The Pull Requests Behind the $21B Drug Gamble

On the GitHub repo for GT4SD/molecular-design, a contributor recently merged a new pipeline—a seven-step workflow that runs state-of-the-art deep learning models for target-based molecular design. It’s a clean, well-documented piece of work, the kind of commit that signals a mature open-source ecosystem. What the commit message didn’t say: its author moonlights from a well-funded AI biotech startup where the real models are locked behind NDAs and partner contracts. And that gap—between the public scaffolding and the proprietary arms race—defines the current moment in AI-driven drug discovery.

We’re in the middle of a funding avalanche. In May 2026, Google DeepMind spinout Isomorphic Labs closed a $2.1 billion Series B, the largest AI pharma financing ever. Two months later, Chai Discovery raised $400 million at a $3.8 billion valuation—just seven months after hitting unicorn status. Even before incorporating, ex-OpenAI researcher Miles Wang was reportedly negotiating a $200 million round at a $2 billion valuation for a drug-repurposing startup nobody had ever heard of. And Xaira Therapeutics, which emerged from stealth in 2024 with a billion dollars, has hoovered up another $300 million and finally started talking to pharma partners. Total investment in generative AI drug discovery now comfortably exceeds $20 billion.

Open-source developers watch this from the sidelines, sometimes cheering, sometimes skeptical. Hacker News commenters have been pointing out that pharma R&D spending dwarfs AI investment by orders of magnitude—and that “AI works on the cheap part.” It’s a recurring theme: the models are flashy, but the clinic doesn’t care about flash.

The Tech That’s Getting Funded

Chai Discovery’s trajectory is a case study in compressed timelines. Founded in 2024, the company jumped from a $30 million seed to a $130 million Series B in under two years, then piled on another $400 million in July 2026. Its secret sauce? A series of models—Chai-1, Chai-2, Chai-3—that predict molecular interactions with increasing precision. Chai-2 already achieved a 16% hit rate in fully de novo antibody design, enough to skip high-throughput screening in many campaigns. Chai-3 claims a 35–40% hit rate. Those numbers are jaw-dropping compared with traditional lab methods, and pharma is paying attention: Pfizer signed a license deal giving it priority access and a custom model trained on its own data; Eli Lilly and Novartis are also partners.

Isomorphic Labs takes a different route. It’s built IsoDDE, a drug design engine that can find cryptic binding pockets from amino acid sequence alone—pockets that sometimes took scientists a decade to discover. In benchmarks, it more than doubled AlphaFold 3’s accuracy on protein-ligand prediction. But IsoDDE is black-box commercial software; unlike AlphaFold, you can’t peek at the code. It’s available only through the company’s partnerships, which already span Eli Lilly (potential $1.7 billion in milestones), Johnson & Johnson, and Novartis.

Xaira’s play is the virtual cell. In March 2026 it unveiled X-Cell, a 4.9-billion-parameter diffusion language model trained on 25.6 million perturbed single-cell transcriptomes. Diffusion—more commonly associated with image generators like Stable Diffusion—is an odd choice for biology, but Xaira claims it outperforms existing models by up to five-fold on perturbation prediction, and it generalizes zero-shot to unseen cell types. The company is now hunting for partners to provide more pre-clinical data; as SVP Rachel Lane told BioSpace, “That’s where we see a sweet spot for partnerships, to help further build the data moat.”

The Talent Exodus

The backstory to these companies is a slow-motion brain drain from big AI labs. Isomorphic was spun out by DeepMind’s Demis Hassabis; Chai’s co-founder Joshua Meier previously worked at OpenAI. Now Miles Wang, another OpenAI researcher, is leaving with a cohort of colleagues to build yet another startup. Lightspeed is reportedly in talks to lead the round. Wang’s company will focus on drug repurposing—finding new uses for existing molecules, which can shave years off development because safety is already established. But the $2 billion pre-revenue valuation drew snickers on HN: “$2B for a pitch deck and an OpenAI alumni T-shirt.”

It’s easy to mock, but the pattern is real. OpenAI and DeepMind are bleeding talent into biotech, and VCs are betting that AI-native founders can crack problems that baffled traditional pharma. Whether the valuation premium for an “OpenAI pedigree” is justified remains an open question.

The Data Moat Wars

Beneath the fundraising headlines, a quieter battle is being waged over data. All these models consume enormous datasets, and the quality of those datasets determines what they can do. Training data bias is endemic: known drug-target interactions skew toward a handful of well-studied proteins; most potential targets lack high-quality annotations. As a 36Kr analysis noted, “No matter how good the model, you’re feeding it the same incomplete textbook.”

That’s why Xaira’s Lane talks so bluntly about building a data moat. Isomorphic’s advantage is access to DeepMind’s research infrastructure and proprietary protein data. Chai’s partnership with Pfizer includes training a model on Pfizer’s internal data—a walled garden no competitor can access. The open-source community, meanwhile, relies on public databases like ChEMBL and the occasional leaked affinity assay. It’s an asymmetric war, and the proprietary players are winning.

Why the Clinic Isn’t Impressed

Despite all this, not a single AI-discovered drug has been approved. Roughly 173 AI-derived programs are in clinical development, but most are still in early stages. Phase I trials often show 80–90% pass rates for AI-designed molecules, but Phase II success drops to ~40%—no better than traditional methods. The poster children for failure include Exscientia’s cancer drug EXS-21546 (halted after Phase I/II modeling suggested it couldn’t reach therapeutic levels) and BenevolentAI’s atopic dermatitis candidate (failed to beat placebo). In 2023 alone, at least six clinical-stage AI drug pipelines were quietly discontinued.

Michael A. Santoro, a professor at Santa Clara University, put it bluntly in ProMarket: “The roughly $2.6 billion it takes to bring a drug to market is dominated not by discovery but by clinical trials and the failures within them, and there AI has so far made almost no difference.” AI can design a molecule faster and cheaper, yes. But that was never the bottleneck. The expensive, risky part is proving the molecule works in humans.

FDA and EMA are watching. In January 2026 they jointly released a set of “Good AI Practice” principles for drug development, covering everything from ethics to lifecycle management. It’s a sign the regulators are taking the technology seriously—and that they’ll demand rigorous validation before anything gets approved.

What the Developers Are Building

While the unicorns duke it out, the open-source community keeps shipping. Repos like Awesome-LLM-Scientific-Discovery track frontier systems from Google DeepMind, OpenAI, and Meta FAIR. The DGDM framework (Darwin-Gödel Drug Discovery Machine) offers a self-improving AI system for small-molecule optimization. StackFeat-RL uses reinforcement learning for biomarker discovery. These projects are small, often maintained by a handful of contributors, and their immediate impact on pharma pipelines is zero. But they represent a collective bet that the field’s foundations should be public.

There’s a pragmatic side, too. Many contributors work at AI biotech firms and use open-source as a way to test ideas, recruit talent, or simply stay connected to the community. A commit merged on a Tuesday might come from someone whose day job is negotiating an eight-figure pharma partnership.

2026: The First Real Test

This year is when the rubber meets the road. Insilico Medicine launched the first Phase III trial for an AI-designed drug—rentosertib for idiopathic pulmonary fibrosis. Generate Biomedicines has two Phase III trials for its AI-designed antibody, GB-0895. Roche is reading out data on five AI-influenced molecules. If even one of these succeeds, the narrative flips overnight. If they all fail, the “$20 billion wasted” headlines will write themselves.

For now, the money keeps flowing, the models keep getting better, and the GitHub repos keep growing. But the ultimate judge is a long way from a pull request. As one HN user commented: “It’s cool that we can design a protein in a weekend now. Call me back when it survives a Phase III run.”