CysLENS combines chemoproteomics with a protein language model to identify targetable cysteines
CysLENS ranks chemical interactions with proteins to guide follow-up experiments
On August 27, the authors posted the CysLENS preprint. On an independent measurement set, the method’s mean precision in selecting reproducible interactions for similar molecules was 1,4–2,2 times higher than that achieved by ranking interactions using a single binding metric.
A small molecule that forms a stable chemical bond with a protein can alter the protein’s function. The molecule must bind to a specific site on the protein, and this site is often a cysteine, one of the amino acids. Chemoproteomics identifies such interactions in protein mixtures derived from cells. After a fragment is added, a chemical probe labels the unoccupied cysteines, and a mass spectrometer, which distinguishes molecules by mass, measures the probe signal. If the fragment has already bound to a site, the probe signal decreases.
A strong signal can sometimes result from the fragment’s intrinsic reactivity or from measurement variability. The authors therefore constructed a broad map in a protein mixture derived from A375 melanoma cells. They tested 940 small fragments arranged into 470 enantiomer pairs, which are mirror-image spatial forms of the same molecule. The screen covered more than 45 thousand cysteines across more than ten thousand proteins.
Enantiomers provide a physical control because they have similar chemical properties, while the protein environment distinguishes between their spatial forms. CysLENS combines this feature with interaction strength and selectivity, fragment similarity, previous data, and the similarity of the sequence environments surrounding cysteines in proteins. The last feature is assessed by ESM-2, a language model that identifies patterns in protein sequences. For each combination of a similar chemical class and a similar cysteine environment, the contribution of each feature remains visible.
The authors tested whether this ranking produced more reproducible findings using an independent measurement set that was not included in the CysLENS calculations. For each interaction in the new screen, they identified a structurally similar fragment and determined whether it bound to the same cysteine. At four chemical similarity thresholds, the highest-ranked CysLENS results were confirmed by this comparison more often than results ranked solely by the original binding metric.
DNMT1, an enzyme that maintains chemical marks on DNA, ranked first. Spiro-oxapiperidine fragment 5 selectively bound to cysteine C1478 of DNMT1 in the protein mixture. A concentration series confirmed that the signal depended on dose, and the interaction persisted in living A375 cells. A computational model placed the fragment near the DNA-recognition surface of DNMT1. The ranking provides a specific choice for the next experiment: which molecule to test, which protein site to examine, and which observations support that choice.