Evaluating protein-protein interactions in AF3 predicted complexes: a PD-1 case study
The latest release of AlphaFold (AF3) has addressed some limitations of the previous version, AF2, potentially expanding the impact of AlphaFold models on drug discovery. We were curious to explore if protein-protein complex models generated by AF3 could effectively serve as starting points for predicting protein-protein interactions and identifying binding hotspots, thus facilitating drug design.
Using our software, 3decision®, we evaluated the accuracy of the AF3 models using Programmed cell death protein 1 (PD-1) complexes as a case study.
In this blog post, we discuss the results of our analysis, highlighting the strengths and limitations of the predictions.
Introduction
The release of AlphaFold2 in 2021 (1) marked a significant advancement in protein structure prediction and established a new important role of Artificial Intelligence (AI) tools in supporting structural biology and drug discovery.(2) Despite its substantial impact, AF2 had notable limitations,(3) such as being unable to predict protein complexes with other biomolecules and ligands, thereby restricting its application in studying protein-protein interfaces or small-molecule binding sites.
AlphaFold3 (AF3),(4) developed by DeepMind and Isomorphic Labs, addresses this limitation by predicting complexes with other proteins, nucleic acids, ions, and small molecules. This extends AF3's utility beyond single-protein predictions, showing promise for broader applications of AF-predicted structures in drug discovery.
In this blog post, we examined the complex of PD-1 with its native ligand, the protein Programmed death ligand-1 (PD-L1), and its predicted protein-protein interaction network. The complex described here was included in the AF3 training dataset, however we still got some unexpected findings.
Then, we extended our analysis to PD-1/antibodies complexes not included in the training, which you can find in our subsequent case study “How accurate are antibody-antigen predicted complexes?”
How accurate are antibody-antigen predicted complexes?
Starting from AF3 PD-1/antibody complexes not included in the training dataset, we assessed the ability of AF3 to agnostically predict protein-protein interfaces and epitopes.
Download our case study to discover:
Where AF3 models showed the biggest accuracy
Importance of post-translational modifications when generating models
Limitations of the AI-predictions
and more!
Given the high relevance of structural analysis of protein-protein interactions (PPI) in drug discovery, we evaluated AF3's accuracy in predicting protein-protein complex interfaces. Our investigation aimed to determine could AF3 models be used effectively to:
investigate protein-protein interactions
identify critical binding hotspots relevant to therapeutic development
Our analysis focused on complexes of Programmed cell death protein 1 (PD-1), a well-established target for cancer therapy.(5) Many experimental structures of PD-1 in complex with its protein partners and various antibodies are publicly available, making PD-1 complexes an ideal case study. Given the high relevance in drug discovery for the structural analysis of protein-protein interactions, we conducted here some tests to evaluate AF3's accuracy in predicting protein-protein complex interfaces and to determine whether these models could effectively be used for this aim.
Curious to know more about antibodies targeting PD-1 for cancer treatment?
Check out our Protein of the Month edition on PD-1 /cemiplimab complex.
PD-1/PD-L1 complex prediction evaluation
The PD-1 complex with its native ligand PD-L1 (Image 1) plays a crucial role in regulating the immune response. Several monoclonal antibodies currently in clinical use target the PD-1/PD-L1 complex interface region binding either to PD-1 or PD-L1, thus preventing their association. This makes the study of the PD-1/PD-L1 complex and its PPI network highly relevant for therapeutic development.
The high-resolution 3D structure of the PD-1/PD-L1 complex is known (Image 1; PDB: 4QZK),(6) and several literature studies have characterized hotspot regions within it.(7,8,9) This structural knowledge provided a robust foundation for our evaluation of protein-protein interaction on the AF3 prediction of this complex.
Image 1. 3D structure of the PD-1/PD-L1 complex (PDB: 4ZQK; (6) PD-1 in purple, PD-L1 in orange). The PPI network is represented as dotted lines, and residues involved in the interactions are represented as sticks. H-bonds are in blue, and polar bonds are in yellow. Water molecules are represented as red spheres. The picture is produced using the 3decision® software.
Since the experimental structure of the PD-1/PD-L1 complex was part of the AF3 training dataset,(4) we expected to observe a very high quality of the AF3-prediction at the protein-protein interface. We produced the AF3 model of the PD-1/PD-L1 complex using the AF3 public server,(4) and we registered it in the 3decision® database (compatible with both PDB and mmCIF files, experimental and in-silico models). We then displayed it using the AF-specific coloring scheme - based on the pLDDT confidence score - to assess the quality of the prediction (Image 2A).
Image 2. Evaluation of AF3-predicted model of PD-1P/PD-L1 complex. A) Structure is colored by the pLDDT confidence score of the prediction: blue for very high, light blue for high, yellow for low, and orange for very low confidence. The protein-protein interface in the zoomed-in panel shows very high confidence in the prediction for most of the residues at the interface. The picture was produced using the 3decision® software. B) Display of the Predicted Aligned Error (PAE) for the AF3-predicted PD-1/PD-L1 complex. PD-1 and PD-L1 chains are labeled respectively in purple and orange. The region of PD-1 that is expected to be in proximity to PD-L1 includes residues 20-150, while for PD-L1, the residue range is 20-230. The picture is produced with the PAE Viewer webserver.(10)
We observed that the local confidence score of the residues at the complex interface was very high (Image 2A, zoomed-in view). Since the PAE plot indicated high confidence in the prediction of the relative positions of the PD-1 and PD-L1 chains within their interface region (Image 2B), we considered this model suitable for protein-protein interactions analysis. The comparison with the experimental structure of the PD-1/PD-L1 complex (Image 3) further validated the accuracy of the structural conformation in the AF3-predicted structure.
Image 3. Superposition of the AF3 model with the experimental structure of PD-1/PD-L1 complex showing high accuracy of the prediction (PDB: 4ZQK; PD-L1 in orange, PD-1 in purple). The view is focused on the protein-protein interface. The picture was produced using the 3decision® software.
This initial assessment validated the use of the AF3 model of PD-1/PD-L1 for calculating PPIs. In 3decision®, we can calculate the PPIs between protein chains on the fly and visualize them in 3D (see Video below).
Analysis of interaction hotspots
Interaction hotspots
Interaction hotspots are the small subset of residues that are essential for binding and stability of the complex, accounting for the highest contribution to the binding free energy.(11) By targeting these key residues, small molecules or biologics can be designed to effectively interfere with protein binding, providing therapeutic benefits.
Some recent papers have extensively characterized the PPIs stabilizing the PD-1/PD-L1 complex and identified the interaction hotspots.(7,8,9) We compared the PPIs obtained for the PD-1/PD-L1 experimental structure and the AF3-predicted model, focusing on those hotspot regions.
1) The analysis of the first PD-1/PD-L1 hotspot region, composed of the residue pairs reported in Image 4A, revealed that the interactions documented in the literature (6,7) were correctly predicted also in the AF3 model (Image 4B): the PPI network obtained for the AF3 model (Image 4D) closely matched the experimental one (Image 4C), demonstrating the model's effectiveness in reproducing known experimental interactions.
Image 4. Analysis of the first hotspot region of PD-1/PD-L1 complex. The experimental structure of the complex is PDB: 4ZQK. The PD-1 chain and residues are purple, while the PD-L1 chain and residues are orange. The AF3 predicted model is colored by the pLDDT confidence score. A) List of residues involved in the analyzed PPIs. The subscript notation “L” before the amino acid indicates residues on PD-L1. Below the table is a legend of the PPI representations: H bonds as long-dashed blue lines, salt bridges as long-dashed yellow lines. B) Superposition of the PPI networks for the AF3 predicted model and the experimental structure, PDB: 4ZQK. C) PPIs calculated for the experimental structure 4ZQK. Residues involved in the PPIs are labeled and color-coded according to the chain: orange for PD-L1 and purple for PD-1. D) PPIs calculated for the AF3 predicted model.
2) When we analyzed a second hotspot reported in the literature (7) (residues reported in Image 5D), we found again that most of the PPIs observed in the experimental structure (Image 5A) could be observed in the AF3 model as well (Image 5B), with the only exception of the water-mediated interaction between the Ile134 on PD-1 and the LTyr56 on PD-L1 (Image 5E). While the experimental structure clearly shows this interaction thanks to the presence of the well-resolved water molecule (Image 5C), in the AF3 model, this water molecule is missing (Image 5B). Therefore, the LTyr56 does not form any contact with the Ile134, and instead, the LTyr56 lateral chain is predicted to be stabilized by an intrachain hydrogen bond with the neighboring LGln58. We can additionally notice that the confidence score for these residues is lower than for other residues at the interface.
Image 5. Analysis of the second hotspot region of PD-1/PD-L1 complex. The experimental structure of the complex is PDB: 4ZQK. The PD-1 chain and residues are purple, while the PD-L1 chain and residues are orange. The AF3 predicted model is colored by the pLDDT confidence score. A) PPIs calculated for the experimental structure 4ZQK. Residues involved in the PPIs are labeled and color-coded according to the chain: orange for PD-L1 and purple for PD-1. The water molecule is represented as a red sphere and labeled. B) PPIs calculated for the AF3 predicted model. Labels for the residues involved in the PPIs are color-coded based on the chain. The position of the water molecule observed in the experimental structure is indicated in the red circle. C) Focus on the water-mediated interaction between Ile134 on PD-1 and LTry56 on PD-L1 and the representation of the electron density map of the experimental structure. D) List of residues involved in the analyzed PPIs. The subscript notation “L” before the amino acid indicates residues on PD-L1. Below the table is a legend of the PPI representation: H bonds are represented as long-dashed light blue lines, and salt bridges as long-dashed yellow lines. E) Superposition of the PPI networks for the AF3 predicted model and the experimental structure 4ZQK. The red arrows point to the water-mediated interaction observed on the experimental structure, and the intra-chain interaction between LTyr56 and LGln58 observed for the AF3 model.
Conclusion
Overall, this analysis of the protein-protein interactions on the PD-1/PD-L1 complex proved that AF3 predicted very effectively the protein-protein interface and the PPI network, with only a few (expected) inaccuracies.
The absence of water-mediated contacts is expected for all AF3 models since the method does not predict the presence and position of water molecules. Even if the PD-1/PD-L1 complex was included in the AF3 training set, the residues involved in the water-mediated contacts were not modeled in the correct conformation. Due to the crucial role of water molecules in PPIs for protein-protein complex formation, this limit of AF3 models should be considered when using them for PPI prediction.
How accurate are antibody-antigen predicted complexes?
In the second part of our case study, we extend our analysis to PD-1/antibody complexes predicted with AF3, for which the experimental structure was not included in the AF3 training dataset, to assess the potential of AF3 in modeling antibody-antigen PPI interactions for complexes without any previous structural knowledge.
If you are curious about the accuracy of predictions, download the case study below.
References
1. Jumper, J., Evans, R., Pritzel, A. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021).
https://www.nature.com/articles/s41586-021-03819-2
2. Akdel, M., Pires, D.E.V., Pardo, E.P. et al. A structural biology community assessment of AlphaFold2 applications. Nat Struct Mol Biol 29, 1056–1067 (2022).
https://doi.org/10.1038/s41594-022-00849-w
3. Perrakis, A. & Sixma, T. K. AI revolutions in biology: The joys and perils of AlphaFold. EMBO Rep. 22, e54046 (2021). https://doi.org/10.15252/embr.202154046
4. Abramson, J., Adler, J., Dunger, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500 (2024). https://doi.org/10.1038/s41586-024-07487-w
5. Javed, S.A. et al. Targeting PD-1/PD-L-1 immune checkpoint inhibition for cancer immunotherapy: success and challenges. Front Immunol 15, 1383456 (2024). https://doi.org/10.3389/fimmu.2024.1383456
6. Zak, K.M. et al. Structure of the Complex of Human Programmed Death 1, PD-1, and Its Ligand PD-L1. Structure 23, 2341–2348 (2015). https://doi.org/10.1016/j.str.2015.09.010.
7. Lim, H., Chun, J., Jin, X. et al. Investigation of protein-protein interactions and hot spot region between PD-1 and PD-L1 by fragment molecular orbital method. Sci Rep 9, 16727 (2019). https://doi.org/10.1038/s41598-019-53216-z.
8. Carter, R. et al. Identification of the functional PD-L1 interface region responsible for PD-1 binding and initiation of PD-1 signaling. J Biol Chem 299, 105353 (2023). https://doi.org/10.1016/j.jbc.2023.105353.
9. Huang, D. et al. Computational analysis of hot spots and binding mechanism in the PD-1/PD-L1 interaction. RSC Adv. 9, 14944 (2019). https://doi.org/10.1039/C9RA01369E.
10. Elfmann, C. & Stülke, J. PAE viewer: a webserver for the interactive visualization of the predicted aligned error for multimer structure predictions and crosslinks. Nucleic Acids Res. 51, W404–W410 (2023). https://doi.org/10.1093/nar/gkad350.
11. Ofran, Y.; Rost, B. Protein–Protein Interaction Hotspots Carved into Sequences. PLoS Comput Biol 3, 1169-1176 (2007). https://doi.org/10.1371/journal.pcbi.0030119.