergence, residuals, robustness, leakage, and uncertainty.
  6. Visually inspect. Verify every table, figure, structure, or viewer state against source data.
  7. Reproduce. Re-run critical outputs from a clean state and compare hashes or tolerances.

Notebooks and Persistent Kernels

Notebook artifacts open in the Workspace with editable cells and captured output. A persistent kernel supports iterative analysis, while notebook versions preserve meaningful checkpoints. UseCtrl/Cmd+Enter to run the focused cell.

Kernel state is convenient but implicit. A notebook that only works after cells are run in a particular undocumented order is not reproducible. Restart and run from a clean state before treating it as a final analysis record.

Figures and Interactive Workbenches

CATO preserves figures as workspace artifacts rather than disposable chat images. Supported figure types open with domain-appropriate controls; editable SVG figures can be refined without losing the original, while 3D and scientific views retain their interactive scene state where supported.

  • Confirm plotted values, transformations, scales, units, and denominators.
  • Use version history to separate analytic changes from cosmetic edits.
  • Check labels, clipping, legends, contrast, and raster/vector export quality.
  • Save the data and code behind a figure along with the visual artifact.

Artifacts and Provenance

A completed scientific run can produce datasets, scripts, notebooks, figures, structured results, reports, or manuscripts. CATO persists those outputs with project and conversation relationships, version history, and available execution metadata. The Workspace restores them when you return.

Treat provenance as a chain: source version → transformation/code → runtime/environment → result version → figure/table/manuscript claim. If one link is ambiguous, resolve it before publication.

External Lookups and Literature

Scientific lookups and literature connectors depend on external source availability and source-specific identifiers. Record accession numbers, database versions or access dates, queries, and source URLs. A transient connector failure should not be interpreted as evidence that a record does not exist.

Method and Safety Boundaries

  • CATO can assist with research analysis; it does not replace clinical judgment, a validated diagnostic device, or regulatory review.
  • Generated code can contain statistical, computational, or security errors. Inspect it before execution or reuse.
  • Do not provide secrets to code or notebooks. Use approved production integrations for credentials.
  • Confirm software licenses and data-use restrictions before redistributing an environment or output.
  • Escalate high-stakes findings to the appropriate statistician, domain expert, privacy officer, or safety reviewer.