Protein language models accurately predict polymorphic peptide–modulated NK cell receptor–HLA class I interaction strengths

Abstract

Killer-cell immunoglobulin-like receptors (KIRs) are key determinants of natural killer cell function and are associated with the outcomes of infective, inflammatory, and neoplastic diseases. They form a polymorphic family of activating and inhibitory receptors that interact with polymorphic class I human leukocyte antigen (HLA-I) molecules. This interaction is dependent on the short peptides bound by the HLA-I molecules, including those derived from viruses and cancers. Identifying these peptides among the vast space of possible peptides based on the sequences of the interacting molecules can provide a valuable tool for developing personalized immunotherapy against infection and cancer. To address this challenge, we leveraged foundation protein language models and trained our model on available datasets for KIR-binding peptide-HLA complexes. Our tool generated excellent predictions with an area under receiver operator characteristic (AUROC) >0.8 for the majority of inhibitory KIRs and performed well (AUROC >0.7) for peptides generated during HIV and HCV infections. Our model holds substantial potential for advancing our understanding of immune regulation and the biophysical factors responsible for it, paving the way for KIR-specific therapeutic interventions.

ML model decodes immune recognition; KIRLinguist predicts KIR-HLA interactions, guiding personalized infection and cancer therapy.

INTRODUCTION

Natural killer (NK) cells are lymphocytes of the innate immunity and provide important protection against infections and tumors. In vivo NK cells are held in check by inhibitory receptors that recognize cell surface molecules, such as human leukocyte antigen class I (HLA-I) that are expressed on healthy cells (1). NK cells sense diseased cells through a combined signal transduced from a paired system of activating and inhibitory receptors (2). In particular, NK cells recognize stress molecules that are up-regulated when a cell becomes infected or malignant but also if there is down-regulation of ligands for the inhibitory receptors. As a result, these activating and inhibitory receptors act in concert to determine whether an individual NK cell becomes activated or remains tolerant when it encounters a target cell. Key to this paradigm are the killer-cell immunoglobulin (Ig)–like receptors (KIRs). These are polymorphic, activating and inhibitory, receptors that recognize HLA-I molecules (3). Target cells deficient in HLA-I are recognized by NK cells expressing inhibitory KIR because of the loss of the inhibitory signal from these KIRs (2, 4), but activating KIR can recognize viruses and cancer-associated antigens if they express specific peptides that bind HLA-I, and the combination of peptide and HLA-I is recognized by KIR.

However, KIRs display a large degree (>90%) of similarity in the sequences of the amino acid residues in the extracellular domain that interacts with HLA-I and peptide complexes (5). Those relatively small differences in the KIR sequences can affect the strength of their interactions with the HLA-I:peptide complexes. For example, KIRs with two Ig extracellular domains, namely, the inhibitory KIR2DL2/3 and activating KIR2DS2 both bind to the group 1 HLA-C molecules that have a serine and asparagine at positions 77 and 80 (S⁷⁷N⁸⁰) of the HLA-C molecule. Conversely KIR2DL1 and KIR2DS1 interact with the group 2 HLA-C molecules, which have asparagine and lysine at positions 77 and 80 (N⁷⁷K⁸⁰), and KIR2DL1 can interact with group 1 HLA-C molecules presenting specific peptides (6). In addition, the three Ig domain KIR (KIR3DL1 and KIR3DS1) bind HLA-A and HLA-B molecules with a Bw4 epitope at positions 77 to 83. The resulting differences in the structure and biophysical properties of the extracellular domains of the KIR arising from these differences in their sequences can produce substantial differences in their interactions with cognate ligands. Figure 1A illustrates these KIR:pHLA pairings.

Fig. 1. Peptide-specific KIR-HLA interactions and KIRLinguist model architecture.

(A) KIR binding HLA-I is both epitope specific and peptide specific. Pairings of the three families of HLA and KIR studied are illustrated. (B) Schematic of the model architecture. Left: The raw sequences of the three interacting moieties (KIR, HLA, and peptide) are translated into embeddings using a pretrained PLM (ProtT5), and the embeddings are averaged over the sequence length. The embeddings of all three entities are further averaged to give a single embedding of the full complex. The resulting embedding is used as input in an MLP model to give the prediction for binding affinity, either as a regressor or a classifier. Right: The model can additionally be supplemented by three-dimensional structure information before the averaging step. AlphaFold2-predicted structures can be translated into Foldseek’s 3Di, which is one-hot encoded into a numerical array to be appended to the PLM embeddings. The remaining steps follow similarly.

Structural and functional studies have shown that this KIR:HLA interaction is also critically dependent on the short peptides presented on the cell surface by the HLA-I molecule. These peptides are 8 to 11 amino acids in length and derive from intracellular proteins, including those associated with diseases such as viruses and cancer. The binding of KIR to HLA is governed by the second and third to the last amino acid residues of these peptides, i.e., residue positions P$\mathrm{\Omega }-2$ and P$\mathrm{\Omega }-1$, where P$\mathrm{\Omega }$ is the terminal residue of the peptide. These residues can promote or prevent binding of KIR to HLA-I in a “peptide-selective” model (7–9). Peptide selectivity has been most comprehensively studied for the inhibitory KIR: KIR2DL1, KIR2DL2, and KIR2DL3, which bind HLA-C, and are permissive for a greater range of peptides than their activating counterparts KIR2DS1 and KIR2DS2, which have a much more narrowly focused binding repertoire (8).

The KIRs have been associated with the outcomes of many different diseases including infection (10–14), cancer (15, 16), pregnancy-associated disorders (17–19), and autoimmune diseases (20–22), and the peptide-selectivity model has recently been proposed to underpin these associations (8, 23). Healthy cells present a large array of self-peptides to the immune system that, in general, have an inhibitory effect on NK cell activation. Viral infections can produce large changes in the peptides presented by HLA (24) and can lead to disinhibition of NK cells. This phenomenon has been modeled using synthetic peptide repertoires (25, 26). Conversely, activating KIRs recognize specific pathogen–derived peptides (10, 27, 28), suggesting that these KIRs are more likely to respond to the presence or absence of pathogen-derived peptides, akin to T cell receptors (TCRs) in T cells.

The KIR and HLA systems form two of the most diverse gene families in the human genome with many thousands of different alleles. This extensive polymorphism, along with the vast array of peptides that can be generated from endogenous proteins in health and disease, results in an immense combinatorial diversity of possible KIR:HLA-peptide interactions. This immense diversity makes it infeasible to empirically determine the binding affinities of every possible combination. Recognizing these challenges, recent studies have leveraged machine learning approaches such as multilabel vector optimization (29) to systematically interrogate the high-dimensional space of KIR:HLA-peptide interactions. By integrating large-scale binding data, peptide-binding motifs, and biochemical descriptors, these computational frameworks have shown great success in predicting KIR:HLA binding, but addressing the complex peptide specificity of this binding interaction has so far remained challenging (29). Particularly prominent models have been protein language models (PLMs) such as ESM and BERT-based models, which generate sequence embeddings that, when paired with machine learning architectures, have been shown to predict binding of TCRs to peptide-HLA complexes with high accuracy (30–34). Building on this progress, we have adopted a machine learning approach, based on the PLM ProtT5 (35), to develop a tool that can interrogate KIR:HLA-peptide interactions in silico (30–32).

In this work, we introduce KIRLinguist, a machine learning framework trained on multiple peptide-specific KIR:HLA-binding datasets. Our model accurately predicts KIR binding for both self and synthetic peptides, achieving area under receiver operator characteristic (AUROC >0.8), and demonstrates robust performance (AUROC >0.7) on clinically relevant viral peptide datasets. Furthermore, cross-dataset analysis among KIR2D receptors reveals that KIR2DS1 shares functional binding tendencies more closely with the sequence-divergent KIR2DL2 than with the more sequence-similar KIR2DL1. Comparison between inhibitory KIR2D and KIR3D receptors further highlights the potential role of the distinct KIR3D D0 domain in modulating binding specificity.

RESULTS

Development of a protein-language model for predicting KIR:HLA-peptide interactions

PLMs encode the structure and biophysical properties of a protein based on its sequence in terms of vectors of real numbers in high (e.g., 1024) dimensions, also known as embeddings. PLMs have been successfully used to model and predict interactions between the TCR and peptide-HLA complexes where the interaction strength depends sensitively on the sequences of TCRs and peptides. We constructed a model, KIRLinguist, based on the PLMs and a multilayer perceptron (MLP) classifier to describe and predict interactions between KIR and peptide-HLA complexes when the sequences of the receptors and ligands are provided as an input.

Our model architecture is shown in Fig. 1B, and below we provide more details about the model. The first half of the model uses pretrained PLMs (ProtT5, ProtBERT, and ESM-2) (35, 36) to obtain a numerical embedding representation of the protein complex. The embeddings were averaged over both sequence length and the individual moieties resulting in embeddings of size 1024 for ProtT5 and ProtBERT and 1280 for ESM-2. All the resulting embeddings were found to perform equally well. All three model embeddings have previously been shown to encode biological properties (37–39), allowing for accurate prediction of protein structure and function, with functionally similar proteins lying closer in embedding space.

The embeddings are then fed into an MLP model where further training specific to binding prediction is carried out. Training was performed on datasets by fitting either the MLPClassifier or MLPRegressor model from the scikit-learn library (40). The fully connected neural network models are trained using the fitting procedure with the default Adam optimizer and a constant learning rate of 0.001. Two different MLP hidden-layer sizes, (500) and (800, 800, 400, 200, and 100), were used to determine the effect on the model performance. The two cases are labeled MLP #1 and MLP #2, and their performances can be seen and contrasted visually in figures showing the performance of the regressor model. Quantitative metrics of performance show that there is no significant difference in performance. Last, hyperparameter tuning was performed to optimize other parameters of the MLP such as choice of activation function and strength of L2 regularization using the AUROC scoring metric.

KIRLinguist generates excellent predictions for binding strengths of inhibitory KIR2DL1, KIR2DL2/3, and KIR3DL1 receptors to canonical peptide–HLA-I complexes

We trained and tested KIRLinguist on several published datasets of KIR:HLA:peptide interactions. The two datasets used in this section are of inhibitory KIR2Ds binding HLA-C molecules (8) and inhibitory KIR3Ds binding HLA-A and HLA-B molecules (9) where the HLA molecules are complexed with a range of peptides. The experiments assayed the strength of interactions of KIR binding by measuring staining of target cells expressing HLA:peptide complexes by KIR-Fc conjugated to a fluorescent label protein A:Alexa Fluor 647; thus, the measured interaction strength quantifies the avidity of KIR binding with HLA-peptide complexes (see text S1 and fig. S4 for details) (7). The KIR3D dataset contained both experimentally tested bindings and bindings predicted on the basis of residues found to drive variance and shared biophysical properties between amino acids (9). Both datasets provide a measure of the binding avidity (BA) of KIR:HLA pairings with all possible variations of peptide positions P$\mathrm{\Omega }-2$ and P$\mathrm{\Omega }-1$, with the exclusion of cysteine in the KIR2D dataset. We quantified the performance of the classifier model’s performance based on the AUROC metric for the best performer in the KIR2D dataset, which paired the KIR2DL1 receptor with its ligand HLA-C2 (C*05:01) and the best performer in the KIR3D dataset pairing KIR3DL1 with its ligand HLA-B*57:01 (Fig. 2, A and B). We used K-fold cross-validation and plotted the AUROC for all $K=5$ fold. For the KIR2D dataset, where the BA was given as a continuous value, we marked $\text{BA}<0.5$ as nonbinder and $\text{BA}\ge 0.5$ as binder. The model performs very well with the K-fold average $\text{AUROC}>0.9$ and an accuracy measure $\text{Acc}.>90\%$ for the positive binders.

Fig. 2. Performance for canonical interactions of inhibitory KIR2D and KIR3D.

Receiver operator characteristic (ROC) curves as a performance metric for (A) the best performer in the KIR2D dataset (KIR2DL1:HLA-C2:peptide) and (B) the best performer in the KIR3D dataset (KIR3DL1:HLA-B*57:01:peptide). The plots show the ROC curves for each K-Fold of the model in addition to the mean ROC curve in blue. The mean AUROC value is stated in the plot legend. (C) AUROC performance of the classifier model for all canonical interactions of inhibitory KIR/HLA bindings (in purple) and the corresponding R² values for the regressor model (in pink). The KIR3D dataset was a classification, and so only AUROC values could be obtained.

We evaluated the performance of our model for additional canonical KIR:HLA interactions including inhibitory KIR (KIR2DL2 and KIR2DL3) binding HLA-C1 and KIR3DL1 binding HLA-Bw4 (HLA-B27, HLA-A24) using AUROC (Fig. 2C), which shows consistently high ($\text{AUROC}>0.8$) model performance. We also evaluated the performance of the regressor model for the KIR2D datasets using the coefficient of determination $R^2$ metric (shown in pink bars in Fig. 2C). In addition, we tested the performance on another dataset measuring binding of KIR2DL2 and KIR2DL3 to HLA-C*0102 (HLA-C1). The synthetic peptides were derived from a screen of 58 P7 and P8 variants of the peptide VAPWNSLSL (VAP), which was eluted from HLA-C*0102 expressing 721.221 cells (41, 42). Binding to both inhibitory KIR2DL2 and KIR2DL3 was measured, resulting in a roughly proportional distribution of high and low binders in both datasets. We applied our model to these datasets, and the performance was excellent with $\text{AUROC}>0.9$ in both cases.

Noncanonical interactions and the role of structural information in the model

Performance of the model for noncanonical KIR:peptide-HLA interactions (KIR2DL1/HLA-C1, KIR2DL2/HLA-C2, and KIR2DL3/HLA-C2) is shown in Fig. 3. The predicted (34, 43) versus true values for the binding avidities of the KIR2DL2/HLA-C2 dataset are shown as an example in Fig. 3A, and the $R^2$ values for all three datasets in Fig. 3B. The regression model performs moderately well for KIR2DL2/HLA-C2 and KIR2DL3/HLA-C2 but performs particularly poorly for KIR2DL1/HLA-C1. We attributed this to the fact that KIR2DL2/3 are significantly more permissive to HLA-C2 for the range of peptides considered in the extensive analysis by Sim et al. (8) than KIR2DL1 is to HLA-C1, which only seemed to bind 1 of the 361 P7P8 combinations. The very low tolerance of KIR2DL1/HLA-C1 therefore offered little information for the model to learn from.

Fig. 3. Noncanonical interactions, P8 molecular size, and structural information.

(A) Performance of the regressor model on the noncanonical interaction KIR2DL2/HLA-C2. The predicted BA is plotted against the true avidity. The blue and orange data points are the result of the smaller MLP #1 and larger MLP #2, with sizes (500) and (800, 800, 400, 200, and 100), respectively. (B) The $R^2$ value for all noncanonical interactions is plotted as a performance metric. (C) The overestimate ratio as described in the Supplementary Materials is plotted for heavy (>Thr) versus all P8 data points. (D) The purely sequence-based model ProtT5 performs at least as well as sequence-structure models where structural information is appended to ProtT5.

One feature of the peptide-specific binding that was particularly noticeable in the dataset from (8) was the sharp dependence on the molecular size of the residue at peptide position P8. Previous literature has established the fact that for KIR2DL2/3 binding HLA-C1, the binding affinity decreases as the size of the residue at peptide position P8 increases, where amino acids larger than Val completely turn off the receptor binding (5, 44–46). Measurements in (8) showed a sharp decrease in the BA averaged over all P7 residues as a function of P8 molecular weight beyond Thr (fig. S1A). This significant restriction on residue size has been attributed to the close proximity of the KIR Gln⁷¹ residue (44) and the HLA-C1 Ser⁷⁷ (8). This size restriction is not exhibited by the P7 residue (fig. S1B), where no such sharp transition occurs and is also absent in KIR2DL1 binding HLA-C2 (fig. S1C). This prompted us to investigate whether this feature is accounted for in our model through the PLM embeddings. If this is not the case, the model would be overestimating the data points with heavy P8 residues at a higher rate than the rest. We performed a t test to compare the k-fold mean of the overestimate ratio (the fraction of data that is overestimated by the model, see fig. S2) for the subset of data classified as having heavy P8 to the overestimate ratio of the full dataset. The results are shown in Fig. 3C. While the heavy P8 mean is consistently larger than the overall mean for every run of the model, the obtained P value remains larger than 0.05, and thus the discrepancy is not statistically significant. This indicates that the PLM embeddings are mostly able to capture such a structural phenomenon.

For a more rigorous determination of the extent of structural information contained in sequence-based PLMs, we compared the performance of the model with purely sequence-based input to the model with structural information added using a recently developed method (Foldseek) to encode the three-dimensional protein structure information into a one-dimensional sequence of twenty letters (3Di) in close analogy with true amino acid sequences (47). This allows the incorporation of the three-dimensional structure information with the PLM, which itself is trained on true amino acid sequences. The specific method of incorporating the two inputs into the model is varied across the literature. While Sledzieski et al. (48) appended the 3Di to the PLM embeddings using simple one-hot encoding (OHE) (Fig. 1B), converting the categorical amino acid identities into a binary column matrix, Heinzinger et al. (49) fine-tuned the ProtT5 PLM to create full embeddings of the 3Di sequences, which are then concatenated to the true amino acid sequence embeddings. The concatenation step in the OHE increases the embedding size from 1024 to 1044. The second method (ProstT5) is similar but replaces the OHE by passing the 3Di “amino acid–like” sequence through a PLM and results in a 2048D embedding. We tested both methods by first running AlphaFold2 structure prediction (50) of the KIR2DL1/HLA-C2 dataset and translating the output PDB file into the 3Di sequence. The performance of the model did not significantly change when the structural information is added (Fig. 3D). This suggests that PLMs for KIRs, peptides, and HLA-I already encode the structural information relevant to the function of protein binding.

The similarity in binding pattern of KIR2D and KIR3D is highlighted in the moderately successful cross-dataset testing between the two receptors

To investigate how well KIRLinguist captures the similarities between distinct KIR-HLA pairings, we trained the model on the entire inhibitory KIR2D dataset and tested it on the KIR3D binding HLA-B*57:01 and HLA-A*24:02 (Fig. 4, A and B). As expected, the model performance dropped significantly compared to training and testing on the same dataset, $\text{AUROC}=0.66$ and 0.60, respectively. However, intriguingly, the model still performs significantly higher than a random classifier ($\text{AUROC}=0.5$), indicating meaningful similarity in the binding function of KIR2D and KIR3D receptors. This decrease in performance is likely related to differences in the direct interaction of the KIR molecules with the HLA and peptide mediated by the D1 and D2 domains. Since the two receptors additionally differ in the presence or absence of KIR3D’s D0 domain, a closer investigation using feature analysis of the model’s learning behavior in such cross-dataset testing could shed light on the role of D0 in KIR3D’s binding pattern.

Fig. 4. Cross-dataset testing KIR3D and KIR2D datasets.

ROC curves for the model’s performance when testing on KIR3DL1 binding (A) HLA-B*57:01 and (B) HLA-A*24:02, after training on the entire inhibitory KIR2D dataset from (8). The performance is higher than that of a random classifier, indicating similarity between the two receptors. Cross-dataset testing was also conducted within the inhibitory KIR2D dataset. Testing was performed on KIR2DL2/HLA-C1 after training on (C) KIR2DL3/HLA-C1 and on (D) KIR2DL1/HLA-C2. The blue and orange data points are the results of MLP #1 and MLP #2 as discussed in the main text. The performance is shown for the regressor model to emphasize the model’s overestimation of BA when training on KIR2DL1 and testing on KIR2DL2.

We also performed cross-dataset testing within the KIR2D dataset to investigate similarities between the receptors of the same family. As expected, the performance was much higher when carried out between the very similar KIR2DL2 and KIR2DL3 (Fig. 4C), which performed unexpectedly well with $R^2=0.70$ than between KIR2DL2 and KIR2DL1 (Fig. 4D), which expectedly overestimated the binding affinities of KIR2DL2 after training on the much more permissive KIR2DL1.

Prediction for activating KIR binding to peptide–HLA-I complexes and cross-dataset testing inhibitory and activating KIR

Activating KIR are known to be significantly more peptide selective than their inhibitory counterparts. The narrower range of permissive peptides for activating KIR means that generalizable sequence features are harder to discern, as was the case with some noncognate interactions like KIR2DL1/HLA-C1. As a result, the model performs poorly for such datasets. We tested the model on binding data for KIR2DS4 from (8) and for KIR2DS2 from (41). The KIR2DS4 data measured its binding to both HLA-C1 and HLA-C2, and the model performed poorly on both. The KIR2DS2 dataset displayed exclusively poor binders, making it difficult to test the model’s performance. An exception to these findings was KIR2DS1 binding HLA-C2, which was found to have generally higher BA (8), and the performance of the model on this KIR2DS1 dataset was significantly better (Fig. 5A), with an $R^2$ value of 0.74.

Fig. 5. Cross-dataset testing on inhibitory and activating KIR.

(A) Performance of the model when trained on KIR2DS1/HLA-C2 and tested on the same dataset. (B) Bar chart comparing the same-dataset testing performance of KIR2DS1/HLA-C2 from (A) (left bar) with cross-dataset testing where the model is trained on KIR2DL1/HLA-C2 (middle bar) and KIR2DL2/HLA-C1 (right bar).

The high sequence similarity between the activating KIR2DS1 and its inhibitory counterpart KIR2DL1 prompted us to investigate cross-dataset testing between the two receptors. However, performance was very low ($R^2=0.22$) when testing between these two receptors of similar sequences (Fig. 5B). This can be explained by the significantly distinct binding pattern between the two datasets. While KIR2DL1 is very broadly permissive to amino acid substitutions in positions P7 and P8, KIR2DS1 is much more selective (8). The binding pattern of KIR2DS1 resembles that of the more sequence-distant KIR2DL2/3, even displaying suppressed binding for larger P8 residues. We find this to be the result of a single KIR residue at position 91 with closest proximity to P8. While KIR2DS1 displays a lysine at that position and KIR2DL2/3 both display methionine, both bulky residues, KIR2DL1 displays the much smaller threonine, thus accommodating larger P8. Previous studies have repeatedly shown that this T91K mutation in activating KIR2DS1 reduces the binding affinity of KIR2DS1 to HLA-C2 compared to KIR2DL1 (51–55). While these studies attribute this mutation to an overall (peptide-independent) drop in binding, its proximity to the peptide P8 seems to play a specific role in KIR2DS1’s sensitivity to the P8 size. The AlphaFold3 structures (56) of these complexes are shown in fig. S3. We tested our model’s performance by training it on the KIR2DL2/HLA-C1 dataset and testing on the KIR2DS1/HLA-C2 dataset (Fig. 5B). This significantly improved the performance of the model with $R^2=0.62$.

Generating predictions for peptides induced by viral infections

We tested the performance of our model on two datasets for the viral peptides of HIV-1 and hepatitis C virus (HCV) (57, 58). The HIV-1 dataset contained 222 overlapping 10mer peptides that spanned the entire p24 Gag protein. The study investigated 721.220 cells expressing HLA-C*03:04, a member of the HLA-C1 group, and the effect of the viral peptides on HLA stabilization and further binding of KIR2DL2 and KIR2DL3 on the pHLA complex. HLA stabilization was measured by staining the peptide-pulsed cells with mouse anti-human HLA-C specific antibody (DT9), and seven peptides were found to have high fluorescence intensity. We identified all seven selected peptides as positive binders to KIR since their measured binding was higher than the positive control self-peptide GAVDPLLKL (GKL). The ROC curve for this dataset is shown in Fig. 6A.

Fig. 6. Performance for viral datasets.

ROC curves for (A) the HIV peptide dataset (57) and (B) the HCV dataset (58). (C) Training on KIR2DL2/HLA-C1 with self-peptides IIDKSGxxV from (8) and testing on the HIV viral peptidome. (D) Training on KIR2DL3/HLA-C1 with self-peptides IIDKSGxxV and testing on the HCV viral peptidome.

The HCV dataset came from a study by Altfeld and colleagues on the effect of HCV viral peptides on HLA stabilization and KIR binding (58). The study used HLA-negative 721.221 cells transfected with HLA-C*03:04 and screened 200 overlapping 15mer peptides covering the nonstructural protein 3 and core protein of HCV genotype 1. They found 31 peptides that stabilized HLA-C*03:04 and further selected the 10 that showed highest stabilization to test their effect on binding of KIR2DL3 to the pHLA complex. We identified all 10 as positive binders since all bound more strongly than the GKL control and the remaining 190 as negative binders. The ROC curve for this dataset is shown in Fig. 6B.

Last, since the viral datasets share the identities of the KIR2DL2 and HLA-C1 with the much larger dataset from (8), we were prompted to perform cross-dataset testing, where the model is trained on the binding measurements of the “self” peptides IIDKSGxxV, and tested on the smaller viral datasets. The model’s performance is shown in Fig. 6 (C and D) for the HIV and HCV datasets, respectively. The AUROC values averaged around 0.65.

The results indicate that our model performs well on viral peptide datasets, achieving consistent discrimination of KIR-binding peptides across both HIV-1 and HCV experiments. Overall, these findings support the model’s capacity to recognize functional peptide binders in diverse viral contexts, although further improvements may be necessary to reach higher predictive accuracy.

Model limitations

While the purely sequence-based approach worked excellently for inhibitory KIR datasets, there are a number of areas of potential improvement of our model. First, the general sequence similarity alone represented in the PLM embeddings could not fully explain the binding pattern of KIR2DS1 in our model. The fact that the binding pattern of KIR2DS1 correlated much better with KIR2DL2 (which shares 92.4% of its sequence with KIR2DS1) than KIR2DL1 (with 96.4% sequence similarity) suggests that the addition of certain structural information, as shown in fig. S3, might enhance prediction of KIR-binding to pHLA. Since KIR2DS1 shared characteristic binding properties with the more sequence-distant KIR2DL2 and KIR2DL3, such as preferentially binding peptides with small residues at position P8, while KIR2DL1 displayed no such behavior, appending more complex biological information focusing on the binding site to the purely sequence-based PLM embeddings might enhance the model’s performance. The shared selective preference for small P8 residues, arising from a shared bulky KIR residue at P91, suggests that the inclusion of structural information could be important for this case.

Furthermore, KIRLinguist has been trained on datasets acquired in different ways for the different KIR. For instance, for the inhibitory KIR, the KIR3DL1 model has been on a dataset of endogenously presented peptides (9), giving rise to variations in P7 and P8 on different peptide backbones. Conversely, for the 2 Ig domain KIR, the datasets have been derived predominantly from P7 and P8 mutations of single peptide (8). The quantity of input data also varies between the activating and inhibitory KIR.

In addition, we have found KIRLinguist to less accurately discern the binding differences between allotypes of a given KIR. For example, we have tested KIRLinguist’s prediction on the allotypes of KIR3DL1 (*005, *015, *086, and *114) from the dataset in (59), and while the dataset showed this KIR polymorphism to significantly affect the binding strength of their cognate ligands, KIRLinguist could not capture this variation (see text S2 and fig. S5). This ultimately stems from the fact that KIRLinguist was trained on a single allotype of KIR3DL1 (*001) (9). In another test, we also looked at KIRLinguist’s prediction on two allotypes of KIR2DL1 (*002 and *004). It is known that KIR2DL1*004 has a D2 domain that resembles that of KIR2DS1 and could therefore share some binding patterns with KIR2DS1. However, we found that KIRLinguist’s prediction of KIR2DL1*004 binding HLA-C2 correlated very strongly with KIR2DL1*002 ($R^2=0.99$) and less so with KIR2DS1*001 ($R^2=0.87$) (see text S3 and fig. S6). This again points to the PLM’s embedding correlating strongly with overall sequence similarity, while the effect on binding pattern requires a more focused determination of sequence similarity in more binding-relevant regions like near the binding site and at locations that can affect protein conformation.

DISCUSSION

We have developed a machine learning prediction tool, KIRLinguist, for identifying short peptides with a strong influence on KIR-HLA–binding interactions. Trained on publicly available binding datasets, our tool accurately predicts peptide-specific KIR-HLA–binding strengths, achieving accuracy rates of up to 90% for a subset of KIR family receptors (inhibitory KIR2D and KIR3D). KIRLinguist learns the binding patterns of the specified KIR, peptide, and HLA directly from their amino acid sequences, thereby eliminating the need for complex biological computations required by traditional approaches such as molecular dynamics, docking, and scoring functions. This sequence-based approach leverages PLMs, which have demonstrated strong performance and are now well established for predicting TCR binding to peptide-HLA complexes with high accuracy (30, 31).

The TCR and KIR have distinct modes of binding to peptide-HLA. In general, the TCR for HLA-I is highly peptide-HLA specific, with potential for cross-reactivity. However, the KIR are peptide-HLA selective, recognizing shared determinants on the HLA-I heavy chain and a range of peptides that share amino acid types at key positions, usually C-terminal -1 and -2, in the peptide chain (60). Thus, the KIR do not share the same exquisite peptide sensitivity of αβ TCR, and so a different model is required. Despite the rapid evolution of KIR, peptide selectivity is a feature retained by KIR that bind classical HLA-I molecules, which is also a key feature that distinguishes activating KIR from their inhibitory counterparts. The prediction tool we developed therefore has the potential to identify previously unknown disease-associated ligands and help to unravel some of the complexity of genetic association studies of KIR with the many different diseases in which they have been explored. In particular, a number of viral infections have been associated with KIR including HCV, HIV, and flaviviral infections such as dengue, Zika and Japanese encephalitis virus infections (61). Recent work has also highlighted that activating KIR can recognize a peptide ligand derived from XPO1, a protein that is commonly up-regulated in cancer (27). Thus, this tool could be used to screen databases of cancer-associated peptides, providing opportunities for targeted NK cell therapies.

The design of peptides with the capacity to modulate the cytotoxic response of NK cells is emerging as a promising candidate in therapeutic strategies against disease. Our results demonstrate that the model we have developed can serve as a practical framework to support both basic and translational efforts by facilitating the identification of peptide candidates with robust immunomodulatory potential. By providing a means to prioritize peptide sequences most likely to result in strong KIR binding, KIRLinguist offers utility for guiding experimental efforts and for informing the selection of peptide modulators to be advanced into clinical investigation

MATERIALS AND METHODS

Data collection

Data were acquired from (8) and (9) by scanning the color plots in the binding data figures and extracting the raw data through a custom-built Python script.

Protein language model

No fine-tuning of the pretrained ProtT5 model parameters was performed for the binding prediction task. Instead, the embeddings were input into an MLP neural network, and the hyperparameters, including choice of optimizer, learning rate, and hidden-layer size, were optimized as explained in Results. Given the relatively small size of the MLP, the training of the optimized MLP was performed on a M3-chip MacBook Pro.

Structure evaluation

The Alphafold2 simulations were run Nationwide Children’s Hospital’s Franklin Cluster on A100 GPUs. The simulations were performed on the KIR2DL1/HLA-C2 binding data as a representative of the inhibitory KIR canonical interaction dataset. The sequence of the KIR, HLA, and the 361 individual peptides were input into AlphaFold2, and the highest-ranked structure was output as a Protein Data Bank (PDB) file. The structural information in the PDB files were then encoded into Foldseek’s 3Di “amino acid” sequences for the KIR, HLA, and peptide. The 3Di sequences encoding structural information were then either one-hot encoded resulting in a 1044-dimensional sequence-structure embedding, or input into ProstT5 to obtain a 2048-dimensional embedding as explained in Results. The Alphafold3 simulations were run on the AlphaFold Server, and the resulting highest-ranked structures were visualized using ChimeraX (62).

Supplementary Materials

This PDF file includes:

Supplementary Text S1 to S3

Table S1

Figs. S1 to S6

References