Z-Screen Pilot Release/ Preprints / Paper 01
PAPER 01 - ActiveSeq

One well shows what changed, and helps explain why.

Drug screens usually give either an image of what happened to cells or a molecular readout of which biology moved. ActiveSeq captures both from the same microwell. In this pilot the combined readout classified compounds better than either view on its own, across every member of the 35-compound control panel.

Paper 01 · 18 pages · April 2026 · CC-BY 4.0
SAME-WELL MULTIMODAL CLASSIFICATION - BALANCED ACCURACY 0% 30% 60% POOLED 55.5% ACT010 60.8% ACT011 45.1% IMAGE ONLY RNA ONLY IMAGE + RNA
Hero figure - Image plus RNA outperformed either readout alone in the control-panel classification task.
TL;DR

ActiveSeq reads an image and an RNA profile from the same microwell. Across 35 control compounds, image plus RNA classified held-out wells better than RNA alone or image alone. The practical implication: a Z-Screen hit arrives with both a phenotype and a molecular fingerprint, instead of the phenotype alone.

Why it matters

In early discovery, the fastest screens often leave teams asking what a hit actually means. Imaging scales well, but morphology alone is hard to translate into biology. RNA is more explanatory, but running it deeply across large libraries is usually too expensive.

ActiveSeq pairs the two readouts in a single well. A Z-Screen hit comes with an image of the cell response and an RNA fingerprint from the same small colony of cells. That does not remove every follow-up experiment, but it makes the first look at a program remarkable rich.

What we did

We analyzed two HEK293 pilot runs covering 20,222 matched wells. A 35-compound control panel tested whether the readout could recover known compound identity, and dummy-bead wells served as a built-in stress test on the assay chemistry.

For each well we compared the image signal, the RNA signal, and the combined signal, and asked four practical questions. Does the assay flag bad wells. Are the readouts reproducible. Do image and RNA organize the panel in related ways. Does combining them actually help on held-out wells.

We also compared the results to scGeneScope, a public treatment-matched imaging-plus-RNA benchmark on 28 control compounds. The experimental designs are different enough that this is not a head-to-head comparison - sGenescope profiled ~600,000 droplets, this dataset is from ~20,000 microwells - but ActiveSeq holds up pretty well even at an order of magnitude lower sampling.

What we found

Five results that make same-well readout useful.

FINDING 01

Bad wells are visible.

Dummy-bead wells should not behave like real compound wells, and they did not. RNA yield collapsed exactly where it should have, so failed chemistry does not get to slip silently into the biology model.

5,167 → 342Median UMIs - ACT010
FINDING 02

The same compounds look like themselves again.

Splitting wells into pseudo-replicates, both RNA and image features recovered compound identity far above chance. The signal also held across the two pilot runs, which is the basic reliability bar for a screening readout.

85.7% / 100%Cross-run top-1 / top-3 - image
FINDING 03

The two readouts agree, but they are not copies.

Image and RNA organized the compound panel in related ways: enough overlap to trust the biology, enough difference that combining them adds information.

ρ = 0.718Image ↔ RNA agreement
FINDING 04

Combining image and RNA helps.

On held-out wells the combined readout beat image alone and RNA alone. The gain was distributed across the panel, not driven by a handful of easy compounds.

35 / 35Compounds improved by fusion
FINDING 05

The public benchmark puts the effect in context.

scGeneScope is larger and built differently, so this is not a head-to-head omparison. After adjusting for task difficulty, the pilot's multimodal effect sits in the same range as this public reference.

21.3x chanceBest run - vs 18.0x scGeneScope
What this enables

A better first look at every hit.

ActiveSeq changes the first question after a screen. Teams stop asking only which wells looked different and start asking what the cells looked like and which molecular programs moved with them. The pilot demonstrates that this is feasible, reproducible, that the two views add orthogonal information.

Access

Preprint, data, and analysis repo.

Public release. The preprint PDF is hosted here; the canonical workspace and per-paper analysis scripts live on Zenodo with a citable DOI.