Optimizing Casanovo for Real Time De Novo Peptide Sequencing

Casanovo is a transformer-based de novo peptide sequencer that is too slow for real-time instrument use. I will profile the new non-autoregressive pipeline, apply CUDA graphs, mixed precision, and FlashAttention-2 to reduce per-spectrum latency, then train lightweight student models via knowledge distillation to reach the 20-100 Hz real-time target required by modern mass spectrometers.