SynNotch-Antigen Analysis

Folding

Methods and Preparation

We began by using AI-based folding algorithms such as Meta’s ESMfold to predict how synNotch folds on the membrane of the CAR T-cell. Using the amino acid sequences provided by Morsut et al. (2016) we were able to to achieve a high confidence in folding for the main binding domain (Figure 1). Using ESMfold, pdb files were constructed for the antigens by isolating the amino acid sequences that make up the ectodomains of each antigen and by implementing them into the ESMfold algorithms with the same assumptions made when creating the synNotch pdb file. We first used the following amino acid sequences to designate the sequence that represents the ectodomains annotated on experimentally confirmed PDBs. We plug the sequence as a string into alphafold’s python jupyter notebook [1,3] which runs on a a type of convoluted neural network that processes geometric information such as R-group coordinate frames and spatial patterns. We then relaxed the torsion angles of side-chains and backbones using pyRoesetta’s relax function to bring the predicted folding to its lowest energy state in order to prep it for the docking stage

Figure 1: Algorithmic flowchart for the creation of binding energies between synNotch and targeted antigens

General synNotch Folding

Because synNotch’s folding ability has already been experimentally verified through Dr. Lim's lab [2], the confidence level of its folding provides a benchmark that can be used for comparison with the folding confidence of HER2 and CD19. Figure 2 shows, in a qualitative manner, that the main binding areas of synNotch such as the extracellular ligand recognition domain and the transmembrane domain, have high folding confidence based on the AlphaFold prediction. This is indicated by the darker purple/blue shading in those regions, meaning a higher likelihood that the predicted structure matches the true native structure. In contrast, some peripheral loop regions of synNotch show colors of red and yellow shading, indicating lower confidence [4]. The high confidence in the core functional regions of synNotch that directly participate in binding increases certainty that the predicted model accurately reflects the real protein's binding properties. This provides a level of folding confidence to aim for in assessing the HER2 and CD19 predictions, especially in their binding sites.

The pLDDT curve for the predicted SynNotch structure displays relatively high accuracy across the majority of the sequence, with a mean predicted lDDT value of approximately 90. the predicted local distance difference test allows for the alphafold algorithm to create a confidence interval on the predictions of R-group placement through local distance differences of all atoms in a model [5]. The highest regions of the curve, indicating the most confident fold predictions, align with the main functional components like the ligand binding domain and transmembrane helix. Some dips in the curve correspond to loop regions with lower folding confidence; however, the overall scale of the plot indicates a high degree of confidence in the global fold. The high mean pLDDT, along with peaks in folding confidence in key binding regions, suggest that synNotch serves as a reliable benchmark for assessing the accuracy of the HER2 and CD19 predictions. Matching or exceeding the folding confidence observed in the experimentally validated synNotch structure should provide greater certainty in the utility of HER2 and CD19 for directing CAR T-cell specificity.

Figure 2(a): Folding ability of synNotch as a homo-oligomer under a greedy pairing strategy represented as a IDDT confidence structure

Figure 2(b): The graphical outputs showing how the predicted IDDT align with the position of residue and error of the aligned residue compared to its predicted placement. the alphafold algorithm had a high confidence

HER2 Folding

There was less uncertainty when investigating truncations of HER2 that have not been experimentally determined. Alphafold, in this case, acted more as a hypothesis engine than it did for synNotch, which adds a new layer of uncertainty to the algorithm. The amino acid seperation between the ectodomain and the rest of the HER2 protein is seen on the rcsb file [3]. The lower confidence near the ectodomain boundaries suggests potential inaccuracies in modeling the transition between the structured ectodomain and the remaining HER2 sequence. Specifically, key charged residues like Arg179, Arg192, and Arg699 exhibit weaker predictions that may alter the electrostatic complementarity with synNotch upon binding [4]. Since there is less certainty in the ectodomain truncation model, experimental validation will play a larger role in checking compatibility with synNotch, especially at the joining regions where AlphaFold shows lower confidence likely due to conformational flexibility distant from the core ectodomain fold.

Figure 3(a): Annotated HER2 ectodomain with highlighted side-chains. the structured was rendered using the same parameters for synNotch excluding the pair mode.

Figure 3(b): Graphical outputs for HER2 truncation shows higher success rate for predicted IDDT measurements from alphafold per residue with steep drop offs near N-terminus and C-terminus.

CD19 Folding

Unlike synNotch, which exhibited high overall folding confidence, the AlphaFold prediction for the CD19 ectodomain shows greater uncertainty, with volatile peaks and dips in the pLDDT plot. In particular, the ectodomain truncation boundaries at both the N-terminal and C-terminal ends display very low confidence, which suggests that AlphaFold struggled to model the conformational change from the ordered ectodomain to the adjoining flexible regions. Key residues like Asp34, Arg44, and Glu160 have been highlighted to have weaker predictions that may disrupt pairing with synNotch upon binding (Figure 4). With such uncertainty in the CD19 ectodomain model, it is difficult to assess compatibility with synNotch, especially regarding the transitional areas linking the structured ectodomain to the complete CD19 sequence. In contrast, synNotch was specifically engineered for ectodomain exchange with HER2 and CD19, implying synNotch binding relies more heavily on its own inherent structural stability rather than accuracy of the binding partner's ectodomain. Nevertheless, the volatility in the CD19 ectodomain prediction warrants future experimental determination to clarify binding affinity with synNotch.

Figure 4(a): CD19 Ectodomain template free folding using ESMfold. the structure shows moderate to low confidence in secondary structure.

Figure 4(b): As seen in the rendered image, the graphical outputs show high levels of error in predicting residue placements.

Docking

With the custom pdbs created, we tested synNotch's ability to bind and interact with the corresponding antigens. The AlphaFold PDBs were prepared by extracting the ectodomain regions and structurally aligning synNotch across models. Then, protein-protein docking was performed using the HDOCK server from the Huang lab between the synNotch ectodomain and each antigen ectodomain to predict their bound complexes [6-9]. Finally, the docked models were evaluated using scoring functions and intermolecular contacts to estimate the relative binding energies between synNotch and its antigens.

synNotch-CD19 Docking

The top models for synNotch docking were best fit with the outer portions of the synNotch binding domains with HDOCK protein-protein docking scores averaging -259.906 kcal/mol, indicating stable binding between the two ectodomains. Visualization reveals key charged interactions between Arg44, Asp184, and Glu224 of CD19 with complementary residues Glu7, Lys108, and Arg140 of synNotch. Literature validates this in vitro, with robust calcium signaling induced in T-cells expressing the synNotch-CD19 receptor upon CD19 antigen stimulation. The docked model provides a structural basis for this signaling, with the observed synNotch-CD19 contacts enabling dimerization to activate the intracellular domains [10]. Additionally, the high HDOCK score agrees with flow cytometry data showing firm binding between antiCD19-synNotch and CD19. We focused the main models with binding to the outer portions of synNotch. This allows for a more confident level in protein expression in vitro with a stable docking score allowing for higher probability of downstream signal amplification.

Figure 5: Depiction of Model 1 from the HDOCK server between synNotch (orange) and CD19 (yellow) pdb files created from alphafold.

PROTEINS	MODEL #	DOCKING SCORE (kcal/mol)	Confidence level
SynNotch-CD19	1	-331.26	0.9740
SynNotch-CD19	2	-260.64	0.9014
SynNotch-CD19	3	-258.90	0.8983
SynNotch-CD19	4	-258.07	0.8967
SynNotch-CD19	5	-252.79	0.8865
SynNotch-CD19	6	-248.76	0.8782
SynNotch-CD19	7	-248.28	0.8771
SynNotch-CD19	8	-247.75	0.8760
SynNotch-CD19	9	-246.47	0.8732
SynNotch-CD19	10	-246.14	0.8725

Table 1: Top 10 models of synNotch-CD19 docking

synNotch-HER2 Docking

The HDOCK simulation reveals a favorable docking score of -269.38 kcal/mol between synNotch and the HER2 ectodomain. Visualization of the docked complex shows key interactions mediated by residues K157, R198, and Y201 on synNotch with D560, E563, and D564 on HER2. This extensive charge-charge complementarity between positively charged lysine and arginine residues on synNotch and negatively charged aspartate and glutamate residues on HER2 enables strong electrostatic interactions and salt bridge formation across the protein-protein interface. Additionally, the aromatic tyrosine residue Y201 on synNotch participates in a cation-pi interaction with R564 on HER2. Overall, these synergistic electrostatic and aromatic interactions lead to high shape and chemical complementarity between synNotch and HER2, resulting in the stable docked complex.

This molecular-level interaction and tight binding between synNotch and the HER2 ectodomain translates downstream into amplified signaling output from synNotch. Literature precedent shows ligand-induced activation of synNotch leads to high expression of target genes driven by the synNotch transcriptional activation domain [11]. Therefore, we would expect binding of the HER2 ectodomain to act as a stimulating ligand that triggers synNotch activation and downstream signaling cascades. However, definitive evidence requires in vitro validation as the synNotch pathway has not yet been tested against the HER2 ectodomain truncation as an activating ligand. Proper protein folding and post-translational modifications may also influence synNotch-HER2 binding affinity. Nonetheless, the computational HDOCK results provide a promising starting point to guide future synNotch-HER2 wet lab experiments as well as proper translation of free energy calculations that can be translated into the mathematical model of synNotch’s signaling pathway.

Figure 6: Depiction of Model 1 from the HDOCK server between synNotch (orange) and HER2 (yellow) pdb files created from alphafold.

PROTEINS	MODEL #	DOCKING SCORE (kcal/mol)	Confidence level
SynNotch-HER2	1	-298.70	0.9514
SynNotch-HER2	2	-282.07	0.9335
SynNotch-HER2	3	-274.87	0.9240
SynNotch-HER2	4	-273.50	0.9220
SynNotch-HER2	5	-272.28	0.9202
SynNotch-HER2	6	-264.06	0.9073
SynNotch-HER2	7	-256.99	0.8947
SynNotch-HER2	8	-252.40	0.8857
SynNotch-HER2	9	-249.55	0.8798
SynNotch-HER2	10	-248.91	0.8785

Table 2: Top 10 models of synNotch-HER2 docking

Supplemental Information

References

1. Lin, Z., Akin, H., Rao, R., Hie, B., Zhu, Z., Lu, W., Smetanin, N., Verkuil, R., Kabeli, O., Shmueli, Y., dos Santos Costa, A., Fazel-Zarandi, M., Sercu, T., Candido, S., & Rives, A. (2023). Evolutionary-scale prediction of atomic-level protein structure with a language model. Science, 379(6637), 1123-1130. https://doi.org/10.1126/science.ade2574
2. Morsut, L., Roybal, K. T., Xiong, X., Gordley, R. M., Coyle, S. M., Thomson, M., & Lim, W. A. (2016a). Engineering customized cell sensing and response behaviors using synthetic notch receptors. Cell, 164(4), 780–791. https://doi.org/10.1016/j.cell.2016.01.012
3. Mirdita, M., Schütze, K., Moriwaki, Y., Heo, L., Ovchinnikov, S., & Steinegger, M. (2022). Colabfold: Making protein folding accessible to all. Nature Methods, 19(6), 679–682. https://doi.org/10.1038/s41592-022-01488-1
4. The PyMOL Molecular Graphics System, Version 1.3, Schrödinger, LLC.
5. Mariani, V., Biasini, M., Barbato, A., & Schwede, T. (2013). lDDT: a local superposition-free score for comparing protein structures and models using distance difference tests. Bioinformatics, 29(21), 2722–2728. https://doi.org/10.1093/bioinformatics/btt473
6. Remmert, M., Biegert, A., Hauser, A., & Söding, J. (2011). HHblits: Lightning-fast iterative protein sequence searching by HMM-HMM alignment. Nature Methods, 9(2), 173–175. https://doi.org/10.1038/nmeth.1818
7. Pearson, W. R., & Lipman, D. J. (1988). Improved tools for biological sequence comparison. Proceedings of the National Academy of Sciences, 85(8), 2444–2448. https://doi.org/10.1073/pnas.85.8.2444
8. Sievers, F., Wilm, A., Dineen, D., Gibson, T. J., Karplus, K., Li, W., Lopez, R., McWilliam, H., Remmert, M., Söding, J., Thompson, J. D., & Higgins, D. G. (2011). Fast, scalable generation of high‐quality protein multiple sequence alignments using Clustal Omega. Molecular Systems Biology, 7(1). https://doi.org/10.1038/msb.2011.75
9. Larkin, M. A., Blackshields, G., Brown, N. P., Chenna, R., McGettigan, P. A., McWilliam, H., Valentin, F., Wallace, I. M., Wilm, A., Lopez, R., Thompson, J. D., Gibson, T. J., & Higgins, D. G. (2007). Clustal W and Clustal X version 2.0. Bioinformatics, 23(21), 2947–2948. https://doi.org/10.1093/bioinformatics/btm404
10. Choe, J. H., Watchmaker, P. B., Simic, M. S., Gilbert, R. D., Li, A. W., Krasnow, N. A., Downey, K. M., Yu, W., Carrera, D. A., Celli, A., Cho, J., Briones, J. D., Duecker, J. M., Goretsky, Y. E., Dannenfelser, R., Cardarelli, L., Troyanskaya, O., Sidhu, S. S., Roybal, K. T., … Lim, W. A. (2021). Synnotch-car T cells overcome challenges of specificity, heterogeneity, and persistence in treating glioblastoma. Science Translational Medicine, 13(591). https://doi.org/10.1126/scitranslmed.abe7378
11. Hernandez-Lopez, R. A., Yu, W., Cabral, K. A., Creasey, O. A., Lopez Pazmino, M. del, Tonai, Y., De Guzman, A., Mäkelä, A., Saksela, K., Gartner, Z. J., & Lim, W. A. (2021). T cell circuits that sense antigen density with an ultrasensitive threshold. Science, 371(6534), 1166–1171. https://doi.org/10.1126/science.abc1855