iiinova — where your digital self keeps evolving

Romain Lopez @biologicalml.org · 01/06/2026

This was a thrilling project with wonderful collaborators Taka Kudo, Ana Meireles, Antonio Rios, JC Huetter, @minetoota.bsky.social, Levi Garraway, @katiegeiger.bsky.social, @avtarsingh.bsky.social, @jkpritch.bsky.social , and Aviv Regev.

040

Romain Lopez @biologicalml.org · 01/06/2026

PerturbPair establishes a generalizable framework for multimodal perturbation atlases. EB-MoCAVI is available as part of scvi-tools: github.com/Genentech/Mo... Explore our study to see how cross-modal integration can make perturbation screening both more scalable and more informative.

130

Romain Lopez @biologicalml.org · 01/06/2026

Integrating measured + imputed profiles with rare-variant burden statistics from UK Biobank, we identified disease-specific macrophage gene programs and key regulatory nodes for monocyte counts, T2D, and IBD. This paves the way towards trait-linked cellular programs.

110

Romain Lopez @biologicalml.org · 01/06/2026

Did EB-MoCAVI generate accurate Perturb-Seq profiles? We ran a secondary Perturb-seq screen on 288 perturbations (including 117 with only imputed profiles). The results confirmed that imputed profiles closely match measured RNA (median Spearman ρ = 0.48).

110

Romain Lopez @biologicalml.org · 01/06/2026

The imputed profiles are biologically coherent: Lamtor3 clusters with measured Lamtor1/4/Rraga (mTORC1), V-ATPase subunits group together, etc. We also found that iron-sulfur assembly pathway components act as a restraint on macrophage interferon tone (a prediction validated experimentally).

110

Romain Lopez @biologicalml.org · 01/06/2026

To exploit these cross-modal relationships, we developed an empirical Bayesian variational inference framework (EB-MoCAVI). Using OPS embeddings as an informative prior, it (1) impute RNA profiles for perturbations measured only by imaging (2) denoise Perturb-seq estimates via adaptive shrinkage.

120

Romain Lopez @biologicalml.org · 01/06/2026

OPS detected substantially more hits than Perturb-seq, partly because it had 10x more cells per perturbation and partly because some PerturbView-only hits included post-transcriptional events invisible to RNA (e.g., mTOR perturbations → elevated phospho-rpS6)

120

Romain Lopez @biologicalml.org · 01/06/2026

The relationship between perturbation profiles was highly concordant across modalities (ρ = 0.85). The same regulatory modules emerged from both, with informative differences: for instance, m6A modification genes cluster with JAK-STAT only in imaging, suggesting post-transcriptional convergence.

110

Romain Lopez @biologicalml.org · 01/06/2026

We profiled LPS-stimulated mouse BMDMs with: (1) 334,000 single-cell transcriptomes across ~1,000 gene perturbations (Perturb-seq) (2) 7.8 million imaging-phenotyped cells across ~3,000 gene perturbations (PerturbView) Both libraries share a common backbone, enabling direct cross-modal comparison.

110

Romain Lopez @biologicalml.org · 01/06/2026

Perturb-seq reveals molecular programs; OPS captures morphology, protein localization, and cell biology at massive scale. But how concordant are these readouts? Can we exploit their relationship to generate one modality from the other? We set out to answer both questions in a single system.

110

Romain Lopez @biologicalml.org · 01/06/2026

🚀 We are introducing PerturbPair (with Taka Kudo) — a platform that combines parallel Perturb-seq and optical pooled screening (PerturbView) in primary cells to systematically map at massive scale how genetic perturbations reshape cellular states across modalities. www.biorxiv.org/content/10.6...

1279

Romain Lopez @biologicalml.org · 29/05/2026

Wonderful work with collaborators Kelvin Y. Chen, @basakeraslan.bsky.social , @anshulkundaje.bsky.social je, @jkpritch.bsky.social , Aviv Regev and Shimon Sakaguchi.

020

Romain Lopez @biologicalml.org · 29/05/2026

Our work establishes a generalizable framework for multi-modal chemical screening at single-cell resolution. MoCAVI is available as part of scvi-tools: github.com/Genentech/Mo... Paper: www.biorxiv.org/content/10.6...

120

Romain Lopez @biologicalml.org · 29/05/2026

CRISPR validation confirmed it: RUNX3 KO strongly reduced GZMB induction by the HDACi abexinostat while TBX21 KO selectively attenuated IFNG — distinct, collaborative roles exactly as predicted.

100

Romain Lopez @biologicalml.org · 29/05/2026

PERCISTRA revealed distinct regulatory mechanisms of BET and HDAC inhibitors. BET inhibitors suppress AP-1 activity and promote a TCF7-driven naive T-cell program, while HDAC inhibitors activate RUNX3/TBX21-associated enhancers, increasing GZMB and IFNG expression and promoting cytotoxic programs.

110

Romain Lopez @biologicalml.org · 29/05/2026

We also introduce PERCISTRA, a framework to infer regulatory networks from perturbation multi-omics data. By integrating coordinated changes in chromatin accessibility and gene expression, it identifies upstream TF regulators, their regulatory loci, and downstream target genes.

141

Romain Lopez @biologicalml.org · 29/05/2026

Wonderful collaboration with Kelvin Y. Chen, @basakeraslan.bsky.social , @anshulkundaje.bsky.social , @jkpritch.bsky.social , Aviv Regev and Shimon Sakaguchi.

000

Romain Lopez @biologicalml.org · 29/05/2026

Applying MoCAVI to the DOGMA-Plex screen, we found that BET inhibitors promoted a ‘naive-like’ T cell state, whereas HDAC inhibitors drove a strong cytotoxic phenotype. These results demonstrate how distinct drug classes can skew T cell fate toward fundamentally different functional programs!

100

Romain Lopez @biologicalml.org · 29/05/2026

In the DOGMA-plex screen, we were able to see how different drug classes differentially affected chromatin accessibility and how this eventually propagated into effects on both gene and protein expression.

100

Romain Lopez @biologicalml.org · 29/05/2026

By co-embedding chemical and genetic perturbation data with MoCAVI, we found that ERK5-in-1 phenocopied BRD4 knockout in a dose-dependent manner — the first functional evidence for an off-target interaction previously only suggested by in vitro binding assays.

110

Romain Lopez @biologicalml.org · 29/05/2026

MoCAVI groups compounds into drug modules by shared phenotypic effects, often matching known targets but with exceptions. Crizotinib clustered with JAK/STAT inhibitors rather than other molecules targeting their annotated targets, potentially a novel off-target annotation at high doses.

110

Romain Lopez @biologicalml.org · 29/05/2026

A core challenge: perturbation effects are often subtle compared to baseline cell state heterogeneity. Existing methods conflate the two. We developed MoCAVI — a contrastive variational framework that separates drug-specific effects from natural variation across any combination of modalities.

130

Romain Lopez @biologicalml.org · 29/05/2026

We developed two platforms: (1) icCITE-plex: RNA + 477 surface proteins + 64 intracellular epitopes. 56 kinase inhibitors × 3 doses in human + mouse T cells (~95K cells). (2) DOGMA-plex: RNA + ATAC + proteins. 290 epigenetic inhibitors × 4 doses in human T cells (~314K cells).

120

Romain Lopez @biologicalml.org · 29/05/2026

Single-cell chemical transcriptomics can now screen thousands of compounds, but only at the RNA level. Mechanisms-of-action are often mediated through protein modifications, chromatin remodeling, and signaling cascades that RNA can't capture directly. We set out to add those missing layers at scale.

130

Romain Lopez @biologicalml.org · 29/05/2026

We built a joint experimental and computational platform for scalable multi-modal single-cell chemical screens — profiling RNA, protein (including phospho-signaling), and chromatin accessibility responses to thousands of small molecule perturbations in parallel. www.biorxiv.org/content/10.6...

14016