Prosit

Note: I first wrote this note in the summer of 2019, right after the Nature Methods paper dropped. When I rebuild my webside and migrate the posts I had an oppurtunity to provide edits in June 2026 with proper citation details and a check on where the models live now, since the original code repository was retired in favor of newer tooling.

Why this one stuck

I had filed spectral prediction under "neat demo" and left it there. Prosit moved it off that shelf. It predicts fragment intensities and indexed retention time from peptide sequence alone, and the predictions are good enough to feed a rescorer or build a DIA library without measuring a single real spectrum.

The shift was not in the math. Predicting spectra from sequence is a straightforward supervised learning problem if you have enough training data. The shift was in the accuracy. The predictions crossed the threshold where people started using them in production pipelines instead of publishing them and moving on.

Mirror plot comparing a Prosit-predicted MS2 spectrum against an observed spectrum, showing close agreement of fragment intensities

Figure

What you actually get

The practical impact shows up in two places. First, rescoring gets better. When you give a rescorer predicted spectra as features, it can actually distinguish good matches from noise instead of gambling on search engine scores alone. Second, you can generate spectral libraries for organisms or proteases nobody has measured. Just predict from sequence.

The interesting part is the architecture. It is not complicated. A bidirectional LSTM with attention, nothing exotic. The win came from scale. Enough training data, enough compute, and suddenly predictions that were too noisy to trust become informative.

When it matters

This matters when you need spectral libraries for something outside the standard human/mouse/yeast space, or when rescoring database searches. It matters less if your peptides sit far from the training set. Early models struggle with unusual modifications or nontryptic cleavage. The newer models on Koina handle more, but the boundaries are still real.

The other place it matters is as proof of concept. Prosit showed that deep learning could do more in proteomics than classify spectra. It could generate them. That opened doors. Since 2019, whole workflows have been built around predicted spectra. The method did not stay in papers. It moved into tools people actually run.