Motif-guided identification of KRAS-interacting proteins

Abstract

Background

For decades, KRAS has always been a huge challenge to the field of drug discovery for its significance in cancer progression as well as its difficulties in being targeted as an “undruggable” protein. KRAS regulates downstream signaling pathways through protein–protein interactions, whereas many interaction partners of KRAS remain unknown.

Results

We developed a workflow to computationally predict and experimentally validate the potential KRAS-interacting proteins based on the interaction mode of KRAS and its known binding partners. We extracted 17 KRAS-interacting motifs from all experimentally determined KRAS-containing protein complexes as queries to identify proteins containing fragments structurally similar to the queries in the human protein structure database using our in-house protein–protein interaction prediction method, PPI-Miner. Finally, out of the 78 predicted potential interacting proteins of KRAS, 10 were selected for experimental validation, including BRAF, a previously reported interacting protein, which served as the positive control in our validation experiments. Additionally, a known peptide that binds to KRAS, KRpep-2d, was also used as a positive control. The predicted interacting motifs of these 10 proteins were synthesized to perform biolayer interferometry assays, with 4 out of 10 exhibiting binding affinities to KRAS, and the strongest, GRB10, was selected for further validation. Additionally, the interaction between GRB10 (RA-PH domain) and KRAS was confirmed via immunofluorescence and co-immunoprecipitation.

Conclusions

These results demonstrate the effectiveness of our workflow in predicting potential interacting proteins for KRAS and deepen the understanding of KRAS-driven tumor mechanisms and the development of therapeutic strategies.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12915-024-02067-w.

Background

As one of the best-known and extensively studied oncogenes, KRAS (Kirsten rat sarcoma) has been implicated in various types of malignancies, making it a critical therapeutic target. KRAS belongs to the RAS family and is one of the most frequently mutated oncogenes across human cancers [1, 2]. KRAS possesses GTPase activity and is one of the guanosine triphosphate (GTP) binding proteins which switches between inactive and active states by binding with GDP or GTP, allowing its interactions with various proteins to regulate downstream signaling pathways to impact cell survival and proliferation [3]. KRAS can bind to numerous effector proteins, such as RAF-kinases, PI3K, and RalGDS, when it is in the GTP-bound state [1]. The activation of KRAS occurs when a guanosine exchange factor (GEF) facilitates the release of GDP from the nucleotide binding site, allowing GTP to bind. KRAS is inactivated when GTPase-activating proteins (GAPs) significantly accelerate the hydrolysis of GTP to GDP [3–5]. KRAS consists of a G domain that performs essential biological functions and a flexible C-terminus called hypervariable region (HVR). The G domain contains six β-strands that form the core of KRAS and are surrounded by five α-helices. The regions linking α1 and β2, as well as that linking α2 and β3, are referred to as switch I and switch II, respectively. KRAS partner proteins often bind to the interface formed by these switch regions [6]. The P-loop, linking α1 and β1, is a mutation hotspot in cancer, in which residue 12 or 13 within this region are common mutation sites, with the most frequent mutations being G12C, G12D, G12V, and G13D, which often lead to the occurrence and development of multiple cancers, such as non-small cell lung cancer (NSCLC), pancreatic cancer (PDAC), and colorectal cancer (CRC) [6–8]. When mutated, KRAS remains bound to GTP due to the disruption of GAP activity, resulting in continuous activation of downstream signaling pathways, the unremitting stimulation of cell proliferation, and the ultimate promotion of tumorigenesis.

Despite its well-established role in oncogenesis, effective targeting of the “undruggable” KRAS has historically been challenging due to its nearly spherical structure with no obvious binding site on the protein surface, strong binding affinity to GTP, and frequent mutations [6, 9]. Additionally, KRAS shares high sequence and structural identity with RAS homologs, NRAS and HRAS, in the G domain, raising the possibility that the inhibitors targeting KRAS may have off-target effects by inhibiting these small GTPases [8, 10].

Over the years, a series of strategies to directly or indirectly target KRAS have been proposed, one of which is to disrupt the protein–protein interactions (PPI) of KRAS to block the RAS pathway. It is estimated that there are about 650,000 pairs of PPIs in human [11]. PPI is one of the pivotal biological processes in life activities for its significance in the realization of protein functions. Most diseases that threaten human health are related to PPIs, such as cancers and neurological diseases [12, 13]. Hence, understanding how proteins interact with each other helps us elucidate disease mechanisms at the molecular level and guide the discovery of disease-specific drugs. PPI is actually a process wherein two motifs from the interface of an interacting protein pair match and interact with each other and the mutual recognition between two motifs follows a certain pattern [14, 15]. Since KRAS regulates the signal transduction via PPI to transmit the upstream signals to the downstream pathways, attention has been focused on targeting its upstream and downstream proteins to disrupt the interaction and subsequently block the relevant pathways. Upstream, the activation of KRAS requires the catalysis of SOS protein, which provides a great perspective for designing inhibitors for KRAS. Sun et al. reported a series of novel small molecules that can directly bind to KRAS between switch I and switch II, inhibiting the activation of KRAS catalyzed by SOS protein [16]. Hillig et al. discovered a SOS1 inhibitor that can interrupt PPI between KRAS and SOS1 with an IC₅₀ of 21 nM [17]. Downstream, for example, RAF is a vital effector protein of KRAS in the MAPK pathway. Recently, a series of potent KRAS^G12D inhibitors has been reported to selectively interrupt its binding with KRAS [18].

Another strategy is to directly target KRAS, with the most explored one being the G12C mutant. Two inhibitors have garnered widespread attention: sotorasib (also known as AMG510) and adagrasib (also known as MRTX849) developed by Amgen and Mirati Therapeutics. These inhibitor function by forming a covalent bond with cysteine 12 of KRAS^G12C, thereby locking KRAS in its inactive state [19, 20]. However, these promising inhibitors have also encountered drug resistance due to various factors, including conformational changes caused by mutations, feedback activation of KRAS upstream and downstream signaling pathways, and other related mechanisms [9]. For instance, the activation of PAK and PI3K pathways has been found to resistance to the first KRAS inhibitor sotorasib [21]. We wonder whether there are other interaction partners of KRAS that transmit downstream signals by regulating certain pathways thereby causing drug resistance.

Therefore, it is vital to explore the interaction modes between KRAS and its known binding partners. In addition to the SOS protein mentioned above, the currently reported KRAS interacting proteins, such as RAF, PDE6D, and NF1, jointly regulate downstream pathways with KRAS through PPI [22]. The BioGRID database, a public resource for genetic and protein interaction data, includes more than 900 interactors that physically bind to KRAS [23].

Current methods for detecting KRAS PPIs predominantly rely on experimental approaches such as co-immunoprecipitation (Co-IP), affinity purification-mass spectrometry (AP-MS), and yeast two-hybrid (Y2H) arrays [24–26]. These traditional methods, while precise and direct, demand significant resources and manpower. For instance, Co-IP requires specific antibodies and cell culture, while AP-MS involves expensive equipment and skilled technicians. Y2H assays can be time-consuming with extensive manual operations.

In contrast, computational methods, though potentially less accurate, offer advantages in terms of efficiency and cost, making them well-suited for large-scale screening and generating preliminary hypotheses about PPI. Additionally, the experimental technologies mentioned above are more effective for strong binding systems, while weaker or transient interactions may not be detected as expected. This further emphasizes the complementary role of computational methods in identifying a broader range of PPIs. Despite the numerous interactors reported, further investigation into KRAS PPIs and their interaction patterns is essential to better understand KRAS signaling and to aid in the discovery of PPI inhibitors.

Our in-house PPI prediction method, PPI-Miner, developed by Wang et al., is a tool designed for identifying and modeling PPI [14]. PPI-Miner leverages both protein sequence and structural features to predict potential interaction partners. The tool comprises two main sub-methods: 3DPPI-Miner and 2DPPI-Miner. 3DPPI-Miner predicts interactions based on three-dimensional structural data, making it suitable for proteins with known structures, while 2DPPI-Miner utilizes protein sequence data for interaction prediction, which is particularly useful for proteins lacking structural information but with abundant sequence data.

Once a potential interacting protein partner is identified using either 3DPPI-Miner or 2DPPI-Miner, we can construct complexes between the receptor protein, such as KRAS, and the potential binding partners. The binding affinities of these complexes are estimated by the designed scores according to the energy terms from Rosetta, with lower scores indicating stronger binding affinities. Specifically, the binding scores from 3DPPI-Miner are obtained by evaluating the binding of the entire KRAS-candidate partner protein complex with consideration of conformational matching between proteins, while the relative difference scores in binding affinities given by 2DPPI-Miner are derived from the comparison between the reference motif-protein complex and the candidate motif-protein complex. This method was successfully employed to unearth the potential substrate of cereblon (CRBN), a widely known E3 ligase [14].

Herein, we have developed a pipeline that applies multiple computational approaches, including the PPI-Miner method, to predict KRAS-interacting proteins, combined with experimental means to carry out verification studies.

Results

Herein, we developed a workflow to computationally predict and experimentally validate potential KRAS-interacting proteins (Fig. 1). The workflow was divided into three main stages: (1) Motif extraction: We collected resolved complex structures containing KRAS from the Protein Data Bank (PDB) [27] and extracted motif structures of partner proteins from these complexes. (2) Motif-guided target protein identification and candidate filtering: 3DPPI-Miner was utilized to identify potential KRAS-interacting proteins and predict complex structures. These candidates were then filtered sequentially through several stages, including subcellular location filtering, redundancy removal, 2DPPI-Miner filtering, and visual inspection. (3) Experimental validation: The predicted interaction partners, refined through this comprehensive filtering process, were subjected to further experimental validation.

Fig. 1.

The workflow of prediction and validation of KRAS-interacting proteins

Motif extraction

As of July 2022, we have searched a total of 67 KRAS-containing complex structures from PDB and filtered non-redundant proteins involved in interactions with KRAS. We then analyzed the interfaces of the complex structures. The interfaces that cannot be used to extract motifs are roughly divided into the following types: (1) The N-terminus/C-terminus of KRAS is inserted as a fragment into a certain pocket of its interacting protein. In our case, KRAS is supposed to provide the pocket rather than the motif. (2) The interface is scattered into pieces in sequence rather than concentrated in a particularly continuous region, which prevents us from intercepting fragments of appropriate length as motifs for fishing. Globally, we obtained 12 appropriate complex structures that mainly occupy five different sites on the surface of KRAS (Fig. 2). The PDB codes and the names of the interacting proteins are listed in Table 1.

Fig. 2.

Known KRAS-interacting partners occupy 5 sites of KRAS. The gray cartoon represents KRAS. Gray sticks represent small molecules binding to KRAS, including GDP and analogs of GTP. Cartoons highlighted in cyan represent the interacting regions of KRAS. Cartoons in other colors are representative interacting proteins from multiple complex structures of each binding site. Pink: R11.1.6 (PDB code: 5UFE); yellow: ARAF (PDB code: 2MSE); limon: (PDB code: 5XCO); slate: H-REV107 peptide (PDB code: 7C41); light magenta: AFFIMER K69 (PDB code: 7NY8)

Table 1.

PDB codes and protein names of KRAS-containing complex structures

Sites	Located regions	PDB codes	Interacting proteins of KRAS (protein name)
Site 1	Switch II	5UFE	R11.1.6
		5WHE	Miniprotein 225–11
Site 2	Switch I	2MSE	ARAF
		6PTS	RAF1
		7LC1	SIN1
		7SCX	RGL1
		7VVB	SIN1
Site 3	Between switch II and α3	5XCO	KRpep-2d
		6WGN	KD2
		6YXW	Affimer K6
Site 4	Among switch I, switch II, and P-loop	7C41	H-REV107 peptide
Site 5	Between α4 and α5	7NY8	AFFIMER K69

We here defined motifs as the continuous or slightly discontinuous fragments located at the PPI interface, with a distance of up to 6 Å. Site 1 includes complexes of KRAS with R11.1.6 and miniprotein 225–11 (PDB code: 5UFE and 5WHE, respectively) (Figs. 2A, 3A) [28, 29]. We selected one motif in each of these two complexes for subsequent calculations. The proteins related to site 2 are ARAF, RAF1, SIN1 (KRAS subtype: KRAS4A), RGL1, and SIN1 (KRAS subtype: KRAS4B) (PDB code: 2MSE, 6PTS, 7VVB, 7SCX, and 7LC1, respectively) (Figs. 2A, 3A) [30–34]. In particular, SIN1, when bound to the two isoforms of KRAS, KRAS4A, and KRAS4B, tends to exhibit different conformations. In order to extend the diversity of the motifs, we selected two complex structures of SIN1 binding to different subtypes of KRAS. As a result, two motifs were selected from each of ARAF and SIN1 (KRAS subtype: KRAS4B), three were selected from RAF1, and three motifs were selected from each of SIN1 (KRAS subtype: KRAS4A) and RGL1. Site 3 consists of complexes containing KRpep-2d, KD2, and Affimer K6 (PDB code: 5XCO, 6WGN, and 6YXW, respectively) (Figs. 2C, 3A) [35–37]. It is the site where SOS protein binds to KRAS and is also relatively close to G12, the highly mutated spot. We selected one motif each of KRpep-2d and KD2 and two motifs of Affimer K6 for subsequent calculations. As for site 4, the interacting protein at this site, H-REV107 peptide (PDB code: 7C41), is a short peptide which is within 6 Å of KRAS so that the whole protein was taken as a motif (Figs. 2D, 3A) [38]. For site 5, one motif of AFFIMER K69 (PDB code: 7NY8) was extracted (Figs. 2E, 3B) [37]. In total, 17 motifs from 12 suitable complex structures were extracted with their structures and sequences listed in Table 2.

Fig. 3.

Structure of motifs from multiple KRAS-interacting proteins binding to KRAS. Gray cartoon stands for KRAS and the cartoons in other colors stands for all the motifs extracted. A Site 1: 5ufe_frag1 (pink), 5whe_frag1 (raspberry); site 2: 2mse_frag1 (yellow), 2mse_frag2 (olive), 6pts_frag1 (lime green), 7lc1_frag1 (pale cyan), 7lc1_frag2 (green cyan), 7scx_frag1 (wheat), 7vvb_frag1 (light blue); site 3: 5xco_frag1 (limon), 6wgn_frag1 (salmon), 6yxw_frag1 (light orange), 6yxw_frag2 (orange); site 4: 7c41_frag1 (slate). B Site 2: 6pts_frag2 (lime), 6pts_frag3 (forest); site 5: 7ny8_frag1 (light magenta)

Table 2.

The structures and sequences of all extracted motifs

Sites	Motifs	Sequence
Site 1	5ufe_frag1	WVIRWGQYIWFKY
	5whe_frag1	EDLHEYWARLWNYLYAVA
Site 2	2mse_frag1	TVKVYLPNKQRTVVT
	2mse_frag2	KALKVRGL
	6pts_frag1	NTIRVFLPNKQRTVVNVR
	6pts_frag2	KTFLKLAFCD
	6pts_frag3	RCQTCGYK
	7scx_frag1	NGNMYKSIM
	7vvb_frag1	AAHGFSLIQ
	7lc1_frag1	HGFSLIQ
	7lc1_frag2	KAVKRRKGS
Site 3	6wgn_frag1	GXFVNFRNFRTFR
	6yxw_frag1	QHSIDIWYDF
	6yxw_frag2	WVKKLNNSHTYK
	5xco_frag1	RRRRCPLYISYDPVC
Site 4	7c41_frag1	LYDVAGSDKY
Site 5	7ny8_frag1	RQLRMGSMNN

Our previous cognition was that sequence determines structure. However, in fact, the motifs that bound to the same site were different in sequence, but relatively conserved in structure, such as site 2 and site 3 (Fig. 2A). Such motif pattern enabled them to bind to the same site even if they were not relatively conserved in sequence.

Motif-guided target identification and candidate filtering

PPI is a process in which two motifs recognize each other in a particular pattern [15]. Therefore, we assume that the PPIs of KRAS are also the consequence of mutual recognition between two motifs on the interface, one of which is from KRAS, and the other is from KRAS’s binding partner, referred to as the query motif. Thus, if a certain protein contains a fragment structurally or sequentially similar to the query motif, this fragment may have a chance to match the motif from KRAS through mutual recognition and form an interaction between KRAS and this certain protein. The 3DPPI-Miner method [14] can assist in finding these certain proteins by taking proteins that contain structurally similar fragments to the query motifs as ligands to simulate interactions with the target protein, KRAS, and filtering those who indeed form an interaction with KRAS. In 3DPPI-Miner, root-mean-square deviation (RMSD) represents the matching degree of the aligned Cα atoms of the residues along the motifs. A complex is filtered only if its RMSD value is below a set threshold. In order to adjust appropriate RMSD parameters for motif-guided target identification, we analyzed the interfaces of the complexes. When the motif skeletons of the partner proteins penetrate into the pocket of KRAS, only smaller conformational perturbation is allowed, necessitating a stricter RMSD (i.e., a smaller RMSD). The deeper the skeleton penetrates, the stricter the RMSD required. For example, we set RMSD parameters of 0.6 Å and 0.8 Å for 5xco_frag1 and 6wgn_frag1, respectively, due to the differences in the depth of their skeleton penetration (Fig. 3A). Additionally, motifs like 2mse_frag1 tend to interact with KRAS via the side chain rather than the backbone (Fig. 3A). Therefore, when performing subsequent calculation, the RMSD parameters need to be set more loosely (i.e., as a larger value) in order to find more potential PPIs. The parameters for all motifs are listed in the “Methods” section. We took the extracted 17 motifs as the query motifs respectively and utilized 3DPPI-Miner to identify the potential target in the human protein structure database. Based on the query motifs, 17 databases of potential binding partners for them were constructed. We obtained all the information regarding to subcellular location of the proteins in the candidate databases and filtered out those that were secreted or occurred only outside the cell or inside the nucleus. Since multiple conformation can be predicted for a protein with KRAS, we removed the structurally redundant PPIs of KRAS with the same protein under the consideration of their binding scores from 3DPPI-Miner as described in “Methods” section after the first filtering step. In the next step, to further narrow down the candidate databases, we extracted sequence motifs from the candidate proteins to form motif databases and applied 2DPPI-Miner [14] to evaluate the relative difference in binding possibilities between the reference motif-KRAS complex and the candidate motif-KRAS complexes from these databases. The binding scores from 3DPPI-Miner are obtained by evaluating the binding between KRAS and the candidate partner protein complex, considering conformational matching for further filtering purposes. In contrast, the scores given by 2DPPI-Miner assess the potential matching degree between the reference motif-KRAS complex and the candidate motif-KRAS complexes. This allows us to determine whether the motif itself, rather than the entire protein, can effectively bind to KRAS.

We selected the top 1000 proteins from each potential binding partner database for further screening based on the scores calculated by 2DPPI-Miner. In the subsequent stages, we considered the binding patterns observed in 3DPPI-Miner, scores from both 3DPPI-Miner and 2DPPI-Miner, and the tissue distribution data of the candidates during the visual inspection of the remaining complexes across 17 candidate databases. Overall, we obtained a total of 78 potential interacting partners of KRAS (Additional file 1: Table S1). Due to the high cost and heavy workload of the experiments, we selected 10 proteins for experimental validation based on their binding modes and scores, including one formerly reported interacting partner of KRAS, BRAF [39]. A reported cyclic peptide, KRpep-2d [35], which inhibits the activity of KRAS, was taken as our positive control (Table 3). Because the motif sequence of UBE3C and KRpep-2d cannot be aligned with the query motif 2mse_frag2 and 5xco_frag1, respectively, 2DPPI-Miner did not calculate the relative binding affinity score for the motif-protein complex.

Table 3.

The potential KRAS-interacting proteins

Query motif	Potential interacting protein	UniProt accession	Motif sequence	Scores
				3DPPI-Miner	2DPPI-Miner
2mse_frag1	GRB10	Q13322	DVKVFSEDGTSKVVE	− 7.229	− 173.297
	GRB7	Q14451	VVKVYSEDGACRSVE	− 1.151	− 168.908
	GRB14	Q14449	VIKVYSEDETSRALD	− 0.803	− 171.363
	BRAF	P15056	IVRVFLPNKQRTVVP	4.719	− 185.228
	SNX27	Q96L92	ELRVALPDGTTVTVR	5.483	− 183.403
2mse_frag2	MAGI1	Q96QZ7	QQAAKQGH	− 10.105	− 80.622
	PDE10A	Q9Y233	DEMKKLGI	− 9.257	− 86.352
	UBE3C	Q15386	TAFQSIGV	− 1.271	/
7ny8_frag1	RAB4B	P61018	VVNVGGKTVK	− 12.529	− 499.884
7scx_frag1	RASIP1	Q5U651	SGANYKSVL	− 26.736	− 154.378
KRpep-2da	/	/	RRRRCPLYISYDPVC	− 55.760	/

^aKRpep-2d is a reported peptide that binds to KRAS. Here, we took this peptide as the positive control

According to our hypothesis, proteins fished by the query motifs are supposed to contain motifs that share similar patterns to the query motifs in both structure and sequence. Therefore, we performed structural and sequence alignment of the 10 selected proteins and their respective query motifs in complex with KRAS. As a result, there were motifs that were conserved both in structure and sequence, such as motifs that were fished by 2mse_frag1 and 7scx_frag1 (Fig. 4A, D, respectively) and motifs that were fished by 2mse_frag2 and 7ny8_frag1 which were structurally conserved but not highly conserved in sequence (Fig. 4B, C, respectively). One possible reason is that we applied 3DPPI-Miner for motif fishing, which gave priority to structural conservation during motif fishing. As mentioned above, the high identity in structure and sequence between KRAS and its homologs may be an obstacle to the discovery of KRAS inhibitors. Unlike interactions between proteins and small molecules, PPI require larger interfaces. Even different mutants of the same protein may lose their ability to interact with original binding partners due to mutation-induced conformational changes, highlighting how even minor side chain differences in homologs like NRAS and HRAS can impact their interactions. Taken BRAF, a positive interaction partner which was verified above, as an example, we applied 3DPPI-Miner to predict the complex structures of BRAF with HRAS and NRAS, two family members of RAS, and evaluate their binding affinities (PDB code: 5J2R, 1CTQ, and 5UHV, respectively) [40–42]. Although these two proteins have high structural and sequence identities with KRAS, subtle conformational changes in the side chains still make their binding to BRAF weaker compared to the binding between KRAS and BRAF, according to the alignment and scoring result (Additional file 1: Fig. S1, Table S2). In addition, it was reported that BRAF displays preferential bindings with G12V or Q61R mutants of KRAS, HRAS, and NRAS, with the strongest interaction being with KRAS [43]. PPI-Miner tends to identify proteins that contain structurally or sequentially conserved regions. Due to the high sequence and structure identity between KRAS and its homologs, PPI-Miner may identify proteins that interact with HRAS or NRAS, such as GRB10 and GRB14 [44, 45]. This aspect of PPI-Miner is a double-edged sword, serving as both an advantage and a limitation. It is also worth noting that GRB7/10/14 all contain a Ras-associating domain [45]. Given PPI-Miner’s focus on conserved regions, it is reasonable that these proteins were identified as potential KRAS interactors.

Fig. 4.

The complex structures and sequence alignment of KRAS and 10 selected potential interaction partners. Wheat cartoon stands for KRAS and the maroon cartoon stands for all the motifs involved. A Proteins fished by the motif 2mse_frag1 of ARAF (dark blue) (PDB code: 2MSE) include GRB10 (sky blue), GRB7 (medium turquoise), GRB14 (royal blue), BRAF (cornflower blue), and SNX27 (cadet blue). B Proteins fished by the motif 2mse_frag2 of ARAF (dark slate blue) include MAGI1 (medium purple), PDE10A (violet), and UBE3C (thistle). C Protein fished by the motif 7ny8_frag1 of AFFIMER K69 (light slate gray) (PDB code: 7NY8) includes RAB4B (light gray). D Protein fished by the motif 7scx_frag1 of RGL1 (teal) (PDB code: 7SCX) includes RASIP1 (light blue)

Validation of peptides and protein interactions with KRAS via biolayer interferometry (BLI) assays

We expressed recombinant human KRAS and loaded it with the non-hydrolysable GTP analog GppCp (Cat.: ab146660, abcom, China), referring to the resulting protein as the GTP-loaded KRAS in all subsequent experiments, and immobilized it on streptavidin (SA) biosensors. The data showed that the peptide KRpep-2d had a high binding affinity with KRAS with a K_D value of 800 nM (Fig. 5A), which was consistent with the previous reports [35]. Subsequently, the binding affinities of 10 other peptides (as listed in Table 3) to KRAS were tested. As illustrated in Fig. 5 (see also Additional file 1: Fig. S2), 4 peptides were identified to exhibit moderate binding affinity to KRAS. The peptide_BRAF, a motif peptide extracted from the BRAF protein that has been reported to interact with KRAS previously, had a K_D value of 150 µM (Fig. 5B) [46, 47]. As reported before, the RAS-RAF interaction plays an essential role in the pathogenesis of many types of cancer [48, 49]. The selection of BRAF’s peptide as a positive control here serves to validate the rationality of our predictive method and testing system. Besides these, peptide_GRB10, peptide_GRB14, and peptide_RAB4B also exhibited interactions with KRAS with K_D values of 78 µM, 87 µM, and 84 µM, respectively (Fig. 5C–E). In contrast, other six peptides, including peptide_GRB7, peptide_SNX27, peptide_MAGI1, peptide_PDE10A, peptide_UBE3C, and peptide_RASIP1, showed no obvious binding response with KRAS (Additional file 1: Fig. S3).

Fig. 5.

Identified peptides directly interact with GTP-loaded KRAS protein by BLI binding kinetics assay. BLI sensorgrams and the binding affinity of A peptide_KRpep-2d, B peptide_BRAF, C peptide_GRB10, D peptide_GRB14, and E peptide_RAB4B to the GTP-loaded KRAS, respectively (n = 2). K_D value was calculated by fitting the maximum BLI response at different concentrations to a steady-state fitting curve

To further validate the effectiveness of our workflow in screening proteins interacting with KRAS, we tested PPI via the BLI binding kinetics assay. Considering the highest affinity of peptide_GRB10 with the GTP-loaded KRAS in the protein-peptide BLI experiments, we chose GRB10 for further verification. Based on the complex structure generated by 3DPPI-Miner, the RA-PH domain of GRB10 was found to be responsible for the interaction with KRAS (Fig. 6A). Combined with previous reports on the structural resolution of GRB10, we expressed the RA-PH domain of protein GRB10 [44]. As is shown in Fig. 6B, GRB10 RA-PH was found to bind to KRAS with a K_D value of 150 µM. To demonstrate the in vivo PPI of KRAS and GRB10, an immunofluorescence experiment was performed. The data demonstrated a co-localization of KRAS with GRB10 (Fig. 7B) and also with BRAF (Fig. 7A), an identified interaction partner of KRAS which was treated as a positive control. To further confirm the interaction between KRAS and GRB10, we performed co-immunoprecipitation using anti-flag antibody. The results showed that KRAS were co-precipitated with BRAF (Fig. 7C) and GRB10 (Fig. 7D), indicating the interaction of GRB10 with KRAS in vitro (the raw data of Co-IP experiments is provided in Additional file 1: Fig. S4). In order to further explore the correlation between GRB10 and KRAS, we tested the binding affinity of GRB10 with the KRAS^G12D mutant using the biolayer interferometry (BLI) method. Our results demonstrate that GRB10 binds to KRAS^G12D, with a K_D value of 51 µM (Additional file 1: Fig. S5). GRB10 is a vital adaptor protein that interacts with multiple tyrosine-kinase receptors, such as insulin-like growth factor receptor (IGF-IR), the insulin receptor (IR), and epidermal growth factor receptor (EGFR), participating in various signaling pathways and physiological functions [50, 51]. Additionally, GRB10 modulates crucial signaling pathways like PI3K-AKT and is involved in embryonic development and tumor formation [52]. Dysregulation of GRB10 has been associated with cancers such as breast, lung, and gastric cancers [53–55]. Given the pivotal role of the RAS signaling pathway in cancer, understanding the interaction between GRB10, the RAS pathway, and tumors could offer new insights for cancer prevention, diagnosis, and treatment.

Fig. 6.

The computational and experimental results of GRB10 RA-PH binding to KRAS. A The complex structure of GRB10 RA-PH and KRAS predicted by the 3DPPI-Miner method (wheat: KRAS; blue: GRB10 RA-PH; maroon: the searched motif). B Binding affinity (left) and steady-state fitting curve (right) of GRB10 RA-PH to KRAS by BLI binding kinetics assay. The K_D and R² are calculated by fitting of maximum BLI response at different concentrations to a steady-state fitting curve

Fig. 7.

GRB10 is determined to be an interacting protein of KRAS by immunofluorescence and Co-IP. A, B Immunofluorescence images (left) of fixed HCT116 cells stained for BRAF (red, top), GRB10 (red, bottom), KRAS (green), and a merged image (yellow) illustrate co-localization of BRAF and GRB10 with KRAS. Intensity-distance profile of fluorescence co-localization (right) for BRAF (red, top)/GRB10 (red, bottom) and KRAS (green). Intensity-distance profiles generated using ImageJ’s Plot Profile tool. C, D The interaction between BRAF (left), GRB10 (right), and KRAS as determined by Co-IP

Discussion

KRAS has always been an essential drug target for that its mutation often leads to abnormal activation of downstream pathways, thereby promoting the occurrence and development of multiple cancer types. Given the important role of PPI in KRAS regulatory pathway, we developed a comprehensive workflow using the PPI-miner method to identify potential KRAS binding partners. Through multiple filtering steps, we screened 78 potential interacting proteins of KRAS. Due to the high cost of experimental validation, we conducted BLI experiments on 10 peptides from these potential proteins. The results demonstrated 4 of the 10 peptides successfully bound to KRAS, underscoring the feasibility and effectiveness of our workflow. In addition, 12 proteins from our potential interacting proteins list have been found to be documented as the interactors of KRAS in BioGRID database, which provides further retrospective validation of the reliability of our PPI-Miner method (Additional file 1: Table S3).

Although we got some positive results, there is still room for improvement. Firstly, when extracting motifs, if the fragments were slightly discontinuous, the two discontinuous fragments were combined into one motif, which would introduce some noise. There is a module in the PPI-Miner method that takes the discontinuity of the fragments into consideration; we may try that function later. However, even if we adopt this parameter, it only optimizes the slightly discontinuous motifs we have extracted so far. For those KRAS-interacting proteins whose interfaces are dispersed into many small fragments, motifs cannot be extracted to search for potential interacting proteins, which is also a limitation of the PPI-Miner method. But this limitation is mainly caused by the nature of discontinuities originating from protein interaction interfaces. In PPI systems, the contribution to binding and affinity usually comes mainly from a few hotspot residues, which are often discontinuous. Theoretically, this search method on the scale of the motif in discontinuous form could therefore also correctly reflect the mechanism of PPI. In addition, the method itself can still be improved, e.g., in terms of scanning such discontinuous sites and conformational changes depending on different amino acid combinations, etc., which we hope to improve in the future. Secondly, due to the conformationally flexible nature of the peptides, they may not form the exact binding modes with KRAS as we predicted, which leads to the relatively low affinity in our experiments. Additionally, we noticed that one of the predicted proteins, MAGI1, which is also listed in BioGRID database as a confirmed interaction partner, failed in our BLI validation. Proximity-dependent biotin labeling (BioID) was used to detect interacting proteins of KRAS in the study that reported KRAS and MAGI1 interactions [56]. BioID allows for the identification of proteins located near KRAS within a ~ 10 nm radius. The relatively low specificity of BioID increases the likelihood of false positives. And due to the broader labeling range of BioID, it may label proteins that are close to, but not directly interacting with each other, thus further introducing false positive results. Though our in vitro tests showed no significant interaction between the predicted peptide_MAGI1 and KRAS, we cannot completely rule out the possibility that MAGI1 may interact with KRAS through other regions. Furthermore, since we tested the binding of KRAS and motif peptides rather than the entire proteins, this may also affect quantification of binding, especially for weak binding. But at least we can say that the binding between KRAS and MAGI1 is a weaker case. More experimental means are still needed to further prove the reliability of our results as we delve deeper into our research.

Conclusions

In order to identify more interaction partners of KRAS, we developed a workflow to computationally predict and experimentally validate the potential interacting proteins. We collected and filtered 12 KRAS-interacting proteins, mainly occupying five sites of KRAS. By analyzing the interfaces between KRAS and these interacting proteins, we extracted a total of 17 relatively continuous fragments on the interfaces as query motifs, which were used as “baits” to search the human protein structural database at different RMSD cutoffs using 3DPPI-Miner. Several steps of filtering were conducted to finally obtain 78 potential interacting partners for KRAS. Ten of them and one positive control, KRpep-2d, were tested experimentally. We found that 4 out of 10 proteins showed binding affinity to KRAS through BLI experiments. Additionally, we selected GRB10 RA-PH protein for further PPI verification and confirmed that it also interacted with KRAS.

These results demonstrate the effectiveness of our workflow in predicting potential interacting proteins of KRAS. Given the importance of KRAS in cancer progression, our findings have significant implications for the development of new treatment options for KRAS-driven tumors. Further studies are in need to validate the effectiveness of our prediction method and to explore the physiological significance of the identified peptides and proteins.

Methods

Motif extraction

Collection of all the KRAS-containing complexes

We utilized PDB to obtain all the experimentally resolved complex structures that contain KRAS up to July 2022.

Extraction of motifs

Not all KRAS-containing complex structures were eligible for motif extraction. We only kept complexes where the binding partner contains relatively continuous fragments on the interface with KRAS. After screening, the qualifying interfaces of KRAS and its partner proteins were obtained. We intercepted the continuous fragments on the interface within 6 Å from KRAS as motifs. If the fragments were slightly discontinuous, only separated by 1 to 2 residues, the two discontinuous fragments were concatenated into one.

Motif-guided target identification and candidate filtering

Motif fishing

3DPPI-Miner uses the motif as a query to identify proteins in the database with similar structures to the query. Then, the searched motifs are aligned with the query to obtain the preliminary binding modes of the searched motifs and the receptor protein, which in our case, is KRAS. After the initial protein–protein complex conformation is determined, the method will optimize the complex conformation of the receptor protein, KRAS, and the proteins containing the searched motifs, and construct PPI complexes. Here, we took the structures of the motifs collected from suitable KRAS-containing complexes as queries to search in the human protein structural database mentioned in the PPI-Miner method which composed of the PDB Database, the SWISS-MODEL database, and the AlphaFold Database to identify potential proteins that interact with KRAS via 3DPPI-Miner. When searching, the RMSD cutoff was set to 1.3 Å for 2mse_frag1, 5ufe_frag1, 6pts_frag1, 6pts_frag2, 6pts_frag3, 6yxw_frag2, 7lc1_frag1, 7lc1_frag2, 7scx_frag1, 7vvb_frag1, and 7ny8_frag1; 1.0 Å for 2mse_frag2, 6yxw_frag1, and 7c41_frag1; 0.8 Å for 6wgn_frag1; and 0.6 Å for 5whe_frag1 and 5xco_frag1.

Candidate filtering

After motif fishing, a large number of candidate complexes form the candidate database. We obtained the final potential KRAS-interacting proteins via a series of filtering method.

Subcellular location filtering

According to the record in the UniProt database [57], the subcellular location of KRAS is distributed in cell membrane and cytoplasm which means that KRAS can only exist in the cell. Based on that, we collected the subcellular location information of all the proteins in the candidate database and retained those that could temporally and spatially interact with KRAS at the cellular level for further analysis.

Removing redundancy

A motif may sometimes match structurally similar fragments at multiple sites on a protein. In this case, 3DPPI-Miner will perform protein–protein docking on these multiple sites with KRAS, respectively. A motif may generate multiple conformations at the same site as well. With this in mind, we only kept the PPI with the best binding score related to this certain candidate protein while other poses were deleted.

2DPPI-Miner

The 2DPPI-Miner method is a sub-method of PPI-Miner, which is a sequence motif searching method. One of the functions requires the user to build a motif sequence database. The relative differences in binding affinities between the reference motif-protein complex and the candidate motif-protein complexes will be measured. With this function, we could better observe whether the motif rather than the whole protein to which the motif belongs can bind well to the receptor protein. We removed the motifs corresponding to the proteins whose binding scores calculated by 3DPPI-Miner were less than 0 from each candidate database to form the motif database, respectively. If there was no protein with binding score less than 0, the motifs of the proteins whose scores ranked in the top 2000 were taken to form the motif database. After obtaining the motif databases, we applied the 2DPPI-Miner to calculate and score the possibilities of binding between the reference motif-KRAS complex and candidate motif-KRAS complexes across motif databases.

Tissue distribution filtering

In Atlas database [58], we found that KRAS is widely distributed in human tissues, mainly in gastrointestinal tissues, liver and gallbladder tissues, bone marrow tissues, and lymphoid tissues, etc. Therefore, in the subsequent filtering, we focused more on proteins that expressed in these tissues.

Visual inspection

After the above filtering steps, combined with our experience in screening poses and the scores of PPIs given by both 3DPPI-Miner and 2DPPI-Miner, we visually screened the remaining PPIs in the candidate database to obtain the final potentially interacting proteins.

Experimental validation

Protein expression and purification

GRB10 RA-PH domain purification: We expressed the GRB10 RA-PH as described by Stevan R Hubbard et al. [44]. Briefly, we subcloned the cDNA encoding residues 106–357 (RA and PH domains) of human GRB10 (Q13322-3) into the expression vector pET28a and introduced some mutations C331S, C232S, C145S, C212S, K270A, and E271A. The plasmid was transformed into Escherichia coli (Rosetta 2(DE3), Cat: EC1014M, Weidi, China). Transformed bacteria were grown in Luria Broth (LB) media containing 50 μg/mL kanamycin at 37 °C until the OD600 reached 0.4–0.6, then cooled to 18 °C and induced with 0.2 mM isopropyl-β-D-thiogalactoside (IPTG, Cat: 10902ES10, Yeasen, China) for 18 h. The bacteria were collected by centrifugation, and the obtained pellet either stored at − 80 °C or used freshly for the subsequent steps.

The pellet was resuspended in lysis buffer (20 mM HEPES pH 7.4, 300 mM NaCl, 1 mM TCEP, 5% glycerol, 10 mM imidazole) containing a complete EDTA-free protease inhibitor cocktail (Cat: C0001, TargetMol, America). The bacteria were lysed by a French pressure cell press and the bacteria lysate was centrifuged at 20,000 rpm for 1 h at 6 °C. The supernatant was incubated with Ni-agarose beads for 3 h, then the loaded beads were washed with wash buffer (20 mM HEPES pH 7.4, 300 mM NaCl, 1 mM TCEP, 30 mM imidazole) and the protein was eluted with elution buffer (20 mM HEPES pH 7.4, 300 mM NaCl, 1 mM TCEP, 300 mM imidazole). The eluted protein was further purified by gel filtration (Superdex200, HiLoad 16/200, Cat: 28,989,335, Cytiva, America) with SEC buffer (20 mM HEPES 7.4, 150 mM NaCl, 1 mM TCEP).

Standard KRAS protein purification: We used the pET15b vector to express the 6xHis-tagged recombinant human KRAS protein. The transformed E. coli Rosetta 2(DE3) bacteria were cultured in LB medium supplemented with 100 μg/mL ampicillin. The induction and purification methods for the KRAS protein are consistent with those for GRB10 RA-PH described above, except that 4 mM magnesium chloride needs to be added to each buffer for purification.

Preparation of biotinylated GTP-loaded KRAS protein

KRAS protein was loaded with the non-hydrolysable GTP analog GppCp (Cat: ab146660, abcam, England). For GppCp loading, the full-length recombinant KRAS protein was incubated with 10 mM and Calf Intestinal Alkaline Phosphatase ((5 units per mg of KRAS protein) CIAP, Cat: 18,009,027, Invitrogen, America) at 4 °C overnight to load the metal and nucleotide site of KRAS fully. The reaction mixture was then desalted by PreCap desalting columns (Smartdex G-25, Cat: SEC018C55, Smart, China) to remove free nucleotide and CIAP. The protein after this step of treatment can be referred to as GTP-loaded KRAS.

We added 20-fold molar excess of Sulfo-NHS-Biotin (Cat: T19952, TargetMol, America) to GTP-loaded KRAS and incubated on ice for 4 h. The reaction mixture was then desalted with PreCap desalting columns (Smartdex G-25) to remove free Sulfo-NHS-Biotin. At this point, we have obtained biotinylated KRAS protein loaded with GTP.

Biolayer interferometry (BLI) binding kinetics assay

All BLI assays were performed on Octet RED96 instrument (FortéBio, Shanghai, China) using a shaking speed of 1000 rpm and a plate temperature of 25 °C. The buffer used for the assays was BLI buffer (20 mM HEPES pH 7.4, 150 mM NaCl, 0.01% Tween 20, 4 mM MgCl₂, 5 µM GPPCP, 1% DMSO (DMSO was used in the BLI experiment to investigate peptide-protein interactions but was not required for detecting PPIs.)). To immobilize the biotinylated GTP-loaded KRAS onto streptavidin biosensors, the streptavidin biosensors were first baseline run for 60 s in BLI buffer, then immersed in a well containing biotinylated GTP-loaded KRAS solution at a concentration of 25 µg/mL for 120 s, followed by another baseline run for 60 s in BLI buffer, and finally stored at room temperature in BLI buffer. The protein loading resulted in an approximate signal shift of 2–3 nm in the BLI measurements.

For the determination of binding kinetics, protein and peptide dilutions at different concentrations were added to a black polypropylene 96-well microplate filled with BLI buffer, leaving one row as the reference well. The total volume of each well was 200 μL. The sensors were incubated in BLI buffer, then moved and immersed into the peptide or protein dilutions to initiate binding and dissociation cycles. The binding and dissociation times were indicated in the legend. The BLI results were analyzed using FortéBio Data Analysis HT 11.1. A single reference subtraction was applied, where only the buffer reference well was subtracted. The data were analyzed based on a 1:1 binding model, and the K_D and R² are calculated by fitting of maximum BLI response at different concentrations to a steady-state fitting curve.

Cell culture and transfection

Human colorectal cancer cell line HCT116 cells were cultured in McCoy’s 5A medium (Cat: CB008, EpiZyme, China) containing 10% fetal bovine serum (FBS, Cat: FSP500, ExCell, China) and 1% penicillin streptomycin sol (Cat: 15,140–122, Thermo, America) at 37 °C incubator with 5% carbon dioxide. HCT116 cells were cultured in 6-well plates, after the cell fusion rate in the dishes reaches to 50%, the medium was replaced with serum-free McCoy’s 5A medium, then mixed 0.5 μg of HA-KRAS and 0.5 μg of Flag-GRB10 (or Flag-BRAF) plasmid DNA and 2 μL of P3000 (Cat: GK20006, GLPBIO, America) with 100 μL OptiMEM (Cat: 51,985,034, GLPBIO, America), left it at room temperature for 5 min, meanwhile, 3 μL lipo3000 (Cat: GK20006, GLPBIO, America) was added in 100 μL OptiMEM, left it at room temperature for 5 min. Then mixed immediately. After standing at room temperature for 15 min, the mixture was added to the wells drop by drop. After 8 h, the medium was changed to serum-containing McCoy’s 5A medium. After 48 h, cells were collected for co-immunoprecipitation analysis.

Co-immunoprecipitation analysis (Co-IP)

Cells were washed with phosphate-buffered saline (PBS, Cat: CB012, EpiZyme, China) and lysed in IP lysis buffer (Cat: 87,787, Thermo, America) containing 1% protease inhibitor and 1% phosphatase inhibitor at 4 °C for 10 min. Cell lysates were collected and centrifuged (12,000 rpm, 15 min, 4 °C), then the supernatant was incubated overnight at 4 °C with 1 μg anti-flag antibody (Cat: F7425, Sigma-Aldrich, America). Meanwhile, for each immunoprecipitation reaction, 20 µL Protein G Agarose (Cat: 20,421, Thermo, America) was blocked with 5% BSA in TBST solution overnight at 4 °C, then washed three times with PBS and combined with supernatants at room temperature for 1 h. The beads were collected and heated to 98 °C for 10 min with 2 × protein loading buffer (Cat: LT101S, EpiZyme, China) after washed with PBS for four times, and finally subjected to western blot analysis.

Immunofluorescence

Cells were seeded in culture slides, washed with PBS, and then fixed with 4% paraformaldehyde (Cat: ABS9179, Absin, China) at room temperature (RT) for 20 min. After washing with PBS, cells were blocked with permeable blocking solution (1% BSA + 0.2% Triton100) at RT for 1 h and incubated with primary antibody flag rabbit antibody (1:1000, Cat: F7425, Sigma-Aldrich, America) and KRAS mouse antibody (1:300, Cat: F234, Santa Cruz Biotechnology) at 4 °C overnight. Cells were washed with PBS, followed by secondary antibody Alexa Fluor 555-labeled Donkey Anti-mouse IgG(H + L) (1:500, Cat: Ab150066, Abcam, America) and Alexa Fluor 488-labeled Donkey Anti-mouse IgG(H + L) (1:10,000, Cat: A-21202, Abcam, America) incubating for 20 min at RT. After treating with DAPI fluoromount-GTM (Cat: 36308ES20, Yeasen, China), the slides were sealed by a coverslip and images were captured in ZEISS LSM 710 confocal microsystems.

Abbreviations

KRAS

Kirsten rat sarcoma

GEF

Guanosine exchange factor

GAPs

GTPase activating proteins

HVR

Hypervariable region

NSCLC

Non-small cell lung cancer

PDAC

Pancreatic cancer

CRC

Colorectal cancer

PPI

Protein-protein interaction

Co-IP

Co-immunoprecipitation

AP-MS

Affinity purification-mass spectrometry

Y2H

Two hybrid arrays

PDB

Protein Data Bank

SA

Streptavidin

IGF-IR

Insulin-like growth factor receptor

IR

Insulin receptor

EGFR

Epidermal growth factor receptor

LB

Luria Broth

IPTG

Isopropyl-β-D-thiogalactoside

FBS

Fetal bovine serum

PBS

Phosphate-buffered saline

RT

Room temperature

Authors’ contributions

F.B., X.Z, X.M. and S.W. conceived and designed this study; S.W. performed investigation, data acquisition, data analysis and data visualization; X.Z. and H.Y. designed experiments; X.G, D.W. and L.L carried out the experiments; S.W., X.G. and D.W. wrote the first draft; S.W., X.Z. and F.B. revised the manuscript; F.B., X.Z, X.M. and H.Y. supervised the study; F.B. provided the funding. All authors read and approved the final manuscript.

Funding

We thank the support from the National Natural Science Foundation of China (No 82003654), National Key R&D Program of China (2022YFC3400501), Shanghai Science and Technology Development Funds (Grant IDs: 22ZR1441400), Start-up Package, Joint Laboratory for Modular High Performance Computing at ShanghaiTech and ShanghaiTech AI4S Initiative SHTAI4S202404 from ShanghaiTech University, and Shanghai Frontiers Science Center for Biomacromolecules and Precision Medicine at ShanghaiTech University.

Data availability

The data including the input structures of complexes, the calculated score files, raw data of BLI assays and raw Co-IP experiment data are available at 10.5281/zenodo.10867430 [59].

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.