README

Description

This application is available on Nprod.net

The dataset is from the LOTUS initiative and Wikidata.

Requirements

• Dash for its web ui.
• EPAM's Ketcher for the molecule editor.
• Rdkit for the molecule massaging.

Install & use

The docker compose way (recommended)

docker-compose run -it backend python update.py  # You do not have to run it everytime, but that's recommended until we have a schema versioning
docker-compose up --build

The web server is then available on http://localhost:3000 and the API on http://localhost:5000.

Run tests

make tests

The manual way

To run it yourself, the source is available at: https://github.com/lotusnprod/lotus-search:

• Install dependencies using uv
  • If you do not have uv installed:
    • curl -LsSf https://astral.sh/uv/install.sh | sh
  • Then:
    • uv sync
• Run python update.py (takes a few minutes)
• Run uvicorn main:app --reload (almost instant)

Performance & Benchmarks

Lightweight, opt-in micro benchmarks are available to help detect regressions in common hot paths (taxonomy lookups, structure searches, SDF extraction).

Run after preparing data (i.e. after python update.py has generated the data/ directory):

make bench

Environment variables:

• BENCH_ITER (default 200) controls repetition count for cached taxonomy lookups.

Sample output (JSON lines):

{"benchmark": "taxa_name_matching_first", "seconds": 0.0123, "size": 42}
{"benchmark": "taxa_name_matching_cached", "seconds": 0.00001, "size": 42}
{"benchmark": "structure_similarity", "seconds": 0.0042, "size": 1200}

Caching / Optimization notes:

• Several pure taxonomy helper methods now use functools.lru_cache to reduce repeat DB round-trips during interactive browsing. These caches are per-process and safe because the underlying dataset is immutable at runtime.
• SDF block retrieval is streamlined to a single join over memory-mapped slices—same ordering/content, fewer Python allocations.
• A single DataModel instance is re-used across Dash pages (legacy import path via model.model retained for compatibility) to avoid redundant large in-memory loads.

All changes preserve existing API outputs and test expectations (see test suite).

Authors

Adriano Rutz, is the one that pushed me into that, we are both part of the team behind LOTUS.

Jonathan Bisson

• Personal website: https://www.bjonnh.net
• You can find me on mastodon too(t): https://mastodon.social/@bjonnh

Adriano Rutz

• Personal website: https://adafede.github.io
• And on mastodon: https://mastodon.online/@adafede

Data safety

Your molecules are never stored unless they make our service crash.

But if your molecules are super secret, it is like your extremities, don't insert them in machines you don't know or understand.

Also, this is an experimental tool meant to test things, you're not supposed to rely on it for anything important, and it comes with no warranty or support whatsoever.

License and legalese

https://raw.githubusercontent.com/lotusnprod/lotus-search/main/LICENSE