When a cell shifts its metabolism toward glycolysis — burning glucose without oxygen, as tumors and stressed tissues do — lactate accumulates and some of it is covalently attached to lysine amino acids on working proteins, a modification called lysine lactylation. Most research has treated this as a slow, gene-regulating process on histones, but recent work shows it also tags enzymes and RNA-binding proteins directly, changing how stable they are and who they bind to, with no change in gene expression. The central problem is that this tagging is now known to be widespread, yet for the vast majority of tagged proteins we have no idea which physical binding partners it gains or loses, because every mechanistic study so far tested one pre-selected partner at a time. The result is a missing layer of the metabolic-reprogramming story: cells may rewire their metabolism not just by turning genes on and off, but by re-stitching the physical protein complexes that carry out metabolism, and current tools cannot see this.
We propose to close that gap with a paired experimental and computational program. Experimentally, we will run the first proteoform-contrasted, unbiased co-immunoprecipitation mass spectrometry study of lactylated proteins: pulling out each tagged bait protein under glycolytic versus oxidative conditions and, in parallel, pulling out a lactylation-dead mutant of the same protein, so that every recovered binding partner can be scored for whether its binding depends on lactylation specifically. Computationally, we will build an edge-level functional enrichment method that interprets these binding changes pathway-by-pathway — something existing enrichment tools, which were built for gene-expression lists, structurally cannot do.
Core Claim
The claim is that non-histone lysine lactylation acts primarily as an edgetic switch: rather than reprogramming transcription, it selectively creates and destroys individual protein-protein interaction edges within metabolic machines, and this rewiring can be measured systematically and interpreted with a purpose-built enrichment analysis. The concept of edgetic perturbation — disease mechanisms that break or form specific interaction links while leaving the protein otherwise intact — comes from the HuRI interactome project, whose authors note that disease variants often act this way (A reference map of the human protein interactome). But HuRI's yeast two-hybrid map of roughly 53,000 binary interactions is, by its authors' own admission, blind to mammalian post-translational modifications, so lactylation-dependent edges are invisible to it. Two mechanistic papers motivate the claim while leaving it untested: the KAT8–PCK2 study in hepatic ischemia/reperfusion injury showed that lactate-driven lactylation of PCK2 stabilizes it and lets it physically bind OXSM, blocking Parkin-mediated degradation and reprogramming mitochondrial fatty acid synthesis — but every interaction there was confirmed by targeted, single-hypothesis co-IP against pre-chosen candidates (The Lactate‐Primed KAT8‒PCK2 Axis Exacerbates Hepatic Ferroptosis During Ischemia/Reperfusion Injury by Reprogramming OXSM‐Dependent Mitochondrial Fatty Acid Synthesis); and the H. pylori study showed lactylation of WTAP at two lysines stabilizes it and changes its RNA binding, but never asked what lactylation does to WTAP's physical partners in the m6A writer complex (<i>Helicobacter pylori</i> infection aggravates hepatic steatosis by lactylation-driven WTAP-mediated m <sup>6</sup> A modification). Reviews of the field name unbiased identification of binding partners for non-histone lactylated proteins as an open question and state that non-histone lactylation mechanisms "require further investigation" (Lysine lactylation in diseases: beyond histone lactylation; Lactylation in cancer: Mechanisms in tumour biology and therapeutic potentials). We will answer it.
Research Program
The experimental arm uses a 2×2 factorial design in a hepatocyte model (the mechanistic anchor studies are both hepatic), crossing two metabolic states with two bait proteoforms. Metabolic state is set by culture conditions that force glycolysis with high lactate (hypoxia plus high glucose, or exogenous lactate as used in the WTAP study) versus conditions that force oxidative phosphorylation, and verified by Seahorse oxygen-consumption and extracellular-acidification rates plus targeted metabolomics, following the flux-phenotyping logic of the metabolic-reprogramming literature (Absolute dynamic and relative static: the relationship of glycolysis and OXPHOS in cancer development; The Lactate‐Primed KAT8‒PCK2 Axis...). Bait proteoform is wild-type versus a lactylation-dead lysine-to-arginine mutant at the mapped lactylation sites, the standard site-directed-mutagenesis approach used in both anchor papers. Each bait is affinity-purified under all four conditions, co-bound partners are digested and quantified by LC-MS/MS, and the writer KAT8 is knocked down as an orthogonal perturbation, since it is the identified lactyltransferase for PCK2 (The Lactate‐Primed KAT8‒PCK2 Axis...). The bait panel starts with PCK2 (positive control — the design must recover the known OXSM edge) and WTAP (whose writer-complex rewiring is a genuine open question), then expands to several substrates from published proteome-wide lactylation maps, such as ALDH2 and FAS (Lysine lactylation in diseases: beyond histone lactylation).
For each prey protein $p$ bound to bait $b$, we define the lactylation-dependent edge score from the condition fold-change in co-IP abundance, computed separately for wild-type and mutant bait and differenced:
$$E_{b,p} \;=\; \Delta_{p}^{\mathrm{WT}} - \Delta_{p}^{\mathrm{K/R}}, \qquad \Delta = \log_2 \frac{\text{prey abundance under glycolysis}}{\text{prey abundance under OXPHOS}}$$
A prey whose binding shifts with metabolic state only when the bait can be lactylated gets a large $|E_{b,p}|$; the mutant arm removes the metabolic-state and total-abundance confounds that make conventional pull-downs ambiguous. Statistically, $E$ is the interaction term of a two-factor model, and we control the known co-IP failure modes directly: lysis artifacts by validating hits with in-situ proximity ligation, which only reports when two proteins sit within about 40 nm inside intact cells (Ian T. Foster et al.), non-specific sticky background by mutant-bait and beads-only controls as characterized in the influenza co-IP-MS study (Proteomics Analysis of Cellular Proteins Co-Immunoprecipitated with Nucleoprotein of Influenza A Virus (H7N9)), and directness by recombinant GST pull-down as in the PCK2 work.
The computational arm turns per-bait ranked prey lists into pathway-level interpretation. The core difficulty is that standard enrichment analysis treats gene sets as flat, unordered bags and was built for expression data; when applied to pull-down proteomics it is typically forced through classical overrepresentation analysis, as in the influenza study where 17 co-purified host proteins were mapped to the spliceosome via GO and KEGG (Proteomics Analysis of Cellular Proteins Co-Immunoprecipitated with Nucleoprotein...). We instead run fast rank-based enrichment over the $E$-scored prey lists — using the fGSEA machinery that processes a full ranked list in under a minute where traditional GSEA takes 5–20 minutes (Gene-set enrichment analysis and visualization on the web using EnrichmentMap:RNASeq) — and then cluster the enriched pathways with a contrastive embedding that aligns gene-set membership vectors with PubMedBERT text descriptions, extending gtCLIP (Contrastive Learning for Gene Set Enrichment Analysis Post-Processing). This extension is not gratuitous: gtCLIP's authors showed that over 80% of pathway pairs share no genes at all, so overlap-based clustering silently fails exactly where we need it, and they explicitly name incorporating protein-protein interaction networks and pathway topology as their next step. Our contribution is to feed the embedding physical edges — endpoints and direction of rewiring — rather than expression ranks. Before any wet-lab data exist, we will pre-test the method on DPPIN, a public suite of 12 yeast dynamic protein-protein interaction networks spanning 36 time-course snapshots over three metabolic cycles (DPPIN: A Biological Dataset of Dynamic Protein-Protein Interaction Networks), asking whether edge-resolved enrichment across snapshots recovers the metabolic-cycle phase structure that static networks miss.
Baselines
The naive reference point is to take our lactylated bait list, map it onto the static HuRI interactome of approximately 53,000 binary interactions (A reference map of the human protein interactome), and run classical overrepresentation analysis (Fisher's exact test over GO and KEGG) on the resulting partners — reproducing what the field does today. It is the strongest naive competitor because it is free, standardized, and would be the first thing any collaborating lab would do with our data. We will run it on the exact same LC-MS/MS prey lists and the same gene-set databases, with no tuning advantage, so the comparison isolates only the analysis method. Beating it means two things: recovering lactylation-dependent edges that are absent from HuRI (predictable, since yeast two-hybrid cannot capture PTM-dependent binding), and recovering pathway-level structure that overrepresentation analysis misses because it ignores edge direction and the zero-overlap problem.
The strongest existing end-to-end competitor is the transcriptomic pipeline: matched RNA-seq from the same 2×2 conditions, run through fGSEA with the Bader Lab gene-set database and visualized with EnrichmentMap network clustering (Gene-set enrichment analysis and visualization on the web using EnrichmentMap:RNASeq), which is the current standard tool for exactly this analysis style. We will collect that RNA-seq ourselves — same cells, same timepoints, same conditions — and give it the same gene sets and compute, so the comparison is between data layers, not pipelines. Beating it means detecting pathway rewiring that transcriptomics misses, with the cleanest test being the known ground truth: the PCK2–OXSM axis, which is regulated entirely post-translationally through lactylation-blocking of Parkin-mediated ubiquitination and should be invisible at the transcript level (The Lactate‐Primed KAT8‒PCK2 Axis...).
On the post-processing side, the strongest competitor is gtCLIP itself, which reports 3.3-fold higher within-cluster gene set overlap, 3.9-fold higher silhouette scores, 92.4% mean NES sign coherence, Recall@5 of 51.3%, and 1.35x higher within-cluster overlap in cross-disease zero-shot transfer than the overlap-based GeneSetCluster 2.0 (Contrastive Learning for Gene Set Enrichment Analysis Post-Processing). We will run unmodified gtCLIP on our enrichment output versus our edge-extended version, on the same corpus of 12,612 gene sets, with matched embedding and training budgets; beating it means higher silhouette and within-cluster overlap on Co-IP-MS-derived enrichment results, where edge endpoints carry information that expression-only embeddings cannot. Finally, two ablations isolate the key ingredients: running the pipeline on total-protein pull-downs without the wild-type-versus-mutant contrast (testing whether proteoform resolution, not just metabolic perturbation, drives the signal), and replacing the embedding clustering with Jaccard-overlap clustering in the EnrichmentMap style (Enrichment Map: A Network-Based Method for Gene-Set Enrichment Visualization and Interpretation), which should collapse on the zero-overlap pairs. For in-silico site triage, the reference is MeToken, which reports an F1-score of 50.40% on its large-scale benchmark (versus 43.94% for ESM-2) and 61.95% on qPTM (versus 57.40% for ESM-2 and 45.20% for CDConv) (MeToken: Uniform Micro-environment Token Boosts Post-Translational Modification Prediction); we use it only to prioritize second-round mutants, and we will report its precision against our mass-spectrometry-confirmed sites.
Why not just…?
Why not just do RNA-seq and GSEA? This is the cheapest and most standard route, and we will in fact collect the RNA-seq as a baseline rather than argue it away. But the mechanistic literature gives direct evidence it falls short: in the PCK2 study, lactylation controls protein stability and physical binding to OXSM by blocking ubiquitination, entirely downstream of transcription (The Lactate‐Primed KAT8‒PCK2 Axis...), and the broader acylome literature shows such modifications tune enzyme activity and machinery without rewriting transcription (Protein succinylome analysis identifies citrate synthase as a central regulator of osteoclast metabolic activity). The head-to-head comparison in our baseline design settles the question empirically: if GSEA on matched RNA-seq recovers the same pathways with the same confidence, the proteoform layer adds nothing and we will say so.
Why not just query static interactome databases like HuRI or STRING? Because the lactylation-dependent edges we are hunting are exactly what those maps cannot contain. HuRI's authors state plainly that their screen tests protein pairs in yeast, so interactions requiring post-translational modifications or third-party proteins are missed (A reference map of the human protein interactome), and the DPPIN benchmarks show static graphs under-represent temporally reconfiguring complexes — temporal modeling reached 95.58% node-classification accuracy on DPPIN-Ito versus 94.35% for the static equivalent (DPPIN: A Biological Dataset of Dynamic Protein-Protein Interaction Networks). That said, HuRI remains genuinely useful as a prior for separating novel edges from known ones, and we use it that way rather than discarding it.
Why not just run Fisher's exact test on the pull-down hit lists, like the influenza study did? This is a real threat for baits with large effect sizes, and we will run it as a baseline rather than dismiss it. The concrete failure modes are documented: over 80% of pathway pairs share no genes, so overlap-based grouping cannot link related processes (Contrastive Learning for Gene Set Enrichment Analysis Post-Processing), and the bag-of-genes treatment cannot represent which edge changed, in which direction, or whether the changed prey sits at a pathway hub or periphery. The decisive check is whether ORA recovers the known PCK2-driven mitochondrial-fatty-acid-synthesis and ferroptosis signal from our data with fewer pathway calls and weaker confidence than the edge-level method, at matched false-discovery control.
Why not just predict the sites computationally and skip the wet lab? Because the best current predictors are screening filters, not replacements: MeToken's 50.40% F1 on the large-scale benchmark and 61.95% on qPTM (MeToken: Uniform Micro-environment Token Boosts Post-Translational Modification Prediction) leave roughly half of predictions unverified, and no sequence-structure model can tell you whether intracellular lactate under a given metabolic state actually occupies the site, let alone what happens to binding partners afterward. We use MeToken for what it is good at — prioritizing which lysines to mutate next — and keep mass spectrometry, mutagenesis, and co-IP as the arbiter.
Proposed Experiments and Decisive Checks
The program proceeds in three gates. First, the positive-control gate: in the PCK2 arm, the design must recover OXSM as a lactylation-dependent edge in wild-type but not K100R bait under glycolytic conditions, and lose it under KAT8 knockdown, matching the targeted co-IP results of the original study (The Lactate‐Primed KAT8‒PCK2 Axis...). Failure here invalidates the platform. Second, the discovery gate: unbiased scoring across the bait panel yields a ranked list of new lactylation-dependent edges, of which we pre-register the top twenty for orthogonal validation by proximity ligation assay in intact cells, targeted co-IP and Western blot, and recombinant GST pull-down for directness; we set a pre-registered threshold of at least half confirmed, because no reported result found on Lacuna gives an existing validation-rate benchmark for lactylation-dependent edges to calibrate against. Third, the interpretation gate: the edge-level enrichment method, pre-tested on DPPIN's 36 metabolic-cycle snapshots, is applied to the full dataset and compared against the ORA, EnrichmentMap, and gtCLIP baselines on identical inputs, with silhouette score, within-cluster gene-set overlap, and recovery of the known PCK2–mtFAS–ferroptosis axis as the primary metrics. If the edge layer beats every baseline on pathway recovery while the total-protein ablation does not, the core claim stands: metabolic reprogramming rewires metabolism one lactylation-dependent interaction edge at a time, and both our maps and our analysis method are needed to see it.
References
- A reference map of the human protein interactome
- The Lactate‐Primed KAT8‒PCK2 Axis Exacerbates Hepatic Ferroptosis During Ischemia/Reperfusion Injury by Reprogramming OXSM‐Dependent Mitochondrial Fatty Acid Synthesis
- Helicobacter pylori infection aggravates hepatic steatosis by lactylation-driven WTAP-mediated m 6 A modification
- Lysine lactylation in diseases: beyond histone lactylation
- Lactylation in cancer: Mechanisms in tumour biology and therapeutic potentials
- Absolute dynamic and relative static: the relationship of glycolysis and OXPHOS in cancer development
- HPF-MPI: A Programming System Supporting Task and Data Parallelism
- Proteomics Analysis of Cellular Proteins Co-Immunoprecipitated with Nucleoprotein of Influenza A Virus (H7N9)
- Gene-set enrichment analysis and visualization on the web using EnrichmentMap:RNASeq
- Contrastive Learning for Gene Set Enrichment Analysis Post-Processing
- DPPIN: A Biological Dataset of Dynamic Protein-Protein Interaction Networks
- Enrichment Map: A Network-Based Method for Gene-Set Enrichment Visualization and Interpretation
- MeToken: Uniform Micro-environment Token Boosts Post-Translational Modification Prediction
- Protein succinylome analysis identifies citrate synthase as a central regulator of osteoclast metabolic activity