Query scientific databases without an API key

What this involves

Querying public biomedical repositories often requires navigating fragmented interfaces, managing diverse query structures, and handling varied data schemas across resources such as UniProtKB, Ensembl, ChEMBL, and gnomAD. When exploring structured scientific datasets, an incorrect assumption about dataset configuration, table splits, or column types frequently returns silent, misleading zeroes or missing records rather than an explicit error.

Doing this work properly demands a systematic discovery process. Before running queries, datasets must be systematically inspected to map specific configurations, partition splits, row counts, and underlying column types. Checking column statistics beforehand reveals issues such as sentinel text strings in numeric fields, heavy class imbalances, or empty strings masquerading as valid data.

The research agent carries out this process remotely on its own machine. It searches public resources, inspects data cards and schemas, calculates summary statistics across candidate fields, executes precise selections, and returns fully attributed numbers paired with their provenance, dataset configuration, and sample size.

How it goes

1. 1Discovery and candidate searchThe agent executes targeted searches across public databases and datasets using find_datasets or find_models, querying by specific subject, biological task, or established community names such as PDBbind, MoleculeNet, or TDC.
2. 2Schema and split inspectionBefore constructing a query, the agent calls inspect_dataset to identify available configurations, splits, total row counts, and typed column definitions.
3. 3Column profiling and data validationThe agent runs dataset_column_stats on candidate filter and reporting columns to check null fractions, cardinality, min and max bounds, and sentinel values.
4. 4Direct query and source retrievalThe agent queries targeted endpoints—including UniProtKB, Ensembl, ChEMBL, ClinVar, gnomAD, GTEx, Open Targets Platform, Reactome, STRING, and GWAS Catalog—projecting explicit columns and unnesting complex structures.
5. 5Data attribution and summaryReturned figures are structured into tables or figures, citing repository owner, dataset name, configuration, split, denominator, and license.

Start it here

Ask in plain language. The agent works on a Linux machine of its own — it installs and runs what the job needs, keeps every file, and reads the results back with the numbers explained. Nothing to set up, and a long run continues after you close the page.

What to watch for

• Multi-config datasets have no usable default configuration, so guessing 'default' frequently queries empty data and misreports it as zero.
• A bare 'SELECT *' pulls megabytes of binary data per row into standard output when image or audio columns typed as struct{bytes, path} are present.
• Integer label columns represent class indices rather than descriptive names, which must be extracted from inspect_dataset features to avoid reporting raw indices.
• Columns typed as strings often contain sentinel characters such as '<0.01' or 'ND' that cause failures if cast directly to numeric types.
• Calculating percentages from queries that use a LIMIT clause misrepresents samples as true population distributions.
• Naming a model from unverified memory frequently causes 404 errors or loads an incorrect checkpoint.

What it runs

Questions