POSH Platform: Pooled Optical Screening in Human Cells

Published in Nature Communications in December 2025, the POSH platform validated a breakthrough innovation integrating pooled CRISPR screening with high-content imaging and self-supervised machine learning. Historically, researchers faced a brutal trade-off. Traditional CRISPR screening could interrogate thousands of genes but provided limited readouts, typically just a single reporter protein—high throughput but low phenotypic depth. High-content imaging captured rich multi-dimensional phenotypes but was logistically prohibitive at genome scale—low throughput but high phenotypic depth.

POSH synthesized both approaches. The platform combines pooled CRISPR screening with Cell Painting, a multi-component fluorescent staining technique, plus Vision Transformer deep learning models trained on raw pixel images. The results speak for themselves: self-supervised models discovered biologically meaningful gene networks directly from morphology, recovering 2.5x more functional gene relationships compared to conventional analyses. The platform identified novel regulators of mTORC1 signaling, including AURKAIP1 and HSD17B10, validated through orthogonal experiments. Across 1,640 genes screened, POSH reconstructed known biological networks like the proteasome and Golgi-ER without any prior hypotheses.

Capital efficiency emerges from eliminating the historical compromise. A single integrated platform simultaneously achieves scale AND phenotypic depth.

ChemML Platform: End-to-End Small Molecule Design

Built through internal development plus the acquisition of Haystack Sciences, ChemML integrates multiple capabilities into a unified engine. Quantitative Adaptive Libraries provide hundreds of millions of binding affinity and selectivity data points. ADMET prediction models were trained on 40 years of Eli Lilly drug chemistry data through strategic collaboration. An iterative design-make-test engine becomes smarter each cycle by combining computational predictions with experimental data. The platform runs on 192 H100 GPUs providing computational power for high-end ML and physics-based simulation.

The strategic differentiation is architectural. Traditional medicinal chemistry optimizes potency first and hopes ancillary properties align. ChemML optimizes binding affinity, selectivity, metabolic stability, brain penetration, and safety profile simultaneously from cycle one. This architectural difference explains why AI-native companies achieve 80-90% Phase I success rates compared to 40-65% for traditionally discovered drugs.

TherML Platform: Modality-Agnostic Therapeutic Design

In January 2026, insitro acquired CombinAbleAI, an Israeli AI therapeutics company from AION Labs, launching the TherML platform. CombinAbleAI brought physics-informed ML models for antibody and complex biologic design, pre-trained on over 100,000 molecular dynamics simulations.

TherML’s integration created full modality coverage: small molecules through ChemML, antibodies through CombinAbleAI, oligonucleotides through insitro’s internal capabilities, plus multi-specific proteins. The capital efficiency advantage is stark. Traditional approaches require building separate discovery engines for small molecules, antibodies, and oligonucleotides. insitro’s unified platform selects therapeutic modality based on biology rather than chemistry constraints, with investment compounding across modalities.

The Partnership Model

insitro doesn’t license single candidates to pharma. Instead, it provides platform access for multi-target discovery within therapeutic areas while retaining core datasets and ML models.

The Bristol Myers Squibb collaboration, announced in 2020, focuses on ALS and frontotemporal dementia. The deal structure included $50M upfront, $20M in near-term milestones, over $2B in potential discovery, development, regulatory, and commercial milestones, plus royalties on net product sales. In October 2025, insitro and BMS extended the partnership, applying ChemML to a novel ALS target discovered in Phase 1. The total potential deal value now exceeds $2B, translating from biology discovery during 2020-2025 to molecule design during 2025-2026 within a single partnership.

Stephen Hitchcock, a scientific advisor with over 30 years in pharma R&D, explained the significance: “ChemML’s approach of simultaneously optimizing across the complete pharmacological profile represents a fundamentally different approach enabling scientists to target indications that would otherwise remain intractable, particularly CNS disorders.”

The Eli Lilly collaboration came in two waves. In September 2025, insitro trained ML models on 40 years of Lilly’s drug chemistry data. The resulting models integrated into ChemML, while Lilly developed TuneLab, a platform providing biotech access to their data. In December 2024, three strategic agreements focused on metabolic diseases, particularly MASLD. insitro secured options to in-license Lilly’s clinical-stage GalNAc delivery technology for two siRNA programs and collaborative antibody therapeutic development for a third metabolic target. insitro retains full global rights while Lilly receives milestones and royalties.

The inaugural partnership with Gilead Sciences in April 2019 validated the platform approach. Focused on NASH and MASLD, the deal included $15M upfront, $35M in near-term milestones, and $200M potential per program across up to five programs. Gilead provided clinical trial data while insitro applied its platform for target identification.

The partnership structure consistently follows the same logic: pharma funds research, insitro retains platform ownership and core datasets, and both parties share downstream commercial upside for specific targets within defined therapeutic areas.