Inferring disruption of directed graphs using LIKA reveals altered protein phosphorylation networks in schizophrenia
Lujing Zhang, Andrew G. DeMarcoiD, Kimia Ghafari, Bernie DevliniD, Matthew L. MacDonaldiD, Kathryn RoederiD
Abstract
MotivationKinases regulate a multitude of protein functions, and their dysregulation is pivotal for many human diseases. Direct measurement of kinase activity, however, is often challenging; therefore, inferring activity from the behavior of their substrates is a widely adopted strategy. Nonetheless, traditional methods typically oversimplify the underlying network, ignoring that any particular substrate can be phosphorylated by multiple kinases. ResultsWe present LIKA, a likelihood-based framework for inferring kinase activity from phosphoproteomic data. By modeling the many-to-many structure of kinase-substrate interactions, LIKA achieves high efficiency, even with limited data, while capturing network complexity. Simulation and cell line analyses confirm the robustness and accuracy of LIKA. Importantly, analysis of a phosphoproteomic dataset from schizophrenia and control subjects reveals novel dysregulated kinases. Availability and ImplementationThe implementation code and publicly available data are provided at: https://github.com/lujingz/LIKA.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 0
- Organism
- Assay
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (7 Aug 2026) · CC BY 4.0
LIKA, a likelihood model over many-to-many kinase-substrate networks, beats KSEA in simulations with shared substrates, recovers EGFR and ABL1 in cell lines, and ranks MAPK3 first in schizophrenia cortex.
Kinase activity is usually inferred from substrate phosphorylation, yet the widely used KSEA scores each kinase's substrates in isolation, so shared substrates can falsely implicate or mask kinases, IKAP needs many parameters and six or more time points, and INKA cannot test pairwise differences. LIKA instead fits all kinase dysregulation levels jointly in one likelihood over the many-to-many network, targeting small samples and two-group comparisons.
Key results
SuggestedIn simulated networks where kinases differ in substrate count or share substrates, LIKA ranked the true dysregulated kinases more accurately than KSEA; with equal-degree, non-overlapping substrates the two methods performed the same.
Three synthetic structures, 5 samples per group (significant substrates: cases N(2, 1), controls N(0, 1)), precision@k and recall@k averaged over 100 runs. Equal degree, no overlap: both recovered all true kinases at the true k. Unequal degree: LIKA kept perfect recovery while KSEA lost precision and recall. Unequal degree plus overlap: LIKA outperformed KSEA over a broad range of k. · Simulation Results; Figure legends
SuggestedWhen K562 was contrasted with HCC827-ER3, two cell lines with known driver kinases, LIKA placed both EGFR and ABL1 in its top 10. It also identified four Src-family kinases downstream of them. KSEA recovered ABL1 only.
Beekhof et al. data with two samples per line plus a technical replicate each; 296 of 504 shared phosphosites are in KSEA, giving a 365-node network; empirical Bayes contrasts. LIKA's top 10 included EGFR and ABL1, KSEA's only ABL1; LIKA also identified FYN, SRC, LYN and HCK, direct downstream targets of the focal kinases, while KSEA identified none. · Experiment on cell line data 5; Conclusion; Figure legends
SuggestedIn postmortem dorsal anterior cingulate cortex from schizophrenia cases and matched controls, the LIKA and KSEA top-10 lists shared MAPK3, CSNK2A1, CDK1 and CDK2. LIKA ranked MAPK3 first; the authors link this kinase to schizophrenia through GWAS colocalization and 16p11.2 duplications.
112 subjects (56 schizophrenia, 56 controls pair-matched on sex, age band and postmortem interval), TMT mass spectrometry; 1,740 KSEA-mapped phosphosites, 1,231 differing at FDR < 0.05; 212 kinases in a 1,952-node network. LIKA's top-10 p-values were uniformly small, so it ranked by influence score and put MAPK3 first; KSEA ranked MAPK3 ninth by p-value. · Dataset Overview; Main findings; Conclusion
SuggestedIn two examples the authors highlight where the rankings diverged, KSEA favoured kinases whose significant substrates are shared with many other kinases (NEK3, SLK), while LIKA instead favoured TTBK2, whose substrates have no other known kinase.
NEK3 rested on one very significant site, SUGT1_T265, which 8 other kinases (including CDK1, CDK2 and MAPK3) can also phosphorylate. TTBK2 and SLK each had three differentially phosphorylated substrates; SLK's are shared with many kinases, and for SLK the reported p-values were 0.40 by LIKA vs 5.56 × 10^−5 by KSEA. · Main findings
SuggestedBoth catalytic subunits of casein kinase 2 (CK2), CSNK2A1 and CSNK2A2, scored highly, and their predicted substrates tended to be dephosphorylated in schizophrenia. In contrast, MAPK3 targets showed a long right tail, with both tails heavier than expected.
CSNK2A2: LIKA p = 1.1 × 10^−16, KSEA p = 5.9 × 10^−7; CSNK2A1 in both top-10 lists. Direction comes from Supplementary Fig. S1A-C. The authors note agreement with earlier postmortem reports of reduced cortical CK2 levels and activity, and that ERK2 (MAPK1), which CK2 regulates, was also among LIKA's top kinases. · Conclusion
Methods, data and reproducibility
Per-phosphosite t-tests (empirical Bayes variance shrinkage for small n) with Benjamini-Hochberg (BH) correction define significant substrates; LIKA numerically maximizes a likelihood in which each substrate's chance of significance is the mean dysregulation of its parent kinases. Kinases are tested against the BH null rate by a profiled signed likelihood-ratio statistic and ties broken by an influence score.
- Simulated kinase-substrate networks with known dysregulated kinases, in three structures (details in Supplementary Material 1) · n = 5 samples per group; 100 runs
- Phosphoproteomic data from cancer cell lines K562 and HCC827-ER3 (Supplementary Dataset EV2 of Beekhof et al.), 296 KSEA-recorded phosphosites · n = 2 samples per cell line plus a technical replicate each
- Postmortem dorsal anterior cingulate cortex (dACC) phosphoproteomics by tandem mass tag (TMT) mass spectrometry, schizophrenia vs pair-matched controls; 1,740 KSEA-mapped phosphosites · human (postmortem brain from schizophrenia and control subjects) · n = 112 subjects (56 schizophrenia, 56 controls)
- Curated kinase-substrate network from the KSEA web resource, subset to measured sites · n = 365 nodes (cell lines); 1,952 nodes from 212 kinases (schizophrenia)
- Methods & tools
- LIKA, KSEA, Empirical Bayes variance shrinkage, Benjamini-Hochberg procedure, TMT-labeled mass spectrometry
- Data & accessions
- https://casecpb.shinyapps.io/ksea/
- Reproducibility gaps
- Subject-level or full schizophrenia phosphoproteomic data are only available on request, subject to institutional approval and data-use restrictions; processed phosphosite inputs are in Supplementary Table S4 and the repository.
Limitations (5)
- The cell-line test rests on two samples per line plus one technical replicate each, with partial missingness; the authors rely on low within-line variability and empirical Bayes shrinkage to obtain enough signal. · stated by the authors
- Kinases with identical substrate sets cannot be told apart; LIKA merges them into one node, so it cannot say which member drives a signal. · stated by the authors
- CSNK2A1's link to schizophrenia is not established with certainty; rare mutations are only weakly associated, though the gene is strongly tied to other neurodevelopmental disorders. · stated by the authors
- Only KSEA-annotated phosphosites enter the model: 296 of 504 shared cell-line sites and 1,740 of 18,372 confidently estimated schizophrenia phosphopeptides, so kinases and sites missing from the curated network cannot be assessed. · noted in this summary
- In the simulations, substrate significance is drawn from a binomial model tied to parent-kinase dysregulation, which is close to LIKA's own assumption. KSEA is the only baseline (IKAP and INKA are discussed but not run), and the schizophrenia kinases lack orthogonal validation. · noted in this summary
Who should read:Read this if you infer kinase activity from phosphoproteomic contrasts with few samples or heavily shared substrates, or want candidate dysregulated kinases in schizophrenia cortex.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- PhosX: data-driven kinase activity inference from phosphoproteomics experiments
Alessandro Lussana, Sophia Müller‐Dott, … Evangelia Petsalaki·BioinformaticsPeer-reviewed·Nov 15, 2024
- Robust Inference of Kinase Activity using Functional Networks
Serhan Yılmaz, Marzieh Ayati, … Mehmet Koyutürk·Nature CommunicationsPeer-reviewed·Feb 19, 2021
- Comprehensive evaluation of phosphoproteomic-based kinase activity inference
Sophia Müller‐Dott, Eric J. Jaehnig, … Julio Sáez-Rodríguez·Nature CommunicationsPeer-reviewed·May 22, 2025
- Improving phosphoproteomic kinase activity inference by contextual refinement of Kinase Library relationships
Yizhi Zhang, Menghuan Zhang·bioRxiv·Nov 5, 2026
- Benchmarking substrate-based kinase activity inference using phosphoproteomic data
Claudia I. Hernández-Armenta, David Ochoa, … Pedro Beltrão·BioinformaticsPeer-reviewed·Feb 13, 2017
- Accurate Prediction of Kinase-Substrate Networks Using Knowledge Graphs
Vít Nováček, Gavin McGauran, … Dirk Fey·PLoS Computational BiologyPeer-reviewed·Dec 3, 2020