PROBIXI - Self-Calibrating (PROB)ab(I)listic Peak Detection for Serial (X)-Ray Crystallograph(I)c Data

Serial crystallography reconstructs a molecular structure from many thousands of single, sparse diffraction images. Data recovery begins by separating Bragg peaks from detector background, a step called peakfinding. Conventional peakfinding algorithms require hand-tuned parameters which must be retuned for every sample and beamline. We reframe peakfinding as online anomaly detection against a learned noise prior. $\text{probixi}$ treats the frame stream as samples from the detector's background distribution, maintaining a running mean and variance as a mixture of panel, pixel, and rotational components, whitening each arriving frame against that prior, and reporting pixels as anomalies under the null $Z \sim \mathcal{N}(0,1)$. The peak contrast, prior, mixture weights, and detection threshold are all solved on a short seed sequence rather than tuned by hand. We evaluate this work on 8.72 million frames of weak diffraction patterns of channelrhodopsin. After peak determination, $\text{probixi}$ indexes the data by searching for an orientation and predicting spots with awareness of the peak likelihood per pixel. $\text{probixi}$ indexed 2.30\% of frames against 0.44\% from a tuned peakfinder8-$\text{xgandalf}$ indexing pipeline, merged $R_{\mathrm{split}}$ fell from 21.83\% to 5.08\%, and $ \langle I/\sigma(I) \rangle$ rose from 2.04 to 13.54.