Drug Discovery

Maximum information per experiment, minimum experiments per conclusion. Halmos Labs builds automated biotech R&D on the principles of validity and interpretabillity. We run research, and simulations answering open questions ranging from wetlab experiment design to full cancer vaccine manufacturing.

Extrapolative AI to unlock First-in-Class drugs. We’re an AI-native biotech, working on replacing needles with pills. Many drugs, like insulin or Ozempic, are wonderful except you have to inject them. We discover new molecules that you can take as a pill instead. Because the pills operate on biology which has been shown to be effective in people, rather than merely in animal models of disease, we avoid what has historically been the single biggest risk in drug discovery. This should dramatically increase our probability of success. There’s $200B of treatable diseases and validated human biology without a pill because discovery programs lack a tractable chemical starting point. This is a zero-shot inference problem, because these targets have no training data, so LLMs and other historical interpolative AI approaches are insufficient. Instead, we are building our own extrapolative frontier

The Intelligence layer for Drug Manufacturing. Abinitio Bio builds foundation models for biomanufacturing, validated against wet lab data. Our first model, Echo, predicts manufacturing outcomes more accurately than standard models, and any pharma or CDMO can adapt it to their own workflow instantly. By making process decisions that take 6-18 months today in hours, we help teams avoid the cost of delay on a single program.

The AI-native platform for next-generation protein characterization.. 10x Science develops frontier AI models with deep memory to redefine how scientists understand and engineer biology across the life sciences, starting with drug development. AI-powered drug discovery is flooding the pipeline with new candidates faster than ever. But before any of them can advance to the clinic, scientists must deeply characterize the protein therapeutic to understand whether it will work. Today, that process takes months of manual, error-prone analysis that cannot keep pace with discovery. Our AI-native platform automates it, giving drug developers the confidence to advance the right candidates and stop the wrong ones early. We have strong commercial traction with enterprise pharma customers. Our founding team built the field of next-generation protein characterization out of Carolyn Bertozzi's Nobel l

AI and Data for Cancer Therapeutics. Creating the largest patient-derived cancer single-cell dataset for licensing to pharmaceutical and AI companies. These datasets are unique as they screen the effects of various drugs and genetic medicines under the control of our DNA switches.

Evolutionary intelligence for autoimmune disease. Ditto Bio discovers drugs for autoimmune diseases by harnessing parasite biology. Viruses, ticks, and worms have experimented over millions of years to create proteins that control the human immune system. We use AI to identify these proteins and engineer them into next-generation autoimmune therapies. Our team is three PhD scientists from UCSF, UCSD, and UC Berkeley who have worked together for 3+ years in biotech. We have deep expertise in parasite biology, computational biology, and software for AI-based protein structure prediction

The agentic drug company. We simulate human biology.. CellType is building an agentic drug company. AI agents that run the full drug discovery pipeline on top of biological foundation models that simulate human biology. Our core technology, developed with Google DeepMind, has already discovered and validated a new cancer treatment signal. We’re working with Top 10 pharma.

Multimodal foundation models to predict uncollected patient biology. Strand AI develops foundation models to generate missing bio-data about patients. With this imputed data, pharmaceutical companies can select better patients for their drug trials and shave months from their drug launch timelines. We’ve already trained a multimodal foundation model that integrates spatial biology modalities, beating SOTA at a fraction of the cost.