To model learning in C. elegans, we used a simple yet robust associative learning paradigm called salt associative learning. Briefly, this assay involves training worms to associate the absence of salt (NaCl) with the presence of food. C. elegans is typically grown in the presence of salt (usually ~50 mM) and displays an attraction toward this concentration when assayed for chemotaxis behaviour on a salt gradient (Kunitomo et al., 2013; Luo et al., 2014). Training/conditioning with ‘no salt +food’ partially attenuates this attraction (group referred to ‘trained’). This is because the presence of abundant food (unconditioned stimulus) is a strong innate attractive cue, and pairing this with ‘no salt’ (the conditioned stimulus) leads to the animals showing the same behaviour towards the conditioned stimulus as they do to the unconditioned stimulus, that is attraction towards no salt, reflected as a preference for lower salt concentrations (Hiroki et al., 2022; Nagashima et al., 2019). Similar behavioural paradigms involving pairings between salt/no salt and food/no food have been previously described in the literature (Nagashima et al., 2019). Here, learning experiments were performed by conditioning worms with either ‘no salt +food’ (referred to as ‘salt associative learning’) or ‘salt +no food’ (called ‘salt aversive learning’).

To identify the learning proteome, we adapted this learning paradigm to incorporate TurboID-catalysed biotinylation of proteins specifically during the learning/conditioning step. We did this by (1) performing the salt associative learning assay on transgenic animals expressing TurboID in the entire nervous system (Prab-3::TurboID) and (2) adding biotin only when the worms are being trained (i.e. exposed to both food and ‘no salt’ in the ‘trained’ group, or to food and ‘high salt’ concentrations in the ‘high-salt control’ group). As an additional control, we performed the same assay on non-transgenic (non-Tg) animals that do not express TurboID (Figure 1A and B). We then isolated proteins from >3000 whole worms per group for both ‘high-salt control’ and ‘trained’ groups, most of which were subjected to a sample preparation pipeline for mass spectrometry, and some of which were probed via western blotting to confirm the presence of biotinylated proteins. The same pipeline was used to generate five biological replicates.

Validation of TurboID-catalysed biotinylation was performed in two ways: First, we compared total protein from naïve/untrained animals that are non-Tg versus TurboID-encoding by western blot and probed for V5-tagged TurboID: as expected, we observed expression in transgenic worms only at the predicted size (39 kDa) (Figure 1C). Secondly, we tested if exposure to biotin increased the biotinylation signal in a TurboID-dependent manner. To do this, we quantified the biotinylation signal in (1) naive non-Tg worms not exposed to biotin, (2) non-Tg C. elegans exposed to biotin for 6 hr, (3) naive TurboID worms not exposed to biotin, and (4) TurboID animals exposed to biotin for 6 hr. Although background biotinylation was present in worms not treated with biotin, we found that biotin exposure increased the signal 1.3-fold for non-Tg and 1.7-fold for TurboID C. elegans (Figure 1—figure supplement 1). Taken together, these findings indicate that there is increased biotinylation of proteins in the presence of both biotin and the TurboID enzyme.

Mass spectrometry experiments were performed with the following experimental groups per replicate: (1) non-transgenic/non-Tg high-salt control, (2) non-Tg trained, (3) TurboID high-salt control, and (4) TurboID trained. We did not include no-biotin treatment controls due to the practical challenges of handling >4 groups in the combined learning assay/mass spectrometry pipeline, for which >3000 worms are required per group. Therefore, all groups were exposed to biotin during the 6 hr exposure period to food and either high salt (for control) or no salt (for trained; Figure 1A).

To confirm that each experimental group displayed the expected phenotype after training, a portion of worms from all groups was tested using a chemotaxis assay. The chemotaxis index (CI) was used as a readout of learning performance: a positive CI reflects high salt preference, a CI close to 0 represents a more neutral response, and a negative CI represents low salt preference (Figure 1—figure supplement 2). We confirmed after each learning assay that naïve/untrained worms had a strongly positive CI (~0.7–0.9), whereas trained animals showed a lower CI (~0.0). We also performed a learning control (indicated as high-salt ‘control’) in which the presence of food (the US) is paired with high salt concentrations – worms in this group are attracted to high salt and showed a strongly positive CI (~0.7–0.9; Figure 1B), displaying a similar behaviour to naïve worms. This behavioural change seen in trained animals, versus the naïve and high-salt control groups, represented successful learning as seen in previous studies (Hiroki et al., 2022; Nagashima et al., 2019). This was observed in both non-Tg and transgenic animals, confirming that introducing the transgene did not perturb learning (Figure 1B).

We also confirmed by western blotting that biotinylated proteins could be observed in TurboID-expressing high-salt control and trained groups (Figure 1C). As in other C. elegans studies utilising TurboID, we saw background biotinylation in non-Tg controls; however, this is visually lower compared with groups from TurboID transgenic worms (Figure 1C; Artan et al., 2021; Sanchez et al., 2021). Quantification of the signal within entire lanes showed a 1.1-fold increase in the ‘TurboID, control’ lane compared with the ‘non-Tg, control’ lane, and a 1.9-fold increase in the ‘TurboID, trained’ lane compared with the ‘non-Tg, trained’ lane. For all replicates, we determined that biotinylated proteins could be observed from total TurboID-positive worm lysate by western blotting before proceeding with downstream proteomic experiments (Figure 1—figure supplement 3, Supplementary file 1B).

Our sample preparation methodology for mass spectrometry is based on similar protocols used in C. elegans and other systems (Artan et al., 2022; Sanchez et al., 2021; Prikas et al., 2020). We performed five biological replicates, in line with other C. elegans studies (Artan et al., 2022; Holzer et al., 2021). To examine the learning proteome, we first subtracted proteins from ‘TurboID, trained’ groups also present in ‘Non-Tg, trained’ samples to generate a protein list specific to ‘TurboID, trained’ animals for each biological replicate. We next subtracted from ‘TurboID, control’ lists any proteins that appeared in ‘Non-Tg, control’ samples to generate a revised ‘TurboID, control’ protein list specific to each replicate. We then compared revised protein lists for ‘trained’ and ‘control’ worms from all biological replicates and examined both unique and shared proteins between these two groups. We found 304 proteins that were shared between ‘trained’ and ‘control’ TurboID groups, 706 proteins unique to the ‘TurboID, trained’ group, and 388 proteins unique to the ‘TurboID, control’ group (Figure 1D). We refer to the learning proteome as proteins unique to samples for ‘TurboID, trained’ worms. When generating the learning proteome, we categorised proteins as ‘assigned hits’ based on the criteria that at least one unique peptide was identified by the MASCOT search engine for the protein identity from at least one biological replicate. We also examined peptide sequences in our peak lists that were considered ‘unassigned’ by MASCOT, as these sequences were not detected as unique for any protein by the software, but specific protein identities could be found by performing a Basic Local Alignment Search Tool (BLAST) query (https://blast.ncbi.nlm.nih.gov/Blast.cgi; see Materials and methods for details). The Venn diagram in Figure 1D shows assigned hits only. Learning proteome lists for both assigned and unassigned hits are in Supplementary file 1C and D.

We assessed overlap between biological replicates for individual candidates (Figure 1—figure supplement 4) using two mass spectrometry systems: Thermo-Fisher Q-Exactive Orbitrap (‘QE’) and Orbitrap Exploris (‘Exploris’). Candidates detected in multiple replicates comprised 17% of assigned hits in QE runs, 13% in Exploris, and 21–23% when including unassigned hits (Figure 1—figure supplement 4A–D). Of the 1,010 assigned QE hits, 17% were also identified with Exploris, increasing to 29% when including all 2065 protein identities (Figure 1—figure supplement 4E–F). Despite modest overlap (<25%), key learning-related pathways (Figure 2, Supplementary file 1) and other biological processes, including metabolic pathways (Figure 3), were consistently represented, supporting the biological relevance of the identified learning proteome.

Our initial analysis of the learning proteome sought to validate our TurboID-based approach by identifying components of biological pathways previously implicated in learning. We then performed a gene ontology (GO) term analysis of ‘cellular component’ to obtain a broad overview of the subcellular localisation of proteins identified in trained animals (Figure 4—figure supplement 1A). To do this, we generated protein-protein interaction (PPI) networks of assigned protein hits within the learning proteome for subcellular components of interest (Figure 4—figure supplement 1B–G), using data from STRING and curated with the Cytoscape ClueGO tool (Supplementary file 1E and F; Bindea et al., 2009). We found that a majority of proteins were categorised as ‘cytoplasmic’ (28.1%) as expected from our approach, which utilised the TurboID enzyme not tagged to any bait protein; this means that we would anticipate the enzyme to be present relatively evenly across the cell body and to catalyse biotinylation of proteins in this space. We saw that an unexpectedly high proportion of proteins were nuclear (18.1%), despite the presence of a nuclear export signal in our TurboID transgene, which should prevent TurboID from entering the nucleus and biotinylating nuclear proteins – this could be due to some ‘leaky’ entry of the enzyme or biotin reactive species into the nucleus, or that some proteins are categorised solely as ‘nuclear’ in the ClueGO database when they are also present in other cellular components. Importantly, we found that a proportion of proteins are categorised as present in neuronal compartments – the pre-synapse (0.5%), cilia/dendrites (2.7%), and in the axon/(synaptic) vesicles (4.0%) – as expected from transgenic expression of TurboID in the nervous system.

Learning and memory formation in organisms with brains of varying sizes has been shown to involve key regulatory pathways including signalling via neurotransmitters/neuromodulators, G-protein-coupled receptors (GPCR), the mitogen-activated protein kinase (MAPK) pathway, and the insulin/insulin growth factor-like pathway (Matsumoto et al., 2018; Myhrer, 2003; Rahmani and Chew, 2021). We next categorised proteins within the learning proteome (Figure 2) based on their known roles within these signalling pathways. This included (1) several GPCR components, including the G_i/o protein subunit GOA-1 and G_α protein subunit GPA-2, (2) regulators of insulin signalling including the DAF-2 insulin receptor, phosphoinositide 3-kinase AGE-1, and serine/threonine protein kinases AKT-1 and SGK-1, which were previously reported to modulate salt-based learning in the worm (Tomioka et al., 2006; Sakai et al., 2017), (3) MAPK signalling components including NSY-1/MAPKKK, MEK-2/MAPKK, and MPK-1/MAPK/ERK, (4) cAMP/PKA (protein kinase A) signalling regulators such as the regulatory PKA subunit KIN-2, and (5) multiple components that modulate synaptic vesicle release, including N-ethyl-maleimide sensitive fusion protein NSF1/NSF-1 and syntaxin/SYX-2. In addition, we identified several proteins relevant to glutamate, acetylcholine, and GABAergic signalling. Several components involved in protein synthesis and degradation were also detected in our learning proteome, in line with studies that suggest changes in total protein composition following memory formation (Inberg et al., 2013; Barzilai et al., 1989). These data are summarised in Figure 2, Figure 2—figure supplement 1, and Supplementary file 1F. In summary, the learning proteome includes both known learning regulators and potentially novel candidates that warrant further study. We focused on proteins functioning within these pathways of interest in our subsequent investigations (highlighted nodes in Figure 2, Figure 2—figure supplement 1, and Figure 4—figure supplement 1).

In addition, we consistently observed enrichment of two metabolic pathways, fatty acid metabolism via the TCA cycle and carbohydrate metabolism (gluconeogenesis and glycolysis), in multiple biological replicates of mass spectrometry data, uniquely in TurboID-trained animals (Figure 3). These metabolic pathways play essential roles, including in cellular energy production and macromolecule biosynthesis (Krebs and Johnson, 1937; Goetsch and Lu, 1993). Consequently, their disruption can severely impair animal health. For example, knock-down of mitochondrial components involved in the TCA cycle led to larval arrest and/or severely reduced lifespan in C. elegans (Artan et al., 2022; Liao et al., 2022). This limits the capacity to assess these processes in learning using single-gene mutants or knockdown tools. Therefore, our TurboID approach reveals biological pathways potentially involved in memory formation that are not detectable through conventional forward or reverse genetic screens.

Aside from identifying relevant biological networks, we also used data from the learning proteome to identify potential neuron classes involved in memory formation, using four databases. This included the Wormbase Tissue Enrichment Analysis (TEA) Tool (Angeles-Albores et al., 2016), based on Anatomy Ontology (AO) terms, and single-cell transcriptomics data from the C. elegans Neuronal Gene Expression Network (CeNGEN; Taylor et al., 2021).

Firstly, we employed transcriptome databases to check representation of the nervous system within learning proteome data. The CeNGEN database confirmed that 87–95% of assigned hits and 89–92% of all hits (assigned and unassigned hits) from the learning proteome show neuronal expression, that is were found in at least one neuron in the database (Table 1). It is important to note that CeNGEN was generated using L4 hermaphrodites, and not young adult hermaphrodites (as used here for TurboID), since transcriptomes differ between the two developmental stages (St Ange et al., 2024). The nervous system transcriptome has been characterised for young adult hermaphrodites, highlighting 7873 genes that are enriched in neurons (versus non-neuronal tissues). Neuron-enriched genes were identified using single-nucleus and bulk neuron RNA-Seq techniques, respectively (St Ange et al., 2024; Kaletsky et al., 2016). The nervous system is highly represented by our proteome data; 75–87% of assigned hits and 75–83% of all hits correspond to neuron-enriched genes identified by St. Ange et al. and Kaletsky et al.

Secondly, we assessed which tissues and neuron classes are most highly represented within the learning proteome. We used the Wormbase TEA tool to search for gene lists corresponding to proteins encoded by (1) assigned hits only and (2) both assigned and unassigned hits within the learning proteome. Anatomical terms were considered enriched when they had a q value <1. We observed enriched terms for pharyngeal neurons (M1, M2, M5, NSM, and I4), sensory neurons (PVD), interneurons (ADA and RIG), ventral nerve cord (VNC) motor neurons (VB2, VB3, VB4, VB5, VB6, VB7, VB8, VB9, VB10, and VB11), and CAN cells from both gene lists. RIS interneurons and DD motor neurons were also enriched when including unassigned hits. Several of these neurons have previously been implicated in learning: RIG interneurons (Zhou et al., 2023) and NSM neurons in butanone olfactory learning (Fadda et al., 2020), VNC neurons through changes in glutamate receptor GLR-1 expression during touch habituation (Rose et al., 2003) and diacetyl aversive learning (Vukojevic et al., 2012), and RIS interneurons in salt aversive learning (Wang et al., 2025). Therefore, neurons enriched within the learning proteome include those known to be required for learning; other neurons not previously identified in this context, such as pharyngeal neurons, may warrant further study.

We complemented this analysis by using the CeNGEN database to search for gene lists encoding proteins (assigned hits only, minus non-transgenic controls) identified in control worms (388 genes) versus trained animals (706 genes) from Figure 1D (Taylor et al., 2021). Using the bulk gene search function in CenGEN (threshold = 2), we determined the number of genes from each list that are expressed in a specific neuron type. Values for the trained gene list were normalized to account for the ~1.8-fold increase in the number of proteins detected in trained samples compared to the high-salt control. For each neuron class that appeared in both datasets (128 in total), we calculated fold-change values between the number of genes from trained vs control gene lists. Neurons were ranked in descending order of fold-change. This ranked list is based on the relative enrichment of training-associated genes compared to control, with higher ranks suggesting neurons that may be more transcriptionally responsive or involved during training. These data are summarised in Supplementary file 1G. Given that CeNGEN utilises transcriptomic data from L4 (juvenile) animals, neuron classes were also ranked using equivalent datasets for young adult hermaphrodites: Worm-Seq (Ghaddar et al., 2023) and CeSTAAN (Princeton University, 2025; see Materials and methods for details). Importantly, CeSTAAN and Worm-Seq provide data for 79 and 104 neuron classes, respectively (vs 128 from CeNGEN); this section therefore focuses on CeNGEN data due to its greater coverage, with other datasets described in brackets. Moreover, as this analysis is descriptive and does not include statistical testing (e.g. bootstrapping), the rankings should be interpreted as indicative rather than definitive, and future work incorporating formal statistical approaches will be important to validate these observations.

Cholinergic and glutamatergic neurons constituted 15% and 55% of neurons ranked #1–20, respectively (45% and 30% for CeSTAAN; 40% and 20% for Worm-Seq). Glutamate signalling components previously have been implicated in C. elegans learning paradigms involving salt (e.g. NMDA-type glutamate receptor subunits nmr-1 and nmr-2; Kano et al., 2008). Acetylcholine has not been explored extensively in C. elegans for its involvement in learning but has been described in other animal models and in humans (reviewed in Huang et al., 2022). Other neuron classes identified have previously been implicated in salt-based associative learning (ranks in brackets): AVK interneurons (rank #7 for CenGEN; #37 for CeSTAAN; #76 for Worm-Seq; Beets et al., 2012), RIS interneurons (rank #14 for CenGEN; #1 for CeSTAAN; #31 for Worm-Seq; Wang et al., 2025), salt-sensing neuron ASEL (rank #18 for CenGEN; #16 for CeSTAAN; #34 for Worm-Seq as ‘ASE’; Beets et al., 2012), CEP and ADE dopaminergic sensory neurons (individually ranked #22 and #39, respectively, for CenGEN, vs ‘CEP_ADE_PDE’ ranked #20 and #80 for CeSTAAN and Worm-Seq, respectively; Voglis and Tavernarakis, 2008), and AIB interneurons (rank #21 for CenGEN; #11 for CeSTAAN; #67 for Worm-Seq; Sato et al., 2021). In summary, although there are some exceptions that may be reflecting expression differences between adult and L4 animals, the same neuron classes are generally seen as highly represented for trained animals (vs control) for all three transcriptomic datasets.

Interestingly, unlike its counterpart ASEL (rank #18 for CenGEN), the salt-sensing neuron ASER was ranked only #104/128 (for CenGEN, Supplementary file 1G). ASER becomes activated in response to a decrease in salt concentration (Suzuki et al., 2008) and its downstream targets likely function to redirect worms toward higher salt concentrations (Appleby, 2012). This activation is suppressed after training that reduces attraction to high salt levels (Sato et al., 2021; Wang et al., 2025). It is possible that this learning-dependent suppression of ASER activity may explain its lower fold-change in trained versus control groups. Alternatively, given it is ranked #4 using the CeSTAAN database, potentially due to the use of adults and not L4, it is equally possible that additional proteomic changes may be required in ASER (vs ASEL, with rank #16) to trigger salt chemotaxis changes. Nevertheless, these findings imply a molecular and cellular switch facilitated by dual ASE neurons to express gustatory learning in the worm, which complements previous research.

Some neurons identified were not previously implicated in learning: IL1 polymodal head neuron class (rank #1 for CenGEN; #44 for CeSTAAN; #42 for Worm-Seq), motor neuron DA9 (rank #2 for CenGEN; #78 for CeSTAAN as ‘PDA_AS_DA_DB’; #95 for Worm-Seq as ‘DA_VA’), and interneuron DVC (rank #5 for CenGEN; #23 for CeSTAAN; #3 for Worm-Seq). IL1 releases glutamate (Pereira et al., 2015) and mainly functions in regulating foraging behaviour (Hart et al., 1995), potentially indicating a role in food-based responses. Separately, cholinergic neuron DA9 and glutamatergic neuron DVC are involved in backward locomotion (Pereira et al., 2015; Ardiel and Rankin, 2015; Chalfie et al., 1985). Changes in locomotion are critical for learning-dependent modulation of chemotaxis: the incidence of sharp turns or ‘pirouette’ movements in C. elegans is influenced by prior experience in salt-based gustatory learning (Kunitomo et al., 2013). IL1 may influence salt-based learning by signalling through interneurons AVE (rank #24 for CenGEN; #10 for CeSTAAN; #87 for Worm-Seq) and PVR (rank #114 for CenGEN; neuron class not available in CeSTAAN; #44 for Worm-Seq) to the DA neurons (Bhatla, 2009), potentially modulating backward locomotion as part of the chemotaxis response. We also identified pharyngeal neurons I3 (rank #4 for CenGEN; data not available in CeSTAAN nor Worm-Seq) and I6 (rank #5; neuron class not available in CeSTAAN nor Worm-Seq), which have not previously been implicated in learning. Figure 4D provides a summary for the neural circuits implicated from these analyses, where neuron classes are highly connected to each other. Investigating the role of specific genes within these circuits opens new avenues for future research into gustatory learning.

Our initial analysis of learning proteome data indicates that there are multiple hits present in biological pathways important for neuron function, and that are potentially relevant to learning and memory formation. To test this directly, we performed salt associative learning experiments on selected learning proteome hits (Figures 4—6, Figure 6—figure supplements 1 and 2). We used the following general rules to interpret our data: if the average chemotaxis indices (CIs) for ‘trained’ worms were higher in a particular strain compared with wild-type, this strain was considered learning-defective, as this reflects a reduced magnitude of the expected behaviour change (an increased preference for low salt demonstrated by CIs closer to 0 or negative CIs). If the average CI for ‘trained’ worms was lower in a strain compared with wild-type, then this strain was considered to display ‘better’ learning, as the lower CI reflects an increased magnitude of the expected behaviour change. In general, we observed no significant difference in CIs between naïve groups for all genotypes, reflecting no gross locomotor or chemotaxis defects in the strains tested (Figures 4—6, Figure 6—figure supplements 1–2).

We tested 26 candidates in total for this study. Although this represents a small subset of the 706 proteins identified in the learning proteome, several proteins in the full list are unsuitable for functional testing due to key constraints: (1) having essential roles, with corresponding single-gene mutants being lethal; (2) involvement in neurodevelopment rather than mature neuronal function; and (3) being required for locomotion, with severe locomotion defects precluding assessment using chemotaxis assays.

Candidates tested were classified as either strong (detected in biological replicates ≥3) or weak (replicates <3) based on the number of mass spectrometry replicates in which they were uniquely identified in TurboID-trained C. elegans (shown in brackets). We determined these numbers by considering both assigned and unassigned protein lists, which contained mostly neuron-expressed proteins (Table 1) including known learning regulators (Figure 2 and Supplementary file 1H). The list of 26 candidates for further testing includes both weak and strong hits. In addition, although candidates tested were mostly detected in more replicates of trained versus control groups, we also assayed seven candidates for which this was not the case. Table 2 summarises the potential learning regulators explored in this study, including strong/weak classifications and replicate numbers between experimental groups.

We first tested the regulatory subunit of PKA, kin-2 (1 replicate), since it is a known regulator of memory and was detected as a weak candidate by TurboID. Adenylyl cyclase is a key signalling effector for Gα_s and Gα_i proteins and regulates levels of the secondary messenger cyclic AMP (cAMP) within the cell. cAMP binding to PKA regulates its activity, and therefore its downstream effects (Sassone-Corsi, 2012). We tested worms with the ce179 mutant allele in kin-2, in which a conserved residue in the inhibitory domain (which normally functions to keep PKA turned off in the absence of cAMP) is mutated to cause an R92C amino acid change – this results in increased PKA activity (Schade et al., 2005). kin-2 has previously been shown to be required for intermediate-term memory in C. elegans (Stein and Murphy, 2014), with cAMP/PKA signalling previously shown to be involved in memory in multiple systems (Kandel, 2012). We found that these kin-2 mutant animals showed enhanced learning compared with wild-type (i.e. Non-Tg worms; Figure 5A). We next re-expressed KIN-2(R92C) in wild-type worms using a pan-neuronal promoter, and these worms showed a similar phenotype to kin-2(ce179) worms, with enhanced learning compared with non-transgenic siblings (Figure 5B). These data suggest that increased PKA activity in the nervous system drives salt associative learning.

We next assessed two strong candidates not previously assessed for their role in learning: putative arginine kinase F46H5.3 (four replicates) or armadillo-domain containing protein C30G12.6 (three replicates). Unlike kin-2(ce179) worms, neither single gene mutant obtained from the Caenorhabditis Genetics Center had been backcrossed. We backcrossed mutant strains four times to N2 and tested both non-backcrossed and backcrossed versions. An improved learning phenotype was displayed by both non-backcrossed and backcrossed F46H5.3(-) worms (Figure 6A and B). In contrast, we found that non-backcrossed C30G12.6(-) animals displayed an enhanced learning phenotype, whereas backcrossed C30G12.6(-) mutants behaved like wild-type (Figure 6C and D). This suggests that the non-backcrossed C30G12.6(-) strain contains a background mutation that impacts learning capacity, a potential avenue for future work.

F46H5.3 is a homolog for creatine kinase B (cytoplasmic) and mitochondrial creatine kinase 1B, but is considered an arginine kinase since C. elegans use arginine instead of creatine. Both creatine kinases are essential for energy metabolism via ATP modulation. Given notable representation of metabolic pathways from the learning proteome identified here (e.g. Figure 3), this provided additional rationale for testing F46H5.3. Moreover, the depletion of creatine kinase B reportedly increases the latency needed for memory encoding of a spatial learning task in mice (Jost et al., 2002). Since F46H5.3 is expressed in most neurons in the worm (Taylor et al., 2021), this putative arginine kinase may affect learning capacity through modulating ATP levels in neurons.

We also tested genetic mutants for candidates involved in neurotransmission, a key function for all neurons and a requirement for learning (reviewed in Myhrer, 2003 and Rahmani and Chew, 2021). Two strong candidates, cationic acetylcholine-gated calcium ion channel ACR-2 (four replicates) and cholinergic receptor interactor ELP-1 (three replicates), were not observed to regulate learning (Figure 6—figure supplement 1A and B). Interestingly, several acetylcholine receptors that were weak candidates showed differences in learning compared with wild-type controls. Those tested include ACC-1 (one replicate), ACC-3 (2 replicates), and LGC-46 (1 replicate) in the subfamily of anionic ‘ACC’ acetylcholine-gated ligand-gated ion channels (Morud et al., 2021). Interestingly, lgc-46 mutants demonstrate a reduced capacity for learning compared with wild-type, whereas acc-1 and acc-3 mutants appear to display better learning (Figure 4A, B and C). Although these ACC receptors are activated by acetylcholine (Park et al., 2000; Putrenko et al., 2005), or have been reported to possess the same protein domain/s as known acetylcholine receptors (Takayanagi-Kiya et al., 2016), they vary substantially in expression pattern in the nervous system (Taylor et al., 2021). We postulate that acetylcholine signalling in specific neurons may therefore contribute to learning in different directions, an interesting avenue for future research. We also tested other genes involved in neurotransmission, such as gbb-2 (for GABAergic signalling, 2 replicates) and glr-1 (for glutamatergic signalling, 1 replicate), neither showed a detectable change in learning capacity compared with wild-type (Figure 6—figure supplement 2A and B). maco-1 (1 replicate) encodes a macoilin family protein that functions broadly in neurotransmission (Arellano-Carbajal et al., 2011), and has previously been shown to regulate memory of olfactory adaptation in worms (Kitazono et al., 2017). The maco-1 mutant tested in our study (nj21) did not show a locomotor defect but also did not show any obvious learning phenotypes (Figure 6—figure supplement 2C). Our data, therefore, indicate that specific components of cholinergic signalling are required for salt associative learning.

We tested several components of GPCR signalling that were identified as learning proteome hits, including gap-1 (one replicate), gap-2 (three replicates), gpa-2 (two replicates), and rho-1 (two replicates; Figure 6—figure supplements 1C, D and 2E, F), but found that worms with single mutations for these genes, when compared to wild-type controls, did not show statistically significant differences in learning capacity. We postulate that although GPCR signalling is broadly important for learning (reviewed in Jong et al., 2018 and Rahmani and Chew, 2021), there may be high levels of redundancy built within this pathway such that single pathway components can be compensated for by other functionally similar genes.

Several components of MAPK signalling have been shown to be involved in different forms of learning (reviewed in Peng et al., 2010; Ryu and Lee, 2016), including NSY-1/MAPKKK, MEK-2/MAPKK, and MPK-1/ERK that were identified as part of the learning proteome in our study (Ohno et al., 2014). We tested the dual leucine zipper MAPKKK-encoding gene dlk-1 (two replicates), a mutant of this gene showed no difference in learning capacity compared to wild-type (Figure 6—figure supplement 2G). The E2 ubiquitin-conjugating enzyme variant UEV-3 (four replicates) has been shown to be a member of the DLK-1 pathway and a potential interactor of p38/MAPK PMK-3 (Trujillo et al., 2010); trained phenotypes between wild-type and uev-3 mutants were not statistically significant (Figure 6—figure supplement 1D). We also tested worms with a mutation in fsn-1 (three replicates), proposed to attenuate synapse growth in a DLK-1-dependent manner (Hung et al., 2013). These animals displayed learning capacity similar to wild-type animals (Figure 6—figure supplement 1E). In summary, these data indicate that mutating single components of the MAPK pathway does not generally perturb salt associative learning.

Finally, we assessed additional hits that do not fit in the pathways above but were considered strong candidates. These include neuronal adhesion/IGCAM gene rig-4 (three replicates) or putative guanyl-nucleotide exchange factor genes (aex-3 or tag-52, both three replicates). Aversive associative learning in worms and mice both relies on IGCAM gene ncam-1/NCAM1 (Cremer et al., 1994; Vukojevic et al., 2020; Doyle et al., 1992). Additionally, guanine nucleotide exchange factors ‘ArhGEF4’ and ‘RapGEF2’ in mice (Jiang et al., 2024; Yoo et al., 2020), as well as unc-73 in C. elegans (Arey et al., 2018), have been linked to learning and memory previously. We found that aex-3(-), rig-4(-), and tag-52(-) single mutants did not show significant differences in salt associative learning compared to wild-type controls (Figure 6—figure supplement 1F, G and H). We also tested elo-6 (three replicates), which encodes a long chain fatty acid elongase that potentially functions together with another elongase encoded by elo-5, although only elo-6 is expressed in neurons (Kniazeva et al., 2004). Fatty acid composition has previously been demonstrated to be important for learning and memory (Wallis et al., 2021; Pershina et al., 2022; Akefe et al., 2024). Our data showed no significant learning defect or improvement in elo-6 mutant animals (Figure 6—figure supplement 1I), although, as mentioned, its role may be masked by functional redundancy.

Other strong candidates tested that did not show a learning phenotype include: (1) ift-139 (5 replicates), a ciliogenesis gene that was explored since hippocampal cilia structures have been seen to be important for memory in mice (Niwa, 2016; Jovasevic et al., 2021; Berbari et al., 2014). (2) tap-1 (three replicates), which encodes an ortholog for TGF- β activated kinase 1 (Meneghini et al., 1999), (3) SAEG-1 (four replicates), a suppressor for protein kinase G (PKG) ortholog EGL-4 activity, which has been implicated in behavioural changes induced by odour-sensory fatigue L’Etoile et al., 2002; cGMP-dependent kinase PKG also promotes long-term memory in rodents (Ota et al., 2008; Paul et al., 2008), and (4) VER-3 (three replicates), which encodes a predicted vascular endothelial growth factor (VEGF) receptor-like protein (Popovici et al., 2002). VEGF/VEGFR was reported to be upregulated in rats following spatial learning (Cao et al., 2004). These data are shown in Figure 6—figure supplement 1.

There are several potential reasons why many mutants tested did not display a learning phenotype. Firstly, as mentioned above, effects may be masked by redundant or compensatory pathways. For example, IGCAM genes in the worm have been reported to act redundantly in axon navigation, including rig-4 (Schwarz et al., 2009). It is also possible that these mutations do not fully knock out protein function. We generally assessed animals with deletion mutations predicted to disrupt protein function, but we did not confirm this through qualitative or quantitative means. Additionally, many mutants used here were not backcrossed as it was beyond the scope of this study, so these lines may have background mutations masking learning phenotypes of the mutations of interest. This was seen for C30G12.6(-) animals in this study, where an enhanced learning phenotype in non-backcrossed worms was lost after backcrossing (Figure 6). Separately, although memory encoding (learning) and retention are interlinked biological processes, they are molecularly distinct (Arey et al., 2018; Kauffman et al., 2010; Watteyne et al., 2020; Rahmani and Chew, 2021). Typically, learning is assayed for C. elegans immediately after training, whereas memory retention is assessed after a post-training rest period – this assessment usually relies on a behavioural change representing learning/memory formation. It is possible that some candidates detected in neurons during learning may correspond to those important for memory retention and not encoding, so it may be worth utilising this proteomic dataset as a resource to explore this concept in future. Finally, the population-wide chemotaxis assays we perform here to validate candidates may not be sensitive enough to capture subtle potential behavioural differences caused by these mutations. Pirouette and weathervane behaviours in C. elegans change based on previously experienced salt concentrations in the presence of food (Kunitomo et al., 2013). These behaviours can be measured through more in-depth investigation of locomotor behaviour through live tracking and analysis, providing a more sensitive measure for learning responses compared to the chemotaxis assays used here. These factors are important considerations for future experiments utilising the learning proteome as rationale to assess novel mechanisms in learning.

Next, we explored whether the learning regulators identified from our learning proteome functioned more broadly in other types of learning. To do this, we assayed salt aversive learning capacity (training with aversive cue i.e. starvation) for candidates seen to affect salt associative learning (training with an appetitive cue i.e. presence of food): PKA regulatory protein KIN-2 (Figure 5), arginine kinase F46H5.3 (Figure 6), and acetylcholine receptor subunits ACC-1, ACC3, and LGC-46 (Figure 4). In the salt associative learning assay used thus far, ‘trained’ worms are exposed to a pairing of no salt + food (‘control’ worms with high salt +food), whereas in salt aversive learning, ‘conditioned’ worms are exposed to high salt + no food (and control ‘mock-conditioned’ worms with no salt + no food). In the salt aversive learning assay, ‘conditioned’ worms therefore learn to avoid high salt (as it is associated with starvation, a strongly negative cue), whereas ‘mock-conditioned’ worms and naïve worms retain a preference for high salt (Nagashima et al., 2019; Hiroki et al., 2022). We found that only F46H5.3(-) mutant worms showed a significant change in learning capacity for salt aversive learning compared with wild-type (Figure 6—figure supplement 3). Specifically, F46H5.3(-) mutants displayed a larger decrease in CI in trained animals compared to wild-type trained worms, demonstrating a potential learning improvement (Figure 6—figure supplement 3A). F46H5.3(-) mutant phenotype also showed enhanced learning for the salt associative learning paradigm (Figure 6A and B). Although kin-2(ce179) mutants were not shown to impact salt aversive learning, they have been reported previously to display impaired intermediate-term memory (but intact learning and short-term memory) for butanone appetitive learning (Stein and Murphy, 2014). These findings therefore suggest (1) a generalised effect for F46H5.3 in gustatory learning paradigms involving salt, (2) a specific role for KIN-2 in appetitive learning paradigms, and (3) a unique effect for ACC receptors ACC-1/2 and LGC-46 in salt associative learning only.

Our TurboID approach provides a unique benefit as a systems-based tool, in that it can be used to map individual candidates onto broader molecular networks. F46H5.3 is mostly uncharacterised beyond its predicted homology to creatine kinase B and mitochondrial creatine kinase 1B. As its role within learning pathways is unknown, we tested for protein-protein interactions between F46H5.3 and other candidates in the learning proteome using the software STRING (version 12.0; Szklarczyk et al., 2023), aiming to infer its potential function within a molecular network. Mitochondrial creatine kinases regulate phosphocreatine synthesis using ATP, which requires calcium influx into mitochondria to induce ATP synthesis (reviewed in Schlattner et al., 2001; Schlattner et al., 2006). Voltage-dependent anion channel VDAC-1 (identified in three replicates of learning proteome data) plays a critical role in calcium homeostasis in C. elegans mitochondria (Shoshan-barmatz et al., 2017) and is predicted interactor of F46H5.3. Moreover, calcium influx into the mitochondria is regulated by calcium/calmodulin kinase II (Nguyen et al., 2018), an established and highly conserved regulator of learning (reviewed in Ataei et al., 2015; Zalcman et al., 2018). The sole calcium/calmodulin kinase II in C. elegans (UNC-43, two replicates) is predicted to interact with VDAC-1. Our learning proteome also includes proteins involved in calcium/calmodulin complex formation (e.g. calmodulin/CMD-1 in four replicates and cyclic nucleotide–gated ion channel TAX-4 in one replicate; Karabinos et al., 2003; Komatsu et al., 1999). Calcium/calmodulin complexes can also modulate cAMP levels (reviewed in Sharma and Kalra, 1994), which influences PKA/KIN-1 activity regulation by KIN-2 and A-kinase anchoring protein AKAP-1 (two replicates; reviewed in Sadeghian et al., 2022). Therefore, cytoplasmic calcium homeostasis is one potential pathway through which both KIN-2 and F46H5.3, validated in our study as learning regulators, modulate learning and memory (Figure 7).

Finally, we can combine the protein network analysis with analysis of neuron representation within the learning proteome, as described above and shown in Figure 4D. There are two neuron classes that express all five candidates shown here to affect salt associative learning (KIN-2, F46H5.3, ACC-1, ACC-3, LGC-46): RIM interneurons and DB motor neurons (Taylor et al., 2021; threshold = 2). These neurons are involved in reversals and forward locomotion, respectively (Guo et al., 2009; Chalfie et al., 1985). It is possible that these learning regulators influence experience-based behaviour through modulating the function of these neurons to alter chemotaxis responses in the presence of gustatory cues.

KIN-2 and F46H5.3 share the same expression pattern in many neuron classes, whereas neurons expressing the three ACC receptors are more diverse (Taylor et al., 2021; Figure 4D). ACC-3 is expressed in salt-sensing ASE, mechanosensory and dopaminergic CEP neurons, polymodal sensory neuron IL1, and two interneuron classes (ADA and RIM) (Taylor et al., 2021). ADA’s function is not well-characterised, but it is predicted to be involved in chemosensation (Sohn et al., 2011). In contrast, ACC-1 and LGC-46 are expressed in several interneurons and motor neurons including those implicated in gustatory or olfactory learning paradigms (AIB, AVK, NSM, RIG, and RIS; Beets et al., 2012; Fadda et al., 2020; Wang et al., 2025; Zhou et al., 2023; Sato et al., 2021) and important for backward or forward locomotion (AVE, DA, DB, and VB; Chalfie et al., 1985). There are also highly represented neuron classes which are not as well defined (ADA, I4, M1, M2, and M5), which may present interesting directions for future research. Cholinergic signalling may therefore regulate gustatory learning through integration of sensory signals as well as direct modulation of the motor circuit. In contrast, KIN-2 and F46H5.3 may play more general functions within the nervous system, such as through modulating calcium homeostasis (Figure 7). As mentioned above, Figure 4D utilises the CeNGEN database generated from L4 animals, so we cross-referenced this expression data with transcriptome studies using young adult hermaphrodites (as in this study). Our five candidates of interest (ACC-1/3, F46H5.3, KIN-2, and LGC-46) were reported in the same neuron classes as in CeNGEN, further suggesting that these neurons may be important for memory encoding in the worm (Ghaddar et al., 2023; Kaletsky et al., 2016; Roux et al., 2023; Smith et al., 2024; St Ange et al., 2024). Overall, our TurboID dataset offers a valuable foundation for future investigations into individual proteins involved in learning and provides a resource for systemic analyses at the level of tissues, neuron types, subcellular localisations, and molecular networks.

Young adult (day 1) hermaphrodite C. elegans were grown using standard conditions on nematode growth medium (NGM) agar in petri dishes at 22 °C for all experiments (Brenner, 1974). This was done for at least two generations for salt associative learning assays involving TurboID, as well as assays involving salt aversive learning involving kin-2(ce179) mutants. Animals were otherwise cultured at 22 °C for one generation and at 15 °C prior to this. For TurboID-based labelling experiments, worms were cultured for at least two generations with the biotin-auxotrophic strain as their food source (Branon et al., 2018). For all other experiments, animals were fed Escherichia coli (E. coli) strain OP50. C. elegans lines used in this paper are listed in Supplementary file 1A. Information about plasmids used to generate lines YLC207 and YLC369 is provided as Source Data.

The biotin treatment strategy was adapted from Artan et al., 2021. Briefly, a solution of 100 mM biotin (in 250 mM KOH, 5 mM K₃PO₄ (pH 6.0), 1 mM CaCl₂, and 1 mM MgSO₄) was diluted 1:100 in E. coli MG1655 bioB::kan washed with modified Luria Broth (LB; 25 mM NaCl, 5 mM K₃PO₄ (pH 6.0), 1 mM CaCl₂, 1 mM MgSO₄, 1.0% (w/v) Bacto Tryptone, 0.5% (w/v) yeast extract, 0.05 mg/mL Kanamycin). Biotin-depleted worms were fed E. coli MG1655 on NGM agar during a 6 hr conditioning period (see the ‘Salt associative learning’ section). The bacterial pH was increased by a negligible amount (i.e. 0.1) from the addition of KOH; thus, it is expected it will not significantly impact worm physiology (Khanna et al., 1997; Cong et al., 2020). At least 3000 worms per group were utilised for downstream proteomic experiments.

The following protocol was adapted from Liang et al., 2014 and Artan et al., 2021: After biotin treatment (end of learning assay), worms were washed twice using ‘washing buffer’ (50 mM NaCl, 5 mM K₃PO₄ (pH 6.0), 1 mM CaCl₂, 1 mM MgSO₄) and then stored as pellets in a –80 °C freezer. Each pellet was suspended in 200 μL Radioimmunoprecipitation assay (RIPA) buffer containing 2 M urea (prepared as in Sanchez and Feldman, 2021), as well as 150 mM NaCl, 50 mM Tris-Cl (pH 8.0), 5 mM EDTA, 10 mM NaF, 2 mM Na₃VO₄, 1 mM NaPP, 1% (v/v) Nonidet-P40, 1% (w/v) SDS, 0.5% (w/v) sodium deoxycholate, 0.1% (w/v) β-glycerophosphate, and 1×cOmplete Mini Protease Inhibitor (Merck). Worm pellets were sonicated (10×4 s total sonication time, 2 s ‘on’ pulse, 3 s ‘off’ pulse’, and 25% amplitude) with the Q125 Sonicator (Q Sonica) in a temperature-controlled room (~2–8°C) and allowed to rest on ice for >20 s between each sonication. All samples were vortexed for ~5 s at room temperature, and then centrifuged (14,000 rcf, 4 °C, 10 min) to separate carcasses/debris from supernatant containing proteins. The supernatant was then used for protein quantification by BCA assay (Thermo Fisher Scientific, #23225).

20–40 µL protein samples, containing 1×sample buffer (0.25 M Tris (pH 6.8), 10% (v/v) β-mercaptoethanol, 10% (v/v) glycerol, 10% (w/v) SDS, 0.25% (w/v) bromophenol blue), were boiled at 95 °C for 3 min. Proteins were electrophoresed through 8% (v/v) polyacrylamide gels for 40–60 min at 80 V, before the voltage was increased to 100 V for an additional 40–60 min, and then increased to 120 V until completion of electrophoresis (Jeong et al., 2018; Stefanoska et al., 2022). The standard protocol was utilised for semi-dry protein transfer (constant 25 V, 40–45 min) onto nitrocellulose membranes and western blotting, using the following antibodies/probes (dilution in brackets, all made in 5% (w/v) bovine serum albumin): rabbit anti-α tubulin (1:1000; Abcam, #ab4074), mouse anti-V5 (1:1000; Cell Signalling Technology, #80076), goat anti-rabbit HRP (1:20,000; Cell Signalling Technology, #7074P2), goat anti-mouse HRP (1:20,000; Thermo Fisher Scientific, #G-21040), and streptavidin-HRP (1:5000; Cell Signalling Technology, #3999; Jeong et al., 2018; Stefanoska et al., 2022). Probed proteins were visualised by chemiluminescence using Clarity Western ECL Substrate (Bio-Rad, #1705060) according to manufacturer’s instructions.

We adapted previously established methods to perform two behavioural paradigms that model associative learning: (1) salt associative learning (Tang et al., 2023; Hiroki et al., 2022; Nagashima et al., 2019), and (2) salt aversive learning (Lim et al., 2018).

Analyses utilised protein lists (containing assigned hits only) from mass spectrometry experiments for the following groups: (1) non-Tg high-salt control, (2) non-Tg trained, (3) TurboID high-salt control, and (4) TurboID high-salt trained (Figure 1A). Briefly, protein identities from non-Tg high-salt control and non-Tg trained were subtracted from TurboID high-salt control and TurboID trained, respectively. Following subtractions, protein lists were then compared to identify those unique to TurboID high-salt control (388 proteins) versus TurboID trained (706 proteins) as in Figure 1D.

For behavioural experiments, we performed three to five biological replicates for most genotypes, consistent with similar high-quality studies (Kitazono et al., 2017; Lim et al., 2018; Sakai et al., 2017; Stein and Murphy, 2014). For key candidates (KIN-2, F46H5.3, ACC-1, ACC-3, LGC-46), 5 biological replicates were performed. The number of biological replicates is indicated in each figure legend; each biological replicate comprised three technical replicates. Randomisation was not applied because experimental groups were defined by genotype or condition. Quantification of chemotaxis assays was conducted without blinding to genotype or condition. Exclusion criteria were pre-determined: Groups were excluded only if bacterial contamination was evident or if fewer than 20 individual animals were present in a technical replicate.

Statistical analyses were performed in GraphPad Prism (version 8.0). The Shapiro-Wilk normality test was used to assess chemotaxis assay data. Following confirmation of normality, an ordinary two-way ANOVA with Tukey’s multiple comparisons post-test (α=0.05) was performed to compare differences between mean CI values for each group. This statistical analysis was chosen given that it aligns with recent publications that employ similar experimental designs and data structures (Beets et al., 2020; Jang et al., 2019; Kitazono et al., 2017; Lim et al., 2018; Lin et al., 2010). Exact p-values for each statistical comparison are reported in Supplementary file 1I.