GRACE predicts collision cross sections using geometric residual adduct conditioning

GRACE utilizes a pretrained molecular geometry encoder with low-rank attention adapters and a learned adduct token to predict collision cross sections. The model incorporates early-fusion adduct conditioning and a residual objective against physical descriptor baselines to better capture adduct-dependent geometric effects in ion mobility mass spectrometry.