Contrastive learning compares brain recordings from mice and patients according to their biological state. When projected into this space, the effects of treatments in animals correspond to their known clinical efficacy.
Animal models do not predict the effectiveness of drugs in humans, especially in neurology. The electrical activity of the brain, measurable in both mice and patients, is a way to bridge this gap, but its richness complicates the task: it is necessary to separate the disease from the species, the recording technique or the protocol.
A common area organized by state of health
The authors designed a double-rule contrasting learning that brings the corresponding states in mice and humans together while keeping different phenotypes at bay. The model finds preserved sensory responses between species and, in epilepsy, separately links three murine models to heterogeneous patient groups.
The most telling result is retrospective: by projecting treated mice in this frozen representation, the shift to the "healthy" state aligned with humans corresponds to the known clinical efficacy of ten model-drug combinations, including a disease-specific deleterious effect. The model also compares the neuronal dynamics of Fmr1-knockout mice (a model of fragile X syndrome) and human carriers with a variation in 16p11.2 copies, despite different genetic causes and recording patterns.
Read the full article on the original source
Go to the source →