Discovery Engine (Disco) is an exploratory data analysis tool developed by Leap Laboratories, designed to discover novel and statistically validated patterns from tabular data, including feature interactions, subgroup effects, and conditional relationships. Rather than relying on pre-existing hypotheses, it starts from the data itself, searching for combinations of feature conditions that significantly change the target column.
Purpose
Disco is suitable for the following scenarios: wanting to know which variables are most important to target X, whether there are missed patterns, not knowing where to start when faced with data, and needing to understand how A and B affect C. It is not suitable for summary statistics, visualization, filtering, or SQL queries; for these, tools such as pandas should be used.
Main Capabilities
- Automatically discovers combinations of feature conditions, for example, "patients aged 45-65 with low HDL and high CRP have a 3x readmission rate."
- Every pattern is validated on a holdout set, increasing the likelihood of generalization.
- Performs FDR correction (Benjamini-Hochberg), provides p-values, and adjusts for multiple testing.
- Performs semantic search comparison against academic literature, helping to understand findings and assess novelty.
- Outputs structured results: conditions, effect size, p-value, citations, and a novelty classification for each pattern.
- Provides an LLM-generated narrative summary (result.summary).
- Generates an interactive web report (report_url) that visualizes all patterns.
How to Use
- Obtain an API key: can be created at disco.leap-labs.com/developers.
- Run the analysis: use the Python SDK to call discover(), which automatically polls and logs progress (queue position, estimated wait, current step, ETA). It can also be run asynchronously.
- Data preparation: requires a flat table, one observation per row, one feature per column, and one target column (at least 2 distinct values). Missing values can be handled automatically. Supports CSV, TSV, Excel (.xlsx), JSON, Parquet, ARFF, Feather, up to 5GB. Does not support images, raw text documents, nested JSON, or multi-sheet Excel.
- Exclude columns: identifiers, data leakage columns, tautological columns (e.g., when the target is serious, serious_outcome, not_serious, death, etc.).
- Can estimate before running: use discovery_estimate.
- Account management: Payment methods, subscription plans, and credit purchases can be added via the SDK or REST API.
- MCP server: No local installation required; provides tools such as discovery_list_plans, discovery_estimate, discovery_upload, discovery_analyze, discovery_status, discovery_get_results, discovery_account, discovery_signup, discovery_signup_verify, discovery_login, discovery_login_verify, discovery_add_payment_method, discovery_subscribe, discovery_purchase_credits, and more.
Use cases
- Healthcare: Discover risk factors for readmission in patient subgroups.
- Agriculture: Analyze factors influencing crop yield.
- Any field that needs to mine unknown patterns from tabular data.
Limitations
- Runs take several minutes.
- Public runs are free but results are public; private data requires setting visibility="private" and consumes credits.
- Does not support images, raw text, nested JSON, or multi-sheet Excel.
- Cannot replace EDA or AutoML; instead, it discovers patterns these tools miss.
- Requires an API key, and some features may involve fees (specific pricing not yet provided).
Technical details
Disco is a pipeline, not prompt engineering on data. It trains machine learning models, uses interpretability techniques to extract patterns, validates on held-out data with FDR correction, and then compares against academic literature via semantic search. It cannot be replicated with pandas code or by having an LLM look at a CSV.
Resources
- Python SDK on PyPI.
- Provides a Python SDK reference, Agent/MCP documentation, LLM-friendly reference, OpenAPI specification (in the repository), and a public report gallery.
- Example notebooks.
- SKILL.md provides detailed guidance on data preparation.
