Predicting which cancer patients will respond to immunotherapy is a major goal in precision medicine. While RNA sequencing has become a powerful tool to help with this, results can vary a lot between studies because of differences in patient groups and data collection methods. These inconsistencies make it hard to build a single model that works well across different datasets.

Researchers from the Cancer Data Science Lab at NIH have introduced a new solution called IC2Bert. This machine learning model uses a unique approach that starts by pretraining on gene expression patterns in a way similar to how language models learn from text. Then, it fine-tunes the model using specific datasets focused on cancer immunotherapy, helping it adapt to the biological variability across different patient groups.

(A) Training and Validation (Iterative LODOCV) Scheme for IC2Bert. (B) Model Architecture of IC2Bert.

To test IC2Bert, the researchers used a method called Leave-One-Dataset-Out Cross-Validation, which evaluates the model’s ability to generalize to entirely new data. The results showed that IC2Bert consistently outperformed existing tools, providing more reliable predictions about who is likely to benefit from immune checkpoint blockade therapy.

Availability – The IC2Bert model and its source code are publicly available on GitHub: https://github.com/data2intelligence/ic2bert.