Ersilia Model Hub

The Ersilia Model Hub is the flagship project of Ersilia Open Source Initiative, a Spanish-Catalan foundation registered as a nonprofit and headquartered at Norrsken House in Barcelona. The organization's stated mission is to strengthen research capacity against infectious and neglected diseases in low- and middle-income countries by making predictive ML tools available without cloud dependency. The hub catalogs models for antibiotic activity prediction, ADMET (absorption, distribution, metabolism, excretion, toxicity) properties, molecular representation, and generative chemistry, drawn from published literature and from custom models the team develops or accepts via contribution templates (eos-template, eos-demo).

Technically the hub is a CLI-driven model registry written in Python (Python 3.8+). Each model ships in standardized form so users can run a small set of commands (`ersilia fetch` to download, `serve` to start, `run` to score a CSV of inputs, `close` to stop, `delete` to remove) without writing ML code. Two packaging paths are supported: a Conda-based source install and a Docker-based container path that the team recommends for production. The catalog is browsed at catalog.ersilia.io and pulled locally so inference runs on the user's own machine rather than against a hosted API.

The deployment story is the differentiator: Ersilia maintains in-country data-science units at the H3D Centre at the University of Cape Town, the Centre for Drug Discovery at the University of Buea, and IMPM in Yaounde, with the explicit goal of fitting an automated virtual-screening cascade onto local infrastructure rather than US cloud platforms. The organization is funded by a mix of Schmidt Sciences, the Bill and Melinda Gates Foundation, the NIH, the Spanish government, and the EU, with the Mozilla award as one stream among many.

The Mozilla Builders Accelerator selection on September 23, 2024 placed Ersilia in the first cohort of 14 projects, each receiving up to $100K in non-dilutive funding plus a 12-week build-sprint program ending in a San Francisco demo day. Within the open-source AI stack, the hub sits at the weights and retrieval-memory layers: it does not train new architectures but packages pretrained models so domain scientists who do not maintain ML infrastructure can pull, run, and reuse them offline.