Detecting Sparse Colorectal Cancer Signals from Multi-Modal Cell-Free DNA Representations Using Modern Hopfield Attention

Next generation sequencing-based early cancer detection from cell-free DNA (cfDNA) presents an extreme-scale multiple instance learning challenge: identifying rare tumor signals amidst millions of instances per sample (with witness rates as low as $<$0.0001\%). It also provides a testbed for translating DNA foundation model embeddings to a clinically important supervised learning task. We propose Fragment-Level Deep Learning (FLDL), an end-to-end deep learning framework that combines multi-modal fragment features, including a HyenaDNA-derived sequence embedding, with Modern Hopfield Networks to perform dense associative retrieval over the massive cfDNA instance space. Using held-out real-world clinical and challenging contrived test sets, we compare FLDL’s performance to a state-of-the-art machine learning model and to a deep learning model without attention (max pooling). Our results demonstrate that only the attention-based FLDL model outperforms the machine learning model, in spite of a modest training set size ($n = 4,394$). FLDL also scales effectively with sample size and with number of instances per sample while offering useful biological insights via attention weights and learned sample representations. This work establishes a new frontier for foundation-model-augmented cfDNA representations and highly scalable attention-based deep learning in clinical cfDNA diagnostics.