Daphne Koller has reached the point at which the strongest argument for insitro can no longer be the elegance of its platform. The company has assembled human genetics, imaging, laboratory automation and machine learning into a system designed to find better therapeutic targets and design molecules against them. In 2026, it is showing more of what that system has produced: additional targets in amyotrophic lateral sclerosis, a preclinical obesity programme and broader capabilities spanning small molecules, oligonucleotides and biologics.
The progression matters because artificial intelligence in drug discovery has moved beyond novelty. Pharmaceutical companies have signed numerous partnerships, investors have funded dozens of platforms and foundation models have made computational biology more accessible. The question has shifted from whether algorithms can generate plausible hypotheses to whether they can improve the probability, speed or economics of delivering a medicine.
Koller, insitro’s founder and chief executive, is trying to answer through a portfolio rather than a single bet. Her leadership challenge is to preserve the advantages of a technology platform while accepting the unforgiving cadence of experimental biology. Models can rank targets quickly. Cells, animals, toxicology studies and patients decide whether those rankings were useful.
Partnership signals become programmes
In March 2026, insitro and Bristol Myers Squibb expanded their work in ALS by nominating two additional targets, known publicly as ALS-2 and ALS-3. They joined an earlier target, ALS-1, identified through a collaboration that uses patient-derived data and insitro’s Virtual Human approach. The new nominations triggered a $10 million milestone payment. Insitro says the underlying work concentrates on the mislocalisation of the TDP-43 protein, a feature observed in the great majority of ALS cases.
One milestone does not validate a drug. It does show that a large pharmaceutical partner was prepared to move more outputs through a defined discovery process. That distinction is important. Platform companies often publicise the number of predictions they can make; partners pay for decisions they consider valuable enough to pursue. Bristol Myers Squibb had already extended the broader collaboration in late 2025 with up to $20 million of additional research funding for one year. Insitro says the relationship carries potential aggregate milestone payments exceeding $2 billion, but those contingent figures should not be mistaken for booked revenue.
The model gives Koller two forms of leverage. Partnerships bring non-dilutive funding and expose the platform to experienced drug-development teams. Internal programmes allow insitro to retain more economics and determine whether it can become a product company in its own right. The tension is familiar in biotechnology: too much service work can turn a differentiated platform into a contractor, while too many wholly owned programmes can consume cash before the science is sufficiently mature.
Insitro reported in 2026 that it had raised roughly $800 million in capital and generated about $150 million in partnership revenue. Those are company-supplied totals, not a substitute for public financial accounts, but they indicate the scale of the resources already committed. Capital that once funded data collection and platform construction must increasingly support candidate selection, manufacturing, regulatory work and clinical development. The unit of progress changes from another model release to a programme that survives successive biological filters.
Human data can narrow the search, not remove uncertainty
A February 2026 obesity study illustrates both the promise and the gap. Insitro analysed 69,598 magnetic-resonance images to derive a phenotype for brown adipose tissue, then combined that signal with human genetic evidence to identify targets. In mice with obesity, knocking down one target reduced body weight by 15 per cent and fat mass by 25 per cent while preserving lean mass, according to the company.
The use of human imaging at that scale is strategically interesting. Brown fat has long attracted attention because it consumes energy, but measuring it reliably across large populations is difficult. Computer vision can turn images collected for other purposes into quantitative traits, creating a bridge between genetic variation and a therapeutic mechanism. If the phenotype is robust, it may reveal biology that conventional diagnosis codes or laboratory measurements miss.
Yet the result remains preclinical. Gene knockdown in a mouse is not the same as a safe, manufacturable treatment in people. The effect must replicate, the target’s role must be understood across tissues and a therapeutic modality must achieve suitable exposure without unacceptable consequences. Obesity is also an increasingly competitive market in which established incretin medicines have raised the efficacy standard. A differentiated mechanism may need to show preservation of lean mass, easier dosing, better tolerability or benefit in a defined patient group rather than weight loss alone.
Koller’s advantage is not that insitro can escape those tests. It is that the company can use human-derived signals to choose where to spend scarce experimental and clinical capital. That is a more defensible claim than promising to automate drug discovery. Even a material improvement in target selection would be economically significant: most programmes fail, late-stage failures are expensive and better evidence at the beginning can prevent years of misallocated work.
A wider design stack increases both reach and integration risk
Insitro broadened its remit in January 2026 by announcing an agreement to acquire CombinAbleAI and launch the TherML platform. The proposed transaction was intended to add computational design across antibodies and other complex biologics to insitro’s existing small-molecule and oligonucleotide capabilities. Insitro said the CombinAbleAI team would continue as an Israeli research-and-development centre, alongside colleagues in the United States, Poland and Malaysia.
The rationale is coherent. A target can be compelling yet unsuitable for a small molecule. Access to several therapeutic modalities allows scientists to match the intervention to the biology instead of forcing every discovery through one design system. It also creates a more complete loop: identify disease mechanisms from human data, perturb them in engineered cells, select a target, design a therapeutic candidate and feed experimental results back into the models.
Completeness can become complexity. Different modalities require different training data, assays, manufacturing processes and development expertise. Integrating an acquisition across scientific cultures and time zones is difficult even when the software works. Koller must decide which capabilities should become shared infrastructure and which should remain specialised. If every programme receives a bespoke stack, the platform loses economies of scale; if tools are standardised too aggressively, teams may miss biological nuance.
Insitro’s publication of its POSH functional-genomics work offers one indication of how the company approaches external validation. The method combines pooled CRISPR perturbations, cell imaging and self-supervised learning to infer relationships among genes. A peer-reviewed study reported that it identified substantially more functional relationships than conventional analysis, and the company released models and code. Open publication can strengthen recruitment and scientific credibility, but competitive advantage will depend on proprietary data, experimental execution and accumulated learning rather than algorithms alone.
The platform must demonstrate an economic advantage
Large pharmaceutical companies have their own data, medicinal chemists and growing AI teams. Cloud providers and specialist software firms are making powerful models broadly available. That environment weakens any moat based purely on access to machine learning. Insitro needs to show that its integrated datasets and laboratory systems generate decisions competitors cannot easily reproduce, and that those decisions improve development economics rather than merely add another analytical layer.
The Lilly partnership announced in 2025 points to a practical use case. The companies are building machine-learning models for absorption, distribution, metabolism, excretion and toxicity. Better prediction in those areas could help chemists reject weak molecules earlier. This is less dramatic than claiming an AI-designed cure, but it addresses a recurring source of attrition. The value can be measured through fewer synthesis cycles, faster lead optimisation or reduced late-stage toxicity.
For Asian life-sciences markets, insitro’s approach carries a related opportunity and warning. The region has large patient populations, expanding genomic resources and sophisticated research centres, but health data are fragmented across institutions and regulatory systems. Models trained on narrow populations can perform poorly when disease prevalence, genetics or clinical practice differs. A presence in Malaysia offers operational reach, yet the more important task is building collaborations that include representative Asian data under credible consent and governance. Scale without diversity would reproduce old drug-development blind spots in digital form.
Koller’s career spans academic machine learning, online education and biotechnology, but the relevant leadership skill now is portfolio judgement. She must allocate capital between platform improvement and programmes, between partner obligations and owned assets, and between acquiring capabilities and integrating them. Scientific ambition is plentiful; sequencing is the scarce resource.
The most useful disclosure would therefore connect platform activity to programme economics: how often a computational nomination survives experimental validation, how much time is removed from design cycles and where candidates fail. Not every figure can be public in a competitive portfolio, but a consistent set of operating measures would help partners and investors distinguish repeatable learning from isolated successes.
Insitro enters the second half of 2026 with more evidence that its engine can produce targets and design options. It still lacks the evidence that matters most: clinical outcomes attributable to that engine and a transparent demonstration that they arrived with better odds or lower cost. The expanded ALS work, obesity findings and multimodal design stack move the company closer to that test. They also raise the burden of proof. Koller’s achievement will not be judged by how comprehensively insitro maps biology, but by whether it can use that map to choose a few routes that reach patients.
