ELM—the Eukaryotic Linear Motif resource—2024 update

1Many eukaryotic proteins are highly modular and have numerous structural and functional components. Multiple globular domains may be separated by substantial intrinsically disordered regions (IDRs). Globular domains provide many functions, such as catalysis and high specificity macromolecular binding. IDR functions can be as simple as acting as linkers spacing out the folded domains. However, IDRs are also heavily involved in regulatory functions in all major signalling pathways, while often being modified by phosphorylation and other post-translational modifications (PTMs) acting as signals/switches, for example, of cell state. The smallest functional modules in eukaryotic proteins are the short linear motifs (SLiMs), typically 3–15 amino acids in length and being present primarily within IDRs and thus simultaneously providing flexibility and accessibility for interactions. When interacting with their partner protein domains, SLiMs are likely to acquire transiently stable conformations by the mechanisms of conformational selection or induced fit to their partner's interaction surface. Short linear motifs are considered to be by far the most abundant class of protein modules, with the corollary that the functions of most SLiM instances are yet to be elucidated.

2SLiMs were first defined in 1990 by Tim Hunt, based on a few examples of subcellular localisation targeting signals, such as the KDEL endoplasmic reticulum (ER) retention motif, the positively charged nuclear targeting sequence and peroxisomal targeting sequences. Over time, it has become clear that essentially all aspects of cell biology involve SLiM interactions. For example, they have key roles in cell cycle, vesicle trafficking, cytoskeleton dynamics, innate immunity and protein and RNA degradation systems. SLiMs bind partner protein domains with moderate affinity and therefore in many contexts they exhibit cooperativity to achieve their functional output. In this regard, multivalent SLiM-domain interactions appear to be important in liquid-liquid phase separation (LLPS) processes.

3Many PTMs on SLiM targets are catalysed by amino acid residue-modifying enzymes such as protein kinases. For heavily researched kinases, the sequence specificities have been determined at the substrate sites, often aided by the use of SPOT array phosphorylation assays. However, for most of the > 530 human kinases, there has been little or no attempt to establish their phosphorylation site motifs. As a result, annotation of kinase specificities in ELM has been hampered. Nonetheless, a recent systematic attempt to provide SPOT arrays for the serine/threonine human kinome has been carried out by Johnson et al.. This will allow this important class of SLiMs to be better defined in future ELM updates.

4The abundant use of SLiMs in cell regulation has a downside because they can be mimicked by intracellular pathogens, enabling the hijacking of cellular systems and thereby repurposing the host cell for pathogen reproduction. Many SLiMs such as the nuclear targeting sequence and the LxCxE-binding motif were first identified in viral proteins, with cellular counterparts only being discovered later.

5Much of the experimental research on SLiM functions is focused on single protein instances. Structure determination and various types of in vitro peptide interaction/modification assays are helpful in establishing the amino acid preferences of the motifs. However, these approaches do not scale well. Therefore, there is growing interest in the deployment of more systematic approaches such as Phage Display, Holdup Assays and large-scale and quantitative screening of linear motif binding specificity, recently reviewed by Davey et al.. These methods enable the discovery of large numbers of candidate SLiMs in the intrinsically disordered proteomes of hosts and pathogens that can be followed up with functional validation methods.

6In this paper, we summarise the current status of the ELM resource in 2024 and highlight areas of data growth since the last report in 2022.

7The ELM resource provides an intensive platform to catalogue and explore the intricacies of SLiMs in proteins. The motif knowledge in the ELM database (ELM DB) is classified into six broad categories: cleavage (CLV), degradation (DEG), docking (DOC), ligand (LIG), modification (MOD) and targeting (TRG). Within themselves, each of these categories comprises high-level functionally similar motif classes/entries.

8The core of each motif entry is built from a detailed compilation of motif instances, which are rigorously extracted from existing scientific studies. In the ELM DB, these instances hold fine-grained details of annotated evidence, including methods used to characterise the motif as well as the curator's inference on the reliability of annotated data. The instances are also analysed from various aspects, such as sequence, structure, function, localisation, evolution and interaction context, to define the primary amino acid residues contributing to the specificity and binding strength of the motif-mediated protein complex. This systematic analysis by the ELM curators enables defining the motif sequence pattern in a standard POSIX regular expression format (detailed in Wikipedia (https://en.wikipedia.org/wiki/Regular_expression) and the ELM website for an in-depth view of their consensus definitions). A motif class also includes local sequence context and molecular interaction features from the core motif positions. Moreover, each class also assimilates critical insights into the biological role as well as the contextual basis for supporting the presence of a motif in the given protein sequence. Overall, every class distributes several key biological insights for the motif such as its functional attributes, cellular context, interacting domain knowledge, cellular locations, and curated knowledge from the scientific literature.

9ELM curations are systematically organised in the PostgreSQL relational database (http://www.postgresql.org/) backend and can be updated as new knowledge or findings become available for any existing motif entry. The ELM DB can be queried through the Django web framework (https://djangoproject.com/), and output is presented via its front-end interface. The ELM resource also provides access to a toolkit for motif exploration, allowing users to identify potential motifs in proteins of their interest. Once the query sequence matches with the known motif patterns in the ELM DB, they are displayed graphically on the protein sequence. The motif matches are overlaid with contextual features, which include accessibility and conservation information, among other functional insights.

10The curated knowledge within the ELM repository can be accessed and downloaded for free in various formats. Detailed specifications on the available formats and datasets are available at http://elm.eu.org/downloads.html. Furthermore, a REST-API is available to facilitate automated queries of ELM class consensus matches against proteins of interest. The access details for utilising the motif search REST-API can be found at http://elm.eu.org/api/manual.html.

11Since its 2022 update, ELM curation has primarily focused on motifs pivotal to RNA and protein degradation, innate immunity, kinase biology, synapse systems, the cell cycle and pathogen-host interactions, among other motif-governed areas. With this update, ELM catalogues 356 motif classes (Figure 1; Table 1)—an addition of 39 since the last release (Table 2)—with a total of 4283 annotated motif instances, resulting in a net increase of 346 instances. Additionally, 17 existing motif classes and 111 instances have been updated based on recent research findings (Table 2). Major advances in the updated motif classes were focused on improvements of the regular expressions defining the motif specificity and on the inclusion of structural and affinity data for each motif. ELM now maintains links to 4274 scientific publications. The growth in protein structure information in ELM has reached a new high and with this update 102 structures have been integrated, bringing the total cross-referenced structures with databases such as PDBe and RCSB-PDB to 718. Additionally, the ELM dataset now defines 2749 motif-partner interactions, of which 688 have curated binding affinities. In-depth insights into ELM’s data are provided in Table 1.

12In eukaryotic cells, numerous RNA processing and modification events are required to produce RNAs in their final functional forms. To remove defective, unstable or unwanted transcripts from the nucleus, RNAs must first be identified and then degraded by the nuclear exosome, a multi-component protein complex of RNA processing enzymes.

13The RNA helicase Mtr4 is an important cofactor of the nuclear exosome. Mtr4 is not just a helicase but rather a multi-domain, modular hub protein that contributes to the processing of ribosomal RNAs and the targeting of the exosome to different degradable RNA species. Mtr4 is part of the TRAMP complex, together with the Trf4 poly(A)polymerase and the Air2 RNA-binding adaptor protein. Most of the hitherto discovered SLiMs associated with nuclear RNA surveillance are either mediating the inter-subunit interactions of the TRAMP complex or facilitating the recruitment of the complex by diverse nuclear exosome cofactors (Figure 2).

14A highly conserved motif within the N-terminal tail of yeast Trf4 interacts with the helicase core of Mtr4 (see the LIG_Mtr4_Trf4_1 motif class;). This interaction is conserved in vertebrates, although with slightly different motif sequence preferences (see LIG_Mtr4_Trf4_2). Within the N-terminal tail of Air2 there are two distinct Mtr4-interacting motifs, one interacts with the helicase core (LIG_Mtr4_Air2_1), while the other can bind to the KOW domain within the arch of Mtr4 (LIG_Mtr4_AIM_1). This newly added, so-called arch-interacting motif (AIM; LIG_Mtr4_AIM_1) is particularly important, because many nuclear exosome cofactors, such as Nop53, Utp18 and Sqs1 also employ this motif for binding to and recruiting Mtr4 and the exosome to specific RNA transcripts to be degraded. Furthermore, this mode of interaction is also conserved in vertebrates. In the TRAMP complex, the Air proteins, Air1 or Air2, interact with Trf4 through a conserved motif within the linker between their fourth and fifth zinc knuckles (LIG_Trf4_IWRxY_1) plus their fifth zinc knuckle itself. Trf4 also contributes to the targeting of complexes for the degradation of incorrectly folded non-coding RNAs. The C-terminal Nrd1-interacting motif (NIM) of Trf4 binds to the NNS (Nrd1-Nab3-Sen1) complex subunit Nrd1. This NIM motif class (LIG_Nrd1CID_NIM_1) has now been extensively revised, since the NNS subunit Sen1 and exosome-associated RNA-binding protein Mpp6 also turned out to employ NIM motifs for the binding of Nrd1.

15Besides the TRAMP complex, additional RNA processing complexes have been characterised. In the yeast S. pombe, the MTREC (Mtl1–Red1 core) complex is responsible for directing the degradation of meiotic, non-coding and unspliced RNAs. Mtl1 (Mtr4-like protein 1) is a homolog of Mtr4. Red1 protein interacts with Mtr1 and additional proteins in the MTREC complex, including zinc-finger containing protein Ars2. Red1 contains an EDGEI motif (see newly annotated class LIG_ARS2_EDGEI_1) that is responsible for its interaction with Ars2.

16In humans, there are a myriad of binding partners that interact with the human homolog of yeast Ars2, the Serrate RNA effector molecule homolog protein, through one or more EDGEI motifs. The binding of these partners facilitates different RNA processing activities. Interestingly, there are examples of EDGEI-mediated interactions even in A. thaliana and D. melanogaster; therefore they seem to be highly conserved across evolution.

17Protein quality control mechanisms are of quintessential importance to all living cells and their representation is becoming increasingly important for the ELM resource. Targeted degradation of proteins can remove unwanted, mis-translated polypeptides, improperly localised, fragmented or misfolded proteins, proteins from viral intruders, and many more (Figure 3). Protein degradation within the eukaryotic cytoplasm and nucleus (and even of proteins translocated from the endoplasmic reticulum) is performed by the ubiquitin-proteasome system. Ubiquitination is orchestrated by at least three families of enzymes, E1 ubiquitin activating and E2 and E3 ubiquitin conjugating enzymes, which act sequentially, in distinct complexes to add lysine 48 (K48)-linked polyubiquitin chains to the target protein. E3 ligases recognise their target proteins through molecular features called degrons. The majority of known degrons are SLiMs, and many of them are located at the protein amino (N-degrons) or carboxy (C-degrons) terminus. Most N‐terminal degrons are well known (explaining the so-called ‘N-end-rule’ determining protein half-life in eukaryotic cells). However, C-degron pathways that selectively degrade proteins depending on the amino acids preceding their C-termini, also referred to as DesCEND (Destruction via C‐End Degron) pathways, have also been identified in recent years. The existence of C-terminal degrons explains why the distribution of C-terminal amino acids is unevenly biased in the proteome, as protein evolution actively selects against destabilising residues unless functionally required. So far, terminal Gly, Ala, Arg and Glu residues (‘destabilising residues’) are known to form parts of C-end degrons, being recognised by specific ubiquitin ligases.

18C‐degron SLiMs consist of a few (typically only 2) conserved residues. A C‐terminal location and high structural accessibility (local disorder) are essential for C‐degrons to trigger degradation. C-terminal degrons can sometimes be present at the native C-termini of proteins or can be introduced by alternative mechanisms including premature translation termination or post-translational modifications. Terminal degron signals can also initially be internal, being latently present within the protein sequence and becoming exposed following proteolytic cleavage. E3 ligases recognising C-end degrons can be single- or multi-subunit molecules, like cullin-RING ligases (CRLs). The list of ubiquitin ligases and recognition subunits known to be associated with DesCEND pathways is extensive: KLHDC1/2/3/10, FEM1A/B/C, DCAF12, APPBP2, TRPC4AP, TRIM7, PIRH2, etc.. Unfortunately, not all E3 ligase substrate preferences have been sufficiently explored. The current ELM update includes new motif class definitions for the C-end recognition motifs of the DCAF12, FEM1A, FEM1B, FEM1C, KLHDC2 and TRIM7 ubiquitin E3 ligases (Table 2). Below we briefly discuss the DCAF12 and KLHDC2 cases.

19The cullin 4-RING (CRL4) ubiquitin ligases are highly modular and associate to various substrate recognition subunits (termed DCAFs) through their shared subunit DDB1. DCAF12 has been found to recruit proteins with a C-terminal di-Glu motif (Glu-Glu-COOH). Known substrates of CRL4DCAF12 Cullin-RING E3 ligase include CCT5, a member of the T-complex protein ring (TRiC) chaperonin complex, as well as the MAGE-A3 and MAGE-A6 melanoma-associated antigens, involved in the regulation of ubiquitylation and autophagy. In all these examples, the extreme C-terminus of the substrate protein is bound tightly at the center of the WD40 domain of DCAF12, by a positively charged coronet of lysine (K91, K108, K254 in PDB: 8AJM/8AJO) and arginine (R203, R256, R344 in PDB: 8AJM/8AJO) residues. However, other, yet-unknown recognition modes might also exist for DCAF12, explaining its interaction with IAPs helping cell survival or MOV10 involved in multiple cellular roles.

20The KLHDC1, KLHDC2, KLHDC3 and KLHDC10 proteins are closely related vertebrate cullin 2 (Cul2) or cullin 5 (Cul5) RING (CRL2/CRL5) ubiquitin ligase substrate recognition subunits, binding C-terminal degrons ending in glycine with distinct but overlapping specificities. Out of this group, KLHDC2 is the best characterised protein, associated with Cul2, responsible for the recognition of substrates ending in a di-Gly (Gly-Gly-COOH) motif. Structural studies reveal that the degron is bound to the center of the Kelch repeat domain, with the twin arginines (R236 and R241 in PDB: 6DO3/6DO4/6DO5) coordinating the C-terminal carboxylate. The size of this pocket is restricted to binding either GG$ or (rarely seen) GA$ motifs. The last 6 amino acids also contact the surface, but there is relatively less restriction imposed on them, except for a preference for small side chains.

21Interestingly, Di-Gly degrons (or Gly-Ala degrons), substrates of CRL2KLHDC2 can be generated by at least three distinct physiological processes. Some rare proteins natively end in two glycines, such as EPHB2. These motifs can also be generated by proteolysis, as they are compatible with the substrates processed by certain ubiquitin-specific proteases, including an autoprocessing site in USP1. Several selenoproteins can also suffer premature termination in case of selenium starvation, precluding the physiological overriding of the stop codons by selenocysteinyl-tRNAs. Selenoprotein K (SelK) and Selenoprotein S (SelS), when terminated prematurely, will end in a di-Gly motif, that is targeted by KLHDC2 (as well as other, related subunits) for ubiquitylation and destruction. Finally, KLHDC2 has also been shown to recognize its C-terminus (Gly-Ser-COOH) due to its resemblance to a degron, leading to the formation of inactive oligomers, adding a further layer of regulation to this already complicated system.

22Similar to degron-mediated protein destruction, the endoplasmic reticulum (ER)-associated degradation (ERAD) pathway functions as a guardian of exported and vesicular protein quality. This pathway differentiates between folded and misfolded proteins inside the ER lumen and shuttles problematic proteins into the cytosol through membrane channels in an energy-dependent manner. Therefore, malfunctioning of the ERAD pathway can lead to the accumulation of defective proteins, a condition implicated in >60 diseases that include neurological disorders, cancer, and cystic fibrosis. Moreover, many viruses, including several Nidovirales species such as coronaviruses, hijack the ERAD machinery during viral replication while simultaneously evading immune detection by the host. Viruses exploit ER membrane-derived structures, such as double-membrane vesicles (DMVs), to conceal viral RNA from cytosolic cellular sensors responsible for initiating interferon production and innate immunity responses.

23Prior to their elimination, most ERAD substrates are tagged with ubiquitin chains, which then serve as docking sites for VCP (Valosin-Containing Protein, also known as P97/TERA). VCP is an abundant protein (among the top 5% of cellular proteins) of the AAA+-ATPase group. It regulates various cellular processes, ranging from protein quality control to supporting DNA damage response and membrane trafficking. Importantly, VCP mutations have been linked to diseases characterised by protein aggregation, including Multisystem Proteinopathy (MSP), Familial Amyotrophic Lateral Sclerosis (fALS), and Charcot-Marie-Tooth Disease Type2Y (CMT2Y). Essential to VCP’s interactions is its N-terminal domain (VCPN), which acts as a hub for binding various adaptor proteins. These interactions happen through three known motifs: VIM (VCP-Interacting Motif), VBM (VCP-Binding Motif), and the SHP box. These three motifs have been added as new ELM classes in the current update (Table 2). Notably, the VIM motif is hallmarked by a sequence module flanked on both sides with arginine preferences and a hydrophobic centre with conserved alanine residues. The VBM motif shows similarities with VIM in its arginine-rich amino acid content and interacts with the exact same VCP surface. In contrast, the SHP box forms a short, antiparallel β-strand, which augments the β-sheet of VCP’s NTD subdomain at a site distinct from the two other VCP-binding motifs. VCP itself has the newly annotated Pub domain Interacting (PIM) motif that docks to the E3 ligase HOIL-1-interacting protein (HOIP), a part of the Linear ubiquitin chain assembly complex (LUBAC) which modulates the NF-κB pathway.

24In the current ELM update, we have added two new classes that define the phosphorylation site specificity of two kinases. One class represents the substrates of two Ser/Thr directed kinases, Adaptor-Associated kinase 1 (AAK1) and BMP-2 Inducible Kinase (BIKe/BMP2K). The second class defines the specificity of the Lymphocyte-oriented kinase (LOK), also named STK10 (Table 2).

25AAK1 and BIKe/BMP2K are endocytic kinases that regulate diverse cellular functions, including receptor-mediated endocytosis, Notch pathway regulation, and dendrite morphogenesis. These kinases preferentially phosphorylate threonine residues and phosphorylate the QxTG motif present in the adaptor-related protein complex mu2 subunit (AP2M1). Endocytic functions of these kinases also enable them to play a crucial role in controlling intracellular trafficking during the SARS-CoV-2 infection and entry process. Hepatitis C virus (HCV) and dengue virus (DENV) are other examples where these kinases are involved in viral entry and assembly/egress processes. The motif preference of the AAK1 and BIKe kinases is known from peptide SPOT arrays and other low-throughput in-vitro kinase assays and indicates a strong preference for a Gly residue at the T + 1 position, Gln at the T – 2 position and Ile at the T – 5 position. Prk1p, the yeast homolog of AAK1/BIKe kinases, phosphorylates the Pan1p, Sla1p, and Scd5p substrates which all contain a similar phosphorylation motif.

26LOK/STK10 regulates cytoskeletal dynamics in addition to affecting the cell shape and movements. An example of the latter includes its role in lymphocyte migration and microvilli formation in epithelial cells, which is achieved via phosphorylating the ERM proteins (Ezrin, Radixin and Moesin). Defining the specificity determinants of the LOK phosphosites is challenging due to the high sequence conservation of the C-terminal ERM protein regions, which contain the target phosphosites. Nevertheless, experiments from Belkina et al. and SPOT arrays from Johnson JL et al. reveal that LOK preferentially phosphorylates Thr over Ser residues. LOK sequence preferences around the phosphosite include Tyr at the T-2 position, large hydrophobic or aromatic amino acids at the T + 1 position and positively charged amino acids in the T – 3 and T + 2 positions flanking the phosphosite.

27In addition to the two kinase classes with Thr site-specificity, a substantial update has been done to three classes describing the specificity for NEK (MOD_NEK2_1 and MOD_NEK2_2) and DNA damage signalling (MOD_PIKK_1) kinases. (Table 2). The ELM kinome update was assisted by the availability of a recent comprehensive assessment of the human Ser/Thr kinome using SPOT array technology, where the phosphosite preferences for 303 kinases have been profiled. This provides a unique opportunity to further improve kinase motif definitions in ELM by integrating these SPOT arrays with existing experimentally identified phosphosites.

28Receptor tyrosine kinases (RTKs) regulate proliferation and actin cytoskeleton remodelling in response to extracellular signals. RTKs become autophosphorylated in tyrosine residues (pTyr) and relay the signal to non-receptor Src-family kinases (SFKs) and adaptor proteins such as NCK1/2, among others. This first step is achieved through binding of pTyr motifs to Src-homology 2 (SH2) domains present in SFKs and NCK1/2. The human proteome contains over 100 SH2 domains which bind defined ligands but present a significant degree of cross specificity. The present ELM update adds an extensive revision of the NCK (LIG_SH2_NCK_1) and SFK (LIG_SH2_SFK_2) specificity classes, as well as a new class describing the autoinhibitory interaction involving the C-tail of SFKs (LIG_SH2_SFK_CTail_3). The loss of the C-tail interaction leads to oncogenic activation in v-SRC. Viral and bacterial pathogens also mimic SFK and NCK SH2 motifs to induce cell proliferation and actin cytoskeleton remodelling.

29The human cell must be able to respond to many different types of pathogens and the tell-tale signs of infection that they provide. Within the cytosol of the cell, these can include both DNA (cGAS-STING) or dsRNA detectors (RIGI-MAVS) that control whether interferon regulatory factors (IRFs) can activate transcription. However, cells that are not already primed to expect pathogen invasion are most likely to succumb to the invader. Additionally, therefore, the cell surface is equipped with Tumour necrosis family receptors, Toll-like receptors, Interferon receptors and so forth so that it can be alerted to the presence of pathogens in the body and rapidly initiate defensive signalling internally, including transcriptional activation of critical defence genes such as interferons and cytokines by transcription factors such as NF-κB.

30The canonical NF-κB activation pathway begins by an activated TNF superfamily receptor trimerising and recruiting a combination of TRAF family E3 ligases via TRAF-binding motifs such as PxQ variants in the cytosolic tails (Table 2). Unlike most E3 ligases, the Lysine-63 ubiquitin chains generated by the TRAFs are activating, rather than destructive signals. The pathway proceeds via IKK protein kinase activation and IκB protein kinase degradation, ultimately releasing the NF-κB RelA component for nuclear entry and gene activation. The non-canonical pathway involves TNFR recruitment of a partially distinct subset of TRAFs and ultimately releasing the NF-κB component RelB into the nucleus for gene activation. These pathways involve communication between several multiprotein complexes but are typically shown as rather linear.

31The RIGI-MAVS detection systems for cytosolic pathogen dsRNA are present on the mitochondrial membrane's outer surface. The cGAS-STING detection systems for cytosolic pathogen DNA are also present on the mitochondrial outer membrane (but also the endoplasmic reticulum). Both of these systems control interferon regulatory factors, using the phosphorylated LxI (pS) SLiM (Table 2) to keep the IRFs in the cytosol as long as there are no pathogenic nucleic acid signals. Most of the IRF protein family have the same SLiM and, when phosphorylated by TBK1 or IKKϵ at the SLiM itself, they homodimerise and escape the cytosol into the nucleus where they can bind and oligomerise the general transcription factors P300/CBP and together stimulate interferon gene transcription.

32There are many complications to the innate immunity systems simplistically described above - for example, they connect to each other via some of their SLiM-mediated interactions. One of the strongest connections is through the TANK Binding Kinase (TBK1). Currently, the only biophysically identified SLiM binder to TBK1 is STING, which interacts via a TBK1 docking motif (DOC_TBK1_STING_1). The acronym TANK is derived directly from NF-κB regulation (TRAF family member-associated NF-kappa-B activator), but TANK and TBK1 are reported to be core components of an IRF3 activation complex. Figure 4 shows a network of twenty close interactors with TBK1: Multiple interactions unite the NF-κB and IRF gene activation pathways. These closely connected anti-pathogen systems illustrate why it is essential to think in terms of regulatory networks (where vertical pathways represent specific routes of information flow that can be controlled by any number of lateral interactions subject to other cell state conditions), as well as the fundamental importance of different SLiM classes in how network state is being modulated.

33AlphaFold has proved to be an effective predictor of IDRs. ELM annotators now routinely use available structure predictions from the AlphaFold database as they aid assessing SLiM instance plausibility, taking advantage of properties such as estimated disorderliness (a low pLDDT score). AlphaFold prediction results have recently been incorporated into LeishMANIAdb to more precisely score the probability of putative SLiMs in Leishmania species proteomes, that might interact with host proteins.

34In addition there is increasing interest in using AlphaFold for de novo modelling of putative SLiM interactions with known or potential interaction domains. For example, in a recent C degron study, AlphaFold was used to model the interaction site of a POLD2 Proline C degron on FEM1B, predicting a distinct binding site to the better studied Arginine C-degron. In a benchmarking study of a protein-peptide interaction dataset, AlphaFold2-Multimer achieved only 40% success rate in modelling the correct binding site and interface by default protocol. However, the accuracy was enhanced with a combination of prediction modes, which included different protocols to incorporate the peptide alignment information while modelling the protein–peptide complex. In a second study, an ELM-derived dataset was used to test the accuracy of AlphaFold-Multimer in modelling SLiM interactions, finding that a fragment-based approach around the interacting SLiM region can enhance the performance of modelling SLiMs. We expect the application of AlphaFold for SLiM prediction will be a growing area of bioinformatics research.

35Research into the biology of SLiMs continues its steady expansion. SLiMs underlie much of the complexity in cell regulation. Additional classes of SLiMs are being revealed while the depth of understanding accrues for the more established SLiMs. The importance of SLiMs in disease and infection continues to be better appreciated. Despite these advances, we consider that the current set of experimentally defined SLiMs only represent a small fraction of the ‘SLiMome’ in eukaryotic proteomes. The adoption of more expansive experimental technologies can identify novel SLiMs and feed them into more directed experimentation allowing further refinement of the recognition patterns for incompletely defined motifs. In addition, artificial intelligence approaches such as AlphaFold hold out the tantalising promise of playing a pivotal role in de novo computational prediction of new SLiM candidates together with their interacting protein domains. This might all mean that resource providers and funders need to work together to further enhance existing tools, develop new ones and better integrate the SLiM resources with each other and with other bioinformatic and biomedical platforms in the way that, for example, InterPro has recently connected with ELM. Only then can the research community fully benefit from knowledge bases such as ELM and the enormous efforts that so many have contributed to their development.