PhosSight

A theoretical spectral library for a phosphoproteomics run contains every peptidoform the digest could produce, and roughly half of them will never reach the detector: the phosphate group shifts ionization efficiency and solvation energy enough that the peptide is simply invisible on the instrument. Searching them costs compute and inflates the chance of a random match. The mirror-image problem shows up in data-dependent acquisition, where the labile phosphoester bond dominates fragmentation, sequence-determining ions are suppressed, and a real low-abundance phosphopeptide gets scored away as noise. Detectability predictors already existed for unmodified peptides, but models trained without phosphate residues do not transfer.

PhosSight, from Shanghai Advanced Research Institute, CAS, closes that gap with PhosDetect — a compact recurrent network that reads a peptide sequence, including pSer, pThr and pTyr tokens, and returns one number: the probability that the peptide is experimentally detectable. The framework then spends that number twice. In DDA it becomes a third deep-learning feature alongside predicted retention time and fragment intensity, driving site relocalization and semi-supervised rescoring. In DIA it becomes a filter, pruning the in-silico library to its most detectable half before the search ever starts.

The catalog holds neighbours that answer adjacent questions — Prosit-PTM predicts fragment intensities for modified peptides, PhosF3C predicts which residues carry a phosphate, and DIA-BERT scores DIA precursors. PhosSight asks whether the instrument will see the peptide at all. It extends the group's earlier DeepRescore2 workflow, and appeared in Advanced Science in May 2026 after a March 2026 bioRxiv preprint.

#Key Features

• Phosphorylation-aware detectability: A learned sequence embedding and an explicit physicochemical channel — hydrophobicity, net charge and polarity, with distinct values assigned to phosphorylated residues — are fused by a sigmoid gate that weighs one stream against the other.
• Detectability as a rescoring feature: Added to retention time from AutoRT and fragment intensity from pDeep3, the score resolves ambiguous positional isomers where spectral evidence alone is inconclusive, then feeds Percolator at a 1% q-value and 0.75 localization probability cutoff.
• Library pruning for DIA: Filtering a DIA-NN library to its top 50% most detectable precursors cut search time by roughly 40% while entrapment false discovery stayed below 0.65%.
• Engine-agnostic: The DDA workflow wraps Comet, MaxQuant, MS-GF+, X!Tandem and MSFragger, and gains held across all of them on both label-free and TMT data.
• Shipped weights with optional adaptation: The trained checkpoint runs on arbitrary sequences as released; fine-tuning on a few pilot runs to capture instrument-specific ionization is optional, not required.

#Technical Details

PhosDetect tokenizes peptides into a 24-character vocabulary — 20 amino acids, three phosphorylated residues and a padding token — embeds them in 10 dimensions, gates that stream against the physicochemical channel, and passes the result through a two-layer bidirectional GRU with 20 hidden units per direction. A self-attention layer aggregates the recurrent outputs, and batch normalization plus a dense sigmoid emits the detectability probability. The released checkpoint holds about 13,000 parameters. Training used binary cross-entropy with Adam at a learning rate of 5×10⁻⁴, plateau-based decay and early stopping. The corpus was balanced to 319,088 sequences — 79,772 each of phospho-positive, phospho-negative, unmodified-positive and unmodified-negative — with positives drawn from a large human phosphoproteome (PRIDE PXD012174) at 1% FDR and class-I localization, and negatives enumerated by in-silico tryptic digestion of PhosphoSitePlus and the UniProt human proteome. Held out on four species datasets from DeepDetect, PhosDetect reached AUC 0.89–0.96 as released and 0.93–0.98 fine-tuned; on two phosphoproteomic datasets it reached 0.86–0.89 and 0.92–0.96, improving precision 1.5- to 2.75-fold over prior detectability models. On a synthetic ground-truth phosphopeptide library it lifted correctly localized PSMs from 8,061 to 9,632 at 1% false localization rate, 164 of which no other method recovered.

#Applications

The framework targets phosphoproteomics groups running cohort-scale clinical studies, where missing values fragment the data matrix and blunt statistical power. Reprocessing 183 CPTAC uterine corpus endometrial carcinoma samples across 17 TMT10-plexes, PhosSight reported 27,237 quantifiable phosphosites against 23,232 from the conventional workflow, a 17% expansion that supplied the substrate coverage needed to tie MARK2 kinase activity to poor survival. The operational cost is real: the DDA workflow runs on Windows with Docker, and the DIA workflow needs Linux, Singularity, DIA-NN 2.2.0 and 200 GB of RAM.

#Impact

PhosSight makes detectability a usable feature for modified peptides, which had been the missing piece keeping library-pruning strategies out of phosphoproteomics. Code, the trained checkpoint and fine-tuning scripts are on GitHub under MIT, with processed datasets on Zenodo under CC BY. The scope is deliberately narrow: PhosDetect covers only pSer, pThr and pTyr, not acetylation, ubiquitination or combinatorial modifications, which the authors attribute to scarce training data elsewhere. It runs strictly after acquisition or before the search rather than inside the instrument's acquisition logic, and its DIA validation covers DIA-NN rather than diaPASEF.