Искусственный интеллект Google репрессирует холодные опухоли: как новая модель меняет подход к терапии рака Google’s Artificial Intelligence Reshapes Cancer Therapy by Targeting Cold Tumors: A New Model Revolutionizes Treatment Approaches

Google, in collaboration with Yale University, has introduced a new foundational model consisting of 27 billion parameters, crafted specifically to comprehend the «language» of individual cells.

C2S-Scale 27B generated hypotheses regarding the behavior of cancer cells, which were later experimentally validated on live organic samples.

«This discovery has opened a promising avenue for developing new cancer therapy strategies,» emphasized the company.

The model is built upon earlier research that demonstrated that biological and linguistic systems adhere to similar scaling laws—increasing size correlates with enhanced effectiveness.

A significant challenge in cancer immunotherapy is the presence of many «cold» tumors that go unnoticed by the immune system. One way to «heat» them up is by facilitating the presentation of signals through a process called «antigen presentation.»

Google tasked C2S-Scale 27B with identifying a drug that acts as a conditional booster: amplifying immune response specifically within a certain «immune-positive» environment, where there is a low level of interferon that is insufficient to independently activate the antigen presentation process.

This task required conditional reasoning, a capability that smaller models struggled to manage.

To achieve this goal, the researchers developed a virtual two-context screening capable of revealing this synergistic effect. It involved two stages:

Next, Google simulated over 4,000 drugs across both contexts and directed the model to determine which ones enhanced antigen presentation only in the first scenario. This allowed for a focused search on clinically relevant scenarios.

Among the numerous candidates, 10-30% were already discussed in scientific literature, while others turned out to be unexpected discoveries.

The model indicated a «striking contextual gap» for a kinase inhibitor named silmitasertib (CX-4945). The neural network predicted a significant enhancement in antigen presentation when using the drug in an «immune-positive» context, but almost no effect in an «immune-neutral» setting.

Notably, this concept is entirely novel and had not been previously mentioned anywhere.

In the following stage, researchers tested the hypothesis in the lab, using human neuroendocrine cells—samples that the model had not «seen» during its training. Results demonstrated that:

In laboratory experiments, this combination led to approximately a 50% increase in antigen presentation, making the tumor more visible to the immune system.

The digital prediction was confirmed multiple times.

C2S-Scale discovered a new conditional interferon booster that could aid in transforming «cold» tumors into «hot» ones—making them more susceptible to immunotherapy.

«Although this is merely the first step, it already provides an experimentally validated foundation for developing new combination therapies, where multiple drugs act together for a stronger effect,» the blog states.

Yale University teams are already investigating the identified mechanism and testing other AI predictions across various immune contexts. With further preclinical and clinical validation, such hypotheses could expedite the development of new treatment methods.

Previously, the biotechnology firm SpotitEarly began developing a home cancer test based on analyzing human breath. This technology integrates the olfaction of dogs and artificial intelligence algorithms.

It’s worth noting that in September, researchers developed an AI tool for predicting over 1,000 diseases and forecasting health changes up to 10 years in advance.