Skip to content

funscoR

Verdict: The first systematic attempt to separate functional phosphosites from noise -- reveals that most detected phosphosites likely have no regulatory role, which is an uncomfortable but important finding.

Citation: Ochoa D, Jonikas M, Lawrence RT, et al. "The functional landscape of the human phosphoproteome." Nature Biotechnology 38:365-373 (2020). DOI: 10.1038/s41587-019-0344-3; Ochoa D, Jarnuczak AF, Vieitez C, et al. "An atlas of human kinase regulation." Molecular Systems Biology 17(3):e10066 (2021). DOI: 10.15252/msb.202010066

Problem Setup

Mass spectrometry detects over 100,000 human phosphosites, but how many are functionally relevant? Many may be non-functional "noise" -- stochastic phosphorylation with no regulatory consequence. Distinguishing functional from non-functional sites is essential for interpreting phosphoproteomics data. funscoR asks: can machine learning predict which phosphosites are likely functional?

Approach

A random forest classifier trained on features that distinguish known functional phosphosites from background. Features include evolutionary conservation (across vertebrates), structural context (surface accessibility, proximity to functional domains), stoichiometry estimates, co-regulation patterns, and overlap with known kinase motifs. Training positives are well-characterized regulatory sites from PhosphoSitePlus (sites with known biological function). Training negatives are randomly sampled detected phosphosites. The output is a continuous functional score (0 to 1) for every detected human phosphosite.

Key Findings

  • Only approximately 15% of detected phosphosites show strong evidence of function, based on the classifier's score distribution.
  • Evolutionary conservation is the single most predictive feature -- functional sites are conserved across vertebrates, non-functional sites are not.
  • High-stoichiometry sites are more likely to be functional than low-stoichiometry sites.
  • The functional score correlates with sensitivity to kinase inhibitor perturbation -- high-scoring sites respond more to targeted inhibition.
  • The score provides a principled way to filter phosphoproteomics datasets before downstream analysis.

Evaluation

Benchmarked via cross-validation on known functional sites. Validated by testing whether high-scoring sites are enriched among sites that respond to kinase inhibitor treatment in independent perturbation datasets. Also assessed correlation with evolutionary conservation and structural features not used in training.

Honest Assessment

Strengths:

  • Addresses a fundamental question that most phosphoproteomics studies ignore: is this site actually doing anything?
  • Provides a practical, downloadable score for every known human phosphosite.
  • The finding that ~85% of phosphosites may be non-functional is sobering and important for the field.
  • Feature importance analysis is transparent and biologically interpretable.

Limitations:

  • Training labels are circular to some degree -- "known functional" sites are biased toward well-studied kinases and pathways, so the model may underestimate functionality of sites in understudied pathways.
  • The binary functional/non-functional framing is a simplification; some sites may be conditionally functional (e.g., only in specific tissues or disease states).
  • Conservation-based features dominate, which means the model is essentially a sophisticated conservation filter with extra features.
  • Cancer-specific phosphorylation events (neomorphic sites created by mutations) are poorly served by a model trained on conserved biology.

Design Decision: Frame functionality as a classification problem using curated positive examples. This is pragmatic -- the alternative (requiring experimental validation for every site) is infeasible -- but it means the model reflects current knowledge biases. The 15% estimate should be treated as a lower bound on functionality, not ground truth.