Skip to content

Aviv Regev (keynote ★) — when quantity becomes quality: toward the virtual cell and the lab-in-the-loop

Speaker: Aviv Regev — Genentech (formerly Broad Institute). (Long-form keynote + extended Q&A with Rahul Satija; the day was framed as a ~15-year retrospective of single-cell genomics.) Official title: "From Cell Atlases to Medicines, with AI" Slot: SCGD26, Jun 12, 2026 — 2:00–2:45 PM EDT keynote Recording: SCGD26 livestream (YouTube) — verify archived link Status: Captured from session transcript

Thesis

Single-cell genomics began as a quality problem (look at everything precisely). Regev's argument: at sufficient scale, quantity becomes a new quality — enough noisy data lets AI models systematically infer the rest. The arc runs from atlases → algorithms that find good representations → systematic perturbation ("lab in the loop") → multimodal generation → drug discovery / clinic in the loop, building toward a mechanistically grounded virtual cell.

Atlases + algorithms (find the biology)

  • Hundreds of millions of single human cells profiled (NucSeq, DroNc-seq, INTACT-seq lineage). Algorithms turn noisy scale into clusters (cell types), trajectories/optimal transport (dynamics), covariation (gene programs), and spatial/interaction structure.
  • sc-linker (Jagadeesh, Dey, … Alkes Price, Regev; Nat Genet 2022) links scRNA gene programs to GWAS/heritability across 60 diseases — cell-type, disease-progression, and pathway programs (e.g. Alzheimer's: monocyte program / microglial ApoE-ApoC1-TREM2 progression / apelin-signaling pathway).
  • Gut/ENS atlas: new dissociation + label-free enrichment recovered ~1,500 enteric neurons; many CNS-disorder risk genes (Hirschsprung, ASD, Parkinson's) are expressed in enteric neurons — a more accessible read-out tissue.
  • Aging brain atlas (Naomi Habib, Phil De Jager): 95 subsets; causal modeling orders microglia-12 → Aβ → microglia-13 → tau → astrocyte-10 → cognitive decline; patient-embedding splits a progressive-AD axis vs. an alternative-brain-aging axis with distinct multicellular communities.
  • DIALOGUE (Livnat Jerby-Arnon & Regev; Nat Biotechnol 2022): cell-type-specific multicellular programs that covary across niches/patients even from single-cell data.
  • Cross-cell-type QTLs (380k ILCs, 274 outbred mice): >half of trans-QTLs harbor genes expressed in a different cell type (NMU in enteric neurons → ILC2; SOX9 in enterocytes → IL-22 program) — causal effects across cells.

Lab-in-the-loop (perturb → model → design)

  • Random-sequence Oracle: ~100M random promoters in yeast → transformer (2018) → design extreme high/low expressers (Karl de Boer/Eilon Sharon lineage). Perturb-seq circuits for dendritic cells; in vivo Perturb-seq for ASD genes (15/35 converge on 6 programs across 4 cell types); brain-organoid ASD models (Levin/Arlotta) converging on accelerated GABAergic neurogenesis with genetic-background modifiers.
  • Autonomous agents: Biomni (with Jure Leskovec, Stanford; preprint) for autonomous biomedical analysis; a spatial agent (Genentech SpatialAgent / STAgent, preprint) that beats average human analysts and pairs best with one. (Phonetic transcripts "CoCo"/"Orion" not separately confirmed — the published spatial agent is SpatialAgent.)
  • Foundation models for generation across modalities:
  • SCHAF (Single-Cell omics from Histology Analysis Framework) — generate spatial transcriptomics / spatially-resolved scRNA from an H&E stain (paired/unpaired; metastatic breast, lung cancer) — "translating languages" H&E→RNA, leveraging the millions of existing H&E slides for pretraining.
  • PerturbView + MOSCATO — optical pooled screens (in situ sequencing of perturbations) in primary cells/tissue (cheaper, ~3–10× more perturbations/cells); AI ties imaging morphology to Perturb-seq RNA; generate unmeasured Perturb-seq from PerturbView, validated by a held-out secondary screen of 117 never-measured perturbations.
  • Iterative Perturb-seq — model-selected ~600 most-informative pairwise combinations per round beat random pairs at equal data (combinatorial space is astronomically large; 5-way over 20k genes = 10²¹ cells).

Clinic-in-the-loop (drug discovery → patient)

Targets (Recursion collaboration: pathways + small molecules), indications (patient samples for liabilities/specificity), modality (encoder/generative models for small molecules + antibodies), and a personalized BioNTech cancer vaccine (neoantigen presentation model; Phase-1 pancreatic — immune responders relapse-free significantly longer). Framing: drug R&D success LI→launch is ~1–2%, so even a 2× improvement is enormous; you're only as strong as your weakest link, so every step merits effort.

Q&A highlights (with Rahul Satija)

  • How many patients for multicellular communities? Not as many as feared — thousands, scaling with variant rarity / effect size; spatial data adds within-sample reproducibility; and H&E gives tens of millions of archival samples for pretraining.
  • Virtual-cell bar: a model whose internals are mechanistically anchored (causal), not just an accurate black box.

Connections to the corpus

Verify / open questions

  • MOSCATO as paired with PerturbView is not yet confirmed. A method named MOSCATO exists (supervised feature selection for multi-omic single-cell data, PMC9351124) but its identity as the imaging↔Perturb-seq integration layer Regev described is unverified — leave flagged.
  • Phonetic agent code-names "CoCo"/"Orion" not separately confirmed (published spatial agent = SpatialAgent/STAgent).
  • Exact patient-count framing for multicellular communities; archived YouTube recording link.

Sources

  • Event/schedule (title, affiliation, slot): https://satijalab.org/scgd26/
  • sc-linker — Nat Genet 2022: https://www.nature.com/articles/s41588-022-01187-9
  • DIALOGUE — Nat Biotechnol 2022: https://www.nature.com/articles/s41587-022-01288-0
  • SCHAF (H&E→omics) — bioRxiv 2023: https://www.biorxiv.org/content/10.1101/2023.03.21.533680v1.full
  • PerturbView — Nat Biotechnol 2024: https://www.nature.com/articles/s41587-024-02391-0
  • Biomni — biomni.stanford.edu / bioRxiv 2025: https://www.biorxiv.org/content/10.1101/2025.05.30.656746v1
  • SpatialAgent (STAgent) — Genentech / bioRxiv 2025: https://github.com/Genentech/SpatialAgent · https://www.biorxiv.org/content/10.1101/2025.04.01.646731