New matverse is now open source

Infrastructure for
AI-native biology

Open analysis libraries that AI agents look up and call — tested, registered functions instead of scripts written from memory. One family, five sciences.

Explore the verses View on GitHub
ov.omicversemv.matversesv.socialversepv.physverseastv.astroverse
pbmc / single-cell
import omicverse as ov
adata = ov.read('pbmc8k.h5ad')
ov.pp.qc(adata, tresh={'mito_perc': 0.2})
ov.pp.preprocess(adata, mode='shiftlog|pearson')
ov.pp.scale(adata); ov.pp.pca(adata, n_pcs=50)
ov.pp.neighbors(adata); ov.pp.umap(adata)
ov.pp.leiden(adata, resolution=1)
# bulk · single-cell · spatial, one AnnData
✓ 8,381 cells · 14 clusters
omicverse
Biology · bulk, single-cell & spatial omics
Explore
screen / battery_cathodes
import matverse as mv
md = mv.datasets.load('battery_cathodes')
mv.pp.standardize(md)
mv.pp.qc(md) # obs['is_valid']
md = mv.pp.filter_materials(md)
mv.feat.element_stats(md)
mv.calc.relax(md, level='emt')
mv.thermo.hull(md, level='emt')
mv.screen.filter(md, e_above_hull_emt__lt=0.05)
✓ still an ordinary AnnData
matverse
Materials · query → relax → rank
Explore
Did the county minimum-wage increase change teen employment?
▸ registry_lookup "staggered difference-in-differences"
▸ check parallel trends · event study
Treatment is staggered across 2014–2019, so I used a heterogeneity-robust estimator. Pre-trends are flat; the post effect is small and not significant.
Write up the evidence chain.
Done — design, diagnostics, estimates and figures are recorded in one object, ready to audit and reproduce.
socialverse
Social science & humanities
Explore
many-body / heisenberg
import physverse as pv
pv.utils.registry_lookup("spin glass ground state")
# quantum many-body: neural quantum states
H = pv.quantum.heisenberg(pv.quantum.lattice_from_bonds("hamiltonian.json"))
rbm = pv.quantum.init_rbm(H.n_sites, 2 * H.n_sites)
res = pv.quantum.vmc_optimize(rbm, H, metric_samples=16384)
# spin glasses, judged against a certificate
cb = pv.opt.cluster_lower_bound(m, cells)
gs = pv.opt.parallel_tempering(m, ladder, target=cb.bound)
# PDE discovery: library → sparse regression
fit = pv.pde.select_model(lib.Theta, ut, lib.names)
physverse
Physics · many-body, statistical, HEP inference
Private preview
exoplanets / transit search
import astroverse as astv
astv.utils.registry_lookup("transit search") # find, do not guess
grid = astv.ts.transit_period_grid(t, minimum_period=1.0, maximum_period=10.0)
res = astv.ts.bls(t, flat, dy, grid)
vet = astv.ts.vet_transit(t, flat, dy, res.best_period, res.best_t0, res.best_duration)
# cosmology: 3×2pt with scale cuts
d = astv.cosmo.theory_3x2pt(params, lay, zs, nzs, zl, nzl)
chain = astv.infer.run_emcee(like.log_prob, p0, n_steps=5000)
astv.infer.convergence_report(chain) # R-hat, ESS, tau
✓ catalog = astv.pp.from_lightcurves(lcs) # an AnnData
astroverse
Astronomy · time series, cosmology, inference
Private preview
One substrate

Cells, crystals, light curves, survey respondents. Same object.

Every verse keeps its data on AnnData. Observations, annotations, descriptors and results travel together, and write to h5ad / zarr — no new file format, no lost rows when you merge, filter or featurise twice.

.X measurements
.obs cells · materials · sources · units
.var genes · elements · bands · items
.obsm embeddings · structures · light curves
.uns provenance · run records

Capability lives in the library, not in the prompt.

Agents are already good at reading papers and writing code. What they lack is a trusted, queryable toolbox for each science. That is what we build.

01

Registered for lookup

Every public function is registered with what it does and what it needs, so an agent asks before it writes code.

02

Tested against a reference

Each function is checked against an independent implementation or known result — not just that it runs.

03

Executed in a tutorial

Every function appears in a tutorial that actually runs, so usage is shown, not described.

04

Grown from failures

When an agent fails a held-out research task, the missing method goes into the library — and is tested on tasks it has never seen.

Askregistry_lookup("transit search")
→ 3 matches, with signatures
Callastv.ts.bls(t, flux, dy, grid)
→ tested result object
Checkvet_transit(...)
convergence_report(chain)
→ diagnostics, not vibes
Recordcatalog.uns["provenance"]
→ every step reproducible

Built in the open,
one science at a time.

omicverse · matverse · socialverse · physverse · astroverse

github.com/scienceinfra