insitro Validates AI-Enabled POSH Platform in Nature Communications, Bridging Critical Gap in Drug Discovery

New publication demonstrates that insitro’s POSH platform — integrating pooled CRISPR screening and self-supervised deep learning — successfully breaks the historic compromise between scale and depth to more rapidly identify novel therapeutic targets

SOUTH SAN FRANCISCO, Calif.–(BUSINESS WIRE)–insitro, the AI therapeutics company built on causal biology, today announced the publication of research in Nature Communications validating its POSH (Pooled Optical Screening in Human cells) platform. The study details the development of CellPaint-POSH, a scalable approach that integrates pooled CRISPR screening, high-content imaging, and advanced machine learning to map gene function at scale without relying on predefined biomarkers or human-engineered hypotheses.

Across multiple large-scale experiments, the study demonstrates that biologically meaningful gene networks emerge directly from morphology-based readouts and AI-derived embeddings. By capturing the subtle, holistic changes in cells rather than isolated metrics, the platform reveals cellular relationships and therapeutic targets that conventional, hypothesis-driven analyses systematically miss.

“Biology doesn’t organize itself according to the features we’ve learned to measure,” said Daphne Koller, Ph.D., founder and CEO of insitro. “insitro’s POSH platform has demonstrated that self-supervised models trained on unbiased cellular morphology can reconstruct gene function and causal relationships without being told what to look for. We’re finally able to interrogate the genome at scale while preserving the phenotypic complexity where disease mechanisms actually live.”

Drug discovery has historically been constrained by a fundamental compromise: researchers were forced to choose between screening thousands of genes using low-resolution metrics (scale) or interrogating a handful of genes with deep biological insight (depth). Prioritizing scale typically meant relying on coarse, low-dimensional readouts — such as simple cell viability or single-protein expression — that inherently obscure complex biology. These “looking under the streetlight” approaches often miss the subtle, multi-factorial changes that drive disease. Conversely, while high-content phenotypes offer rich insights, generating and interpreting them at scale has been logistically prohibitive, making it impractical to map the functional architecture of the entire human genome.

insitro’s approach resolves this trade-off by industrializing a method that is simultaneously broad and deep. By synergizing the massive throughput of pooled CRISPR screens with the high-dimensional, multiplexed data of Cell Painting, CellPaint-POSH identifies shared cellular states, pathways, and functional groupings de novo. This capability allows the company to profile the effects of genetic perturbations across thousands of phenotypic features while capturing intricate details of organelle shape, texture, and organization. The result is a high-resolution map of disease biology that moves beyond binary “hit-calling” to capture a holistic phenotypic signature for every gene screened.

Interpreting Cellular States with Self-Supervised AI

The study highlights the power of self-supervised machine learning — specifically Vision Transformers — to interpret the “visual language” of cell biology without human labels. Unlike supervised models constrained by human-defined features (such as “nucleus size” or “mitochondrial shape”), these self-supervised models analyze raw pixel data to learn intrinsic patterns. They map each cell into a high-dimensional feature space that preserves subtle biological relationships, allowing the system to group genes with similar functions based on patterns invisible to the human eye.

These models recovered 2.5 times more functional gene relationships than conventional, expert-designed analysis, detecting subtle shifts in cellular texture and organization that signal therapeutic opportunities.

“High-content phenotyping has always promised to capture the full complexity of cellular biology, but we’ve lacked the tools to extract meaning from that complexity at scale,” said Philip Tagari, Chief Scientific Officer at insitro. “We now have the tools to identify therapeutic targets based on comprehensive cellular signatures rather than predefined readouts. This fundamentally changes how we approach target identification.”

“With insitro’s POSH platform, we no longer have to choose between the scale of a screen and the richness of the data,” said Max Salick, Ph.D., Director of Integrated Technology Exploration at insitro and a founding member of the POSH research team at insitro. “Cellular life is intricate, and with these tools, we are transforming it into a quantitative science. Seeing this platform mature from an ambitious concept into a proven engine that uncovers novel disease mechanisms affirms that the most promising discoveries come when you let unbiased data lead the way.”

About the study

“A pooled Cell Painting CRISPR screening platform enables de novo inference of gene function” appears in the Dec. 16, 2025 issue of Nature Communications. Technical highlights:

• Mitoprobe chemistry: To enable compatibility between Cell Painting and in situ sequencing, researchers engineered a structural biology-inspired RNA probe for mitochondria. The stain avoids issues associated with standard dyes and supports robust high-content imaging at scale.
• Self-supervised learning outperforms expert-engineered features: The study benchmarks a self-supervised Vision Transformer approach against classical image analysis, showing improved extraction of biologically meaningful features without human labels. This was validated in a screen of 1,640 genes, in which the platform reconstructed known biological networks (e.g., proteasome, Golgi–ER) without prior hypotheses.
• Target discovery at scale: The work resulted in the identification of AURKAIP1 and HSD17B10 as novel regulators of mTORC1 signaling, validated via orthogonal experiments.
• Resources shared to advance the field: The team also released pre-trained models and analysis code on Hugging Face and GitHub under an open license.

Authors:

Srinivasan Sivanandan, Bobby Leitmann, Eric Lubeck, Mohammad Muneeb Sultan, Panagiotis Stanitsas, Navpreet Ranu, Alexis Ewer, Jordan E. Mancuso, Zachary F. Phillips, Albert Kim, John W. Bisognano, John Cesarek, Fiorella Ruggiu, David Feldman, Daphne Koller, Eilon Sharon, Ajamete Kaykas, Max R. Salick, Ci Chu

About insitro

insitro is the AI therapeutics company built on causal biology. By generating an integrated, multimodal corpus of human and cellular data and analyzing it with machine learning, insitro’s platform aims to reveal how disease begins, progresses, and can be resolved. The company applies this approach to identify genetic drivers, prioritize targets, and design medicines intended to treat disease at its root, with programs focused in metabolic disease and neuroscience. insitro is backed by world-class investors and has raised approximately $800M in capital, including approximately $150M from non-dilutive pharma partnerships.

Contacts

Media contact
Eric McKeeby
eric@insitro.com