Researchers at Columbia University have developed a machine learning model called PFNet that calculates localized protein stabilities from standard mass spectrometry measurements, according to a study published in Nature Communications. The peer-reviewed paper, published on September 1, 2026, details how the tool extracts residue-level free energy of opening, designated as ∆Gop, directly from peptide-level hydrogen exchange-mass spectrometry (HX-MS) datasets.
The residue-level free energy of opening serves as a thermodynamic descriptor for localized stability across protein ensembles under physiologically relevant timescales and conditions. PFNet determines ∆Gop instantly for arbitrarily large proteins and complexes. The researchers reported that the computational system establishes a quantitative, scalable, and accessible analysis framework for conventional HX-MS experiments.
Chenlin Lu and Kyle C. Weber contributed equally as lead authors on the project, working alongside Savannah K. McBride, Andrew Reckers, and senior author Anum Glasgow in Columbia University's Department of Biochemistry and Molecular Biophysics. The authors declared no competing financial interests in the publication.
The study utilized data and samples from several external research groups during development and testing. Dr. Gabriel Rocklin and his group supplied NMR datasets alongside EEHEE_rd4_087 and HHH_rd4_0557 samples, while Drs. Susan Marqusee, Shawn Costello, and Sophie Shoemaker provided raw experimental data files for the SARS-CoV-2 spike and ACE2 interaction. Dr. Mohammed AlQuraishi at Columbia University Medical Center contributed through technical discussions.
The National Institutes of Health funded the research through grants R35GM157185 and R21EB035208 awarded to Glasgow, with additional support from National Science Foundation Graduate Research Fellowships awarded to Weber and McBride. Nature Communications received the original manuscript on June 3, 2026, and accepted it for publication on July 7, 2026.
