The company has raised roughly $800 million to build a single machine-learning engine that designs small molecules, antibodies, oligonucleotides, and complex biologics. In January 2026, Insitro launched TherML, a platform that unifies four drug modalities under one AI architecture, trained on data its own automated labs generate at scale.
The CombinAbleAI acquisition adds physics-informed, AI-driven design for complex biologic therapeutics, completing a full-stack system spanning every major drug modality.
Early validation from Eli Lilly and Bristol Myers Squibb suggests large pharma sees the platform as infrastructure, not just another point-solution tool.
Bringing a single drug to market now exceeds $2 billion. Nine of ten clinical candidates fail, most during Phase 2, after the expensive part has started. The industry has tried to make R&D cheaper through AI, but most AI-native drug discovery companies operate inside one modality: small molecules (Atomwise, Exscientia) or antibodies (Absci, BigHat).
Integration across modalities is the hard part. Each class has its own rules for binding, stability, delivery, and manufacturability. Building a platform that handles all four requires not just separate models for each, but a unified training pipeline that can transfer insights between them.
That is what TherML attempts. As we wrote in June, the convergence of AI and biology is reshaping drug discovery. Few companies are trying to build the infrastructure layer rather than a single therapy. ai-drug-discovery-platforms-2026
Insitro capital raised to date
Includes $150M from non-dilutive pharma partnerships. Latest valuation: $7B. · Insitro press releases, 2026
The TherML Architecture
TherML is not a single model. It is a stack of specialized engines connected by a shared training infrastructure, each responsible for one modality but feeding into a common representation space. The small-molecule engine uses Insitro's Quantitative Adaptive Libraries, algorithmically designed chemical libraries that map local regions of chemical space and generate training data on demand. The oligonucleotide engine industrializes siRNA candidate design across targets using AI and automation. The biologics engine, inherited from CombinAbleAI, applies physics-informed optimization pre-trained on more than 100,000 molecular dynamics surrogates to predict protein structure and flexibility. The antibody engine designs multi-specifics and T-cell engagers.
The key architectural choice: potency and developability are optimized simultaneously, not sequentially. Traditional drug design optimizes for binding affinity first, then checks whether the molecule can be manufactured. TherML generates designs that are already manufacturable. The platform estimates stability, solubility, and synthesis complexity alongside target engagement predictions.
"Drug discovery has traditionally optimized molecules for potency before assessing developability — often discovering that highly potent candidates face manufacturing constraints. By integrating CombinAbleAI's physics-informed, AI-driven design with our causal biology platform, we treat potency and manufacturability as interdependent design criteria from the outset."— Philip Tagari, Chief Scientific Officer, Insitro
How TherML differs from other AI drug discovery platforms
The closest competitor is Isomorphic Labs, but DeepMind's spinout focuses on AlphaFold-derived protein structure prediction, not end-to-end drug design. Recursion operates at scale on the phenotypic imaging side but outsources chemistry. Insitro's bet is that owning the entire stack (data generation, model training, and modality design) produces better candidates than any single layer alone.
Pharma Validation
The thesis is not theoretical. Eli Lilly signed three separate agreements in 2024-2025, including a September 2025 collaboration to build first-in-kind ML models predicting key pharmacological properties of small molecules. Bristol Myers Squibb extended its collaboration in 2025 after an earlier option exercise in 2022, leveraging the ChemML platform to discover molecules for ALS, a target where traditional approaches have produced few viable candidates.
The partnership structure is unusual. The company retains full global rights to its research programs; the pharma partner receives milestone payments and royalties on approved products. It is positioning itself as a platform owner rather than a contract research organization. The $150 million in non-dilutive partnership revenue already on the books supports that framing.
Market Position
The AI drug discovery market has fragmented into roughly 200 startups globally. Most will fail because drug discovery's failure rate is structural, not because their models are wrong. TherML addresses the most common cause of clinical failure: developability issues that were predictable before trials began. If the platform reduces Phase 2 attrition by even 10 percentage points, the economic impact across a partnered pipeline would be measured in billions.
The risk is execution. The CEO (Daphne Koller, Stanford professor, Coursera co-founder, MacArthur fellow) has the credibility to attract talent and capital, but the company has not yet advanced a wholly-owned candidate to clinical trials. All pipeline assets are at discovery or preclinical stage. TherML is broad in scope. The question is whether it produces molecules that survive Phase 3.
What happens to AI drug discovery by 2028?
Probability: 65%. Three platform companies (Insitro, Recursion, Isomorphic Labs) will have generated at least one Phase 2 readout each by end of 2027, creating a feedback loop between real clinical data and model training.
✅ Arguments for
The multi-modality bet is defensible: if the platform works across drug classes, it captures more of the pharma value chain than any single-modality competitor.
Confirmation criteria: The company or a peer platform-company advances a wholly-owned candidate to Phase 1 within 12 months.
❌ Arguments against
Multi-modality increases scope, but also complexity. A platform that does everything well may turn out to do nothing exceptionally, especially against narrow specialists who optimize for one chemistry class.
Disconfirmation criteria: No Insitro candidate enters the clinic by end of 2027, or an early-phase trial fails on safety. This is the most common developability failure TherML is designed to prevent.
Key signals to track
IND filing for a wholly-owned candidate (pipeline progress)
Partner option exercise on a TherML-designed molecule (external validation)
TherML-designed molecule enters Phase 2 without major safety revision (developability claim test)
A competitor launches a comparable multi-modality platform (competitive signal)
Development scenarios
🟢 Optimistic scenario (25%)
Implications: Platform-based drug discovery becomes the default model for pharma R&D partnerships, and the company is positioned as the infrastructure layer.
🟡 Base-case scenario (50%)
Implications: Insitro becomes the preferred AI partner for pharma, profitable on partnership revenue but years away from an approved drug.
🔴 Pessimistic scenario (25%)
Implications: TherML's multi-modality scope is seen as a liability. Too broad to excel at any single modality. Capital shifts toward focused platforms with clearer paths to market.
