Our understanding of biology has always been shaped by our instruments. From the microscope to the microarray, new scientific tools have allowed us to see life at increasing resolution – giving us more granular definitions of disease and more targeted drugs for treating them. Nowhere is that more true than in cancer, which has evolved from a singular disease with a local phenotype to a constellation of conditions defined by tissue, cell, gene and increasingly by mechanism and pathway.

This molecular view of biology has led to discoveries like the PDL-1 and blockbuster drugs like Keytruda and Opdivo that target them. These first generation immunotherapies have shown remarkable results - sending previously untreatable cancers into remission - but only for a small subset of patients. Nearly a decade after their approval, attempts to reproduce their success around other pathways have mostly failed – creating an urgent need for new approaches to immunotherapy development.

This is the premise of our latest investment in NOETIK - a new AI-native biotech using self-supervision and spatial biology to learn new cancer subtypes from human data and develop the next generation of cancer immunotherapies, working backwards from the patient.

NOETIK was founded by Ron Alfa and Jacob Rinaldi on the understanding that biology is complex and the belief that a “reductionist” approach to drug discovery is at odds with everything we know about the disease. That cancer biology is dynamic, evolutionary and governed by complex interactions across far flung networks that can only be observed at the system level.

Humans struggle to understand complex systems - especially at biological scale - but recent advances in self-supervised learning suggest that this is where modern AI systems excel. Large pre-trained models have shown an impressive ability to encode complex biology and predict dynamic systems - from protein folding, to gene regulation - that have eluded scientists for decades.

The power of large, self-supervised systems - or foundation models - comes from the rich internal representations they build during pre-training; billions or trillions of connections, represented as vectors in high-dimensional latent space or embeddings. Unlike supervised learning, which is task-specific and relies on human labels for ground truth, self-supervision allows models to learn directly from data and build general representations that are untethered to existing classifications and therefore not bound by current knowledge.

While much of the attention on foundation models in biology has been on “generative” tasks like protein design and sequence generation, the internal representations themselves may provide a new lens through which to learn novel biology – if we can find ways to understand what these large models are learning.

NOETIK is the culmination of ideas that Ron and Jacob have been working on for over a decade. First at Stanford where they met as PhD students and then at Genentech where Jacob developed new ML methods for personalized cancer vaccines and at Recursion where Ron led R&D and recruited Jacob to help build the oncology team at what was then a rare disease focused company. It was at Recursion, where they saw the potential of large-scale data generation for AI and the opportunity opened up by advances in self-supervision and multi-modal learning.

The company was founded with a simple mission, to improve the lives of cancer patients through breakthrough medicines. To do that, they are leveraging advances in machine learning, spatial biology and high-throughput in vivo experiments to learn from human tumor biology and “reverse translate” those insights into new targeted cancer immunotherapies.

It is an ambitious vision, which upends many of the assumptions of basic research and inverts the traditional drug development process. It requires new techniques for generating data, training large-scale models and extracting insights from machine-learned representations. It requires new organizational design and new ways of thinking about target discovery, clinical development and translation, supported by new tools for generating and testing hypotheses at scale. If successful it could offer a new path forward for drug development and have an outsized impact on the lives of millions of cancer patients and their loved ones.

To do this Ron and Jacob have assembled an impressive, interdisciplinary team from organizations like The Parker Institute, Genentech, ImmunAI and GRAIL and are leveraging research and researchers from the leading labs of Brian Brown Lab at Mt. Sinai, Dana Pe’er at Sloan Kettering and Dan Yamins at Stanford.

In just over a year they have generated over 200 Terabytes of rich multi-modal data from over 1,000 patient tumors, phenotyping over 22 million spatially resolved cells which they have used to train some of the first self-supervised models of cancer biology. And they are just getting started.

Science is slow and drug development takes time but the NOETIK team is moving incredibly fast and with extraordinary focus and purpose. We are delighted to be working with them as early investors alongside DCVC, 11.2 Capital, Catalio, EPIC Ventures, InterMountain Health and many more. If you are interested in being part of this ambitious vision — they are hiring!