Software / Models / Environments
MAGICS LAB · NORTHWESTERNScientific intelligence.Built to be used.
Foundation models. Research agents. Protein-evolution policies. Executable worlds. We turn ideas in scientific AI into systems others can use, extend, and build on.
01Genome foundation models
Code & model weights availableDNABERT / DNABERT-2 / GenomeOcean
From understanding DNA
to generating it.
A sustained line of genome AI: DNABERT brings language-model pretraining to DNA; DNABERT-2 extends representation learning across species; GenomeOcean adds generative modeling at the 4-billion-parameter scale. Together, these projects give researchers reusable starting points for sequence analysis and generation.
GenomeOcean is developed with the DOE Joint Genome Institute and collaborators.
02Scientific research agents
Public interface & open-source codeSciSciGPT
From a research question
to an analytical workflow.
An AI collaborator for the science of science. SciSciGPT coordinates literature retrieval, scholarly data access, executable analysis, visualization, and evaluation—bringing specialist agents together around a researcher's question.
Published in Nature Computational Science ↗
A Northwestern CSSI collaboration.
03Protein-evolution policies
Research previewPEV
The next protein edit.
One forward pass.
A general mutation policy that uses a parent protein and assay context to score allowed substitutions in parallel—or choose STOP. PEV learns from measured outcomes across protein families, with calibrated estimates of whether an edit will improve the parent.
Current evidence is strongest for folded domains. See the project page for evaluation settings and limitations.
Candidate-scoring throughputProject-reported, at 1,000 candidates per request,
versus a matched per-mutant encoder.
04Executable environments
Preprint · Open-source release plannedMagicSim
Where an agent's plan
becomes an executed episode.
A unified infrastructure for embodied interaction. MagicSim brings world construction, robot execution, evaluation, and multimodal trajectory collection into a shared runtime—grounding high-level commands in actions, rather than simulator-side shortcuts.
Specify executable worlds
YAML-defined scenes, behaviors, sensors, and robot embodiments.
Evaluate in a shared runtime
Benchmarking, reinforcement learning, and agent-facing interaction.
Collect grounded experience
Structured trajectories connecting commands, actions, observations, and outcomes.
The paper is available now. Open-source software and environments are planned; release links will be added here when available.
Build on the science.
Help shape what comes next.
Explore the releases, reproduce the research, or work with us on the next generation of models, agents, and scientific environments.
Selected projects developed with our collaborators. Resource links lead to the official repositories, model pages, and publications. DNABERT download milestone reported by the lab, September 2026. Availability is shown for each project.