Bringing AI drug discovery tools to underserved labs
Drug development is slow and expensive — a single new medication typically takes around 10 years and $1.3 billion to bring to market. Because of that cost, pharmaceutical companies tend to focus on diseases that promise a strong return on investment in high-income countries, leaving work on infectious diseases prevalent in lower- and middle-income regions underfunded.
AI and machine learning models have emerged as a way to compress that timeline. Investments in AI/ML for drug discovery have soared over the last five years, and scientists use these models to better identify promising drug candidates before expensive lab work begins. But access to those models is not evenly distributed: running them requires data science expertise and computational resources that many research groups simply don't have.
The nonprofit Ersilia Open Source Initiative is working to close that gap. Through its Ersilia Model Hub, the organization publishes AI/ML models relevant to antibiotic drug discovery as public GitHub repositories, giving researchers anywhere in the world free access to computational tools that would otherwise be out of reach.
Who benefits from the hub
The hub is aimed squarely at researchers outside Europe and North America. Most of the people contributing models are data scientists and developers; the people running those models are mostly biomedical researchers at institutions across Sub-Saharan Africa. Ersilia plans to establish the hub at 15 African institutions over the next two years.
Its largest partner is the University of Cape Town's Holistic Drug Discovery and Development (H3D) Centre, founded in 2010 as Africa's first integrated drug discovery center. Ersilia also works with newer groups like the University of Buea's Center for Drug Discovery in Cameroon, which launched in 2022 with seed funding from a Bill & Melinda Gates Foundation fellowship and now has 25 members.
Ersilia co-founder Miquel Duran-Frigola, PhD, describes the model hub as a practical starting point for resource-constrained teams. Computational models are relatively cheap to run compared with laboratory experiments, so researchers can use them to simulate how candidate molecules might treat a disease before committing to wet-lab work.
How models get into the hub
The hub currently holds about 150 models. Most are open source and pulled from scientific literature. For example, biochemists at the David Braley Centre for Antibiotic Discovery contributed a model that predicts whether a chemical compound is likely to inhibit Acinetobacter baumannii, a drug-resistant pathogen often acquired in healthcare settings. Ersilia also builds some models in-house, including one that predicts the efficacy of chemicals against lab-grown Mycobacterium tuberculosis using data from Seattle Children's Hospital.
Ersilia's technical team manually approves each model before it enters the hub, but the submission process itself is automated with GitHub Actions. Contributors request a model by opening an issue, and a self-service process collects key information:
- The model's schema — what input it expects and what output it returns.
- Open source license information.
- Whether the model runs on CPUs or GPUs.
- A link to the model's open source code.
- A link to the peer-reviewed paper or preprint.
- Labels describing the use case, with tags like
malaria,classification,regression,unsupervised, orsupervised.
Once Ersilia approves a model, the contributor submits a pull request. That triggers a test suite in GitHub Actions; when the tests pass, the pull request is merged and the model is added to the hub. Researchers can then pull ready-to-use models from their public repositories, submit candidate molecules, and get predictions back about how well those molecules perform against specific diseases — all free and entirely online.
The process is transparent by design. Users should note that any activity on the repository, including predictions stored as action artifacts, is public.
Loosening the bottlenecks
Duran-Frigola says Ersilia is currently focused on disseminating models and information. Next, the team wants to attach confidence metrics to each model's predictions. And as the user base grows, Ersilia could aggregate the inputs researchers submit, building a picture of which candidate molecules are being tested against infectious diseases.
That sort of aggregated data could be powerful. By seeing which drugs are already available in various countries, researchers could explore repurposing them against different microbes — potentially treating neglected diseases without starting from zero.
Automating model access with GitHub Issues and Actions
Ersilia's breakthrough came when GitHub's customer success architect Jenna Massardo and social impact engineer Sean Marcia stepped in to build a self-service system. The work was organized through GitHub's Skills-Based Volunteer program, which pairs employees with social sector organizations on strategic and technical projects.
Massardo began by getting the Ersilia Model Hub running locally to understand the researcher experience. From there, she and Marcia broke the problem into stages: How would a researcher request a model? How would input data be processed? What notifications should users receive?
The solution: a standard IssueOps pattern that leverages GitHub issues to trigger GitHub Actions. Massardo deliberately ruled out a pull request flow because researchers would be adding files for processing rather than contributing code changes.
"People are contributing to the repository but they're not actually making code changes. Using pull requests would have meant a lot of noise in the repository's history. But issues are perfect for this sort of thing."
Researchers, biologists, and students can initiate runs by creating an issue from a template that asks for the target model and candidate molecules in SMILES format (Simplified Molecular Input Line Entry System), a computer-readable standard for representing molecular structures.
A series of simple workflows
Early plans called for a custom GitHub Action, but Massardo — who has written several — recommended against it due to the maintenance burden of managing dependencies and security updates. Instead, she scoped the work as a series of individual tasks using readily available actions from GitHub Marketplace.
When a researcher files an issue, the following automated process kicks off:
- GitHub Actions provisions a GitHub-hosted runner to execute the workflow.
- The GitHub Issue Forms Body Parser action parses issue content from Markdown into structured data.
- The workflow retrieves the requested model and invokes Ersilia's software.
- The software configures the model and writes the user's input to a file for processing.
- A CSV output is generated and saved as an artifact in GitHub Actions.
- The workflow comments on the issue with a downloadable link to the artifact.
Artifacts follow a 30-day retention policy. Five days before expiration, stale bot reminds users to download results, and at the end of the period, it automatically closes the issue. Massardo highlights that the entire lifecycle happens on GitHub — researchers simply submit an issue and receive completed output, with no maintenance burden on Ersilia staff.
Docker-based model distribution
Ersilia standardizes model packaging with a Dockerfile template. When a new model is submitted, the template is copied to its repository, triggering a Docker image build on GitHub-hosted runners. The resulting images live in public repositories where GitHub Actions runs at no cost, enabling free model execution for researchers. Models are optimized for CPU environments, ensuring they can also run locally on researchers' own machines.
Ersilia CTO Dhanshree Arora notes that the models are purpose-built for specific use cases and therefore compact. The team is working to further shrink image sizes to reduce network transfer overhead, disk usage, and container spin-up times. Docker packaging also creates consistent, reproducible environments that simplify collaboration between researchers.
Updating the model list daily
A maintenance workflow Massardo built keeps the issue template's model selection current. Every day, it fetches the Ersilia team's master model list, runs a Python script to extract new, updated, or deprecated entries, and refreshes the issue template.
"This is another example of how we built this process to be as hands-off as possible while still making it as easy as possible for researchers to actually use the tool," Massardo says.
Expanding the open source mission
Ersilia remains committed to open source development and was drawn to GitHub Actions because it is free in public repositories. Founder Duran-Frigola is now exploring creative applications for GitHub Actions beyond CI/CD to broaden researcher access. He also wants Ersilia's interns to gain familiarity with GitHub Copilot and modern AI-assisted development workflows.
Contributing tips for newcomers (from Massardo and Ersilia's contribution guidelines):
- Choose a project you genuinely care about — personal interest keeps you engaged.
- Scan repository issues for approachable improvements — many projects tag these with
good first issue. - Plan to iterate — maintainers may expect small contributions first, clarity on edits for distributed collaborators, and early discussion via issues before major changes.
Ersilia was recently designated a Digital Public Good (DPG) by the United Nations. DPGs are open source solutions — including software, data, AI systems, and content — that are freely accessible, address societal challenges, and promote sustainable development. Outreach for contributing to Ersilia and other DPGs runs through For Good First Issue, GitHub's curated directory of recognized projects seeking technical volunteers. The tool connects resource-constrained nonprofits with technologists worldwide who can drive positive change.
