PerTurbo

in-silico perturbation · real targets
⬡ running live on AMD Instinct MI300X
⌄ scroll or click to enter
connecting…·kimi-k2 · AMD Instinct MI300X
Analysis

Ready

The workspace is quiet.

Ask PerTurbo a question and it will present its findings here — ranked targets, gene-level detail, spatial maps, and cross-sample comparisons. Each analysis is kept so a new question never erases the last.

scroll to zoom · drag to pan · click outside to close

Agent endpoint

URL of the Perturbo agent server. If it's unreachable, the dashboard automatically plays a built-in saved analysis so a demo never breaks.

PerTurbo

PerTurbo reads real Celcomen spatial couplings from one patient's pancreatic tumour and its liver metastasis, and an LLM agent walks you through which populations look like targets, which genes drive them, and what public databases already know. Everything it shows is a prediction that needs lab validation, not a proven target. Built for the AMD Developer Hackathon, Track 3.

Team
Method & infrastructure
EngineCelcomen, extended to learn signed, sparse gene–gene couplings across matched Visium sections. The model is symmetric, so the couplings are associations, not directed regulation.
DataOne PDAC patient: primary (T11) and liver metastasis (HM11). From GEO GSE272362 (Khaliq et al.).
Agent LLMkimi-k2 running on AMD Instinct MI300X, served through Fireworks and deployable with vLLM.
FramingEvery output is a hypothesis that needs experimental validation, never a proven target. The direction is laid over the data from literature, not learned from it.
PaperMegas et al., Celcomen, Nature Communications (2026) · arXiv:2409.05804.
External evidence · public databases
Open Targets
druggability
DGIdb
drug–gene
Europe PMC
literature
ClinicalTrials.gov
trials
⚠ Everything here is predicted, causal-given-model: prioritised in-silico hypotheses that need experimental validation, not proven targets. Couplings are undirected associations, and external records are shown exactly as returned (a drug or trial appearing here does not mean it works against this target). Every figure is generated live by the agent; none of it is hardcoded.

Export results

A complete, thorough report of every analysis this session — findings, tables, the comparison plot, spatial maps, the coupling network and the agent's own answers.