Agentomics

Agentomics takes a folder-based dataset split and an optional data description as input, then autonomously experiments with different machine learning models and strategies to produce a trained model ready for inference, along with a PDF report summarizing the development process and the metrics achieved. Made for biomedical data, it is reported to outperform human experts and produce new state-of-the-art models for problems in protein engineering, drug discovery and regulatory genomics, and was presented at ISMB 2026 and published in the Bioinformatics journal.

It can be driven by a local Codex subscription through its own login, by a supported provider API key such as OpenRouter, OpenAI or Anthropic, or by local Ollama models for fully offline, private runs. Agents execute code inside a Docker container with read-only mounts to the host filesystem and are only allowed to write to a Docker volume, and when a held-out test set is supplied, the agent is fully prevented from accessing it, so reported test-set metrics come from a separate, programmatic evaluation rather than the agent's own claims. Outputs include the trained models, scripts and conda environments needed to reproduce inference or retraining with one command.