Bio Foundation Models Predict Therapy Responders in Clinical Trials

TARIO-2 is a foundation model trained on aligned tiles of human H&E pathology images and spatial RNA transcripts. Each modality is encoded into a sequence of tokens with a modality-specific encoder; then the sequences from both modalities are concatenated.

During pretraining, a transformer-based decoder is used to predict later tokens in the sequence from earlier tokens. At inference time, the model operates on clinical H&E images alone.

With TARIO-2 pretrained patient embeddings, we can predict overall survival accurately from historical pretreatment biopsy H&E alone.

This means that from a simple pathology image, the model can give a probability of a patient benefiting to a therapeutic.

[ C-index reaches a p-value of 0.04 by a 2-sided test relative to null distribution. ]

After pretraining, TARIO-2 can accurately predict spatially resolved, sub-cellular gene expression for thousands of genes across the whole genome.

This allows us to bring standard clinical pathology samples into the regime of high dimensional molecular read-outs, to decode the complexity of tumor biology.

Our mission at @NOETIK_ai is to use AI to actually cure cancers. We've been focusing on predicting therapeutic response because this is the most direct application of AI on patient outcomes.

This is just the beginning.

Check out more details in our white paper here: https://www.noetik.ai/tario2

There has been interest in solving this problem with standard pathology models in the past. We think that more biological context from spatially aligned whole transcriptome data is a massive unlock, and benchmarks prove it.

TARIO-2 outperforms large pathology foundation models at this task.

A challenge in clinical development is that inevitably N is too small - there are not enough patients to decode heterogeneity of response.

This means that successful medicines need to work in a large fraction of patients to demonstrate clinical signal. Ca we use bio foundation models to understand the complexity of patient biology in the clinic, to position the right drugs in each patient?