TCRmodel server can be used to model the unbound TCR structure and TCR structure in complex with peptide-MHC structures. Both Class I and Class II MHC complexes are supported. The main server page has three sections:

• TCR-pMHCI complex modeling
• TCR-pMHCII complex modeling
• Unbound TCR modeling

Users can switch between the tabs to choose the input forms for TCR-pMHC modeling and unbound TCR modeling. Example input sequences are provided at the bottom of each form for the user to test it out. Alternatively, users can upload a FASTA file with the input sequences, as detailed below.

Users can provide the input for the TCR-pMHC complex modeling in three ways:

• Enter the sequence in the text box
• Select the germline gene from dropdown menu
• Upload a FASTA file

TCR and pMHC sequences for the complex modeling can be submitted as amino acid sequences for alpha and beta chains encompassing at minimum the TCR variable domains. The peptide sequence can be submitted and the program will truncate the peptide if needed. Users can choose the MHC sequences for both Class I and Class II by selecting the gene from the dropdown menu.

The peptide sequence input needs to fulfill the length requirement. For Class I peptides, the length of the peptide can range from 8-13 (inclusive), excluding 12. For Class II peptides, the length of the input peptide should be greater than or equal to 9. Our modeling algorithm will first trim the Class II to 9 amino acid "core" peptide sequences. Only the core peptide of class II will be modeled. Our server will trim user-input MHC sequence to alpha 1 & alpha 2 domains for class I MHCs, and alpha 1 & alpha 2 domains for class II, prior to modeling.

TCR sequences for modeling can be submitted as amino acid sequences for alpha and beta chains, encompassing at minimum the TCR variable domains. The TCRmodel server will automatically identify germline genes in the input sequences, if there is an exact match in the database.

Users can opt for uploading a multiple FASTA file instead of filling the sequence fields manually. The FASTA file must contain the TCR:pMHC sequences in the following order:

• For TCR-pMHCI complex
1. TCRα
2. TCRβ
3. peptide
4. MHCα
Check example FASTA input
• For TCR-pMHCII complex
1. TCRα
2. TCRβ
3. peptide
4. MHCα
5. MHCβ
Check example FASTA input
• For Unbound TCR
1. TCRα
2. TCRβ
Check example FASTA input

Despite we provide a suggestion for the FASTA headers names in example files, we don’t require any convention.

Users can opt to skip the energy minimization step performed by Amber. This additional refinement step can remove steric clashes in the structure prediction without compromising the prediction accuracy.

Users can include or exclude pMHC PDB structures to be used as templates in the TCRmodel pipeline.

Users can also limit the use of TCR and pMHC PDB structures as templates by using a PDB release date cutoff.

Five models are generated. They are ranked in the order of "model confidence" from AlphaFold-Multimer (iptm+ptm). After modeling finishes, the result page includes an interactive viewer showing the top-ranked modeled TCR or TCRpMHC structure, and downloadable model coordinates for top-ranked or all predictions.

Modeling statistics are diplayed in the collapsible table under the viewer port. Confidence measurements provided by AlphaFold, including predicted local difference distance test (pLDDT), predicted TM-score (pTM), interface pTM score (ipTM), and model confidence are shown for all predictions. If the target is a TCR-pMHC complex, we provide a modified version of ipTM score, TCR-pMHC ipTM. The meaning of these scores are provided below:

• pLDDT, ranging from 0-100, is indicative of local structural accuracy
• CDR3α pLDDT, ranging from 0-100, is an average of the pLDDT of only CDR3α residues
• CDR3β pLDDT, ranging from 0-100, is an average of the pLDDT of only CDR3β residues
• pTM, ranging from 0-1, is indicative of overall topological accuracy
• ipTM, ranging from 0-1, is a topological accuracy assessment of the interface between all chains.
• TCR-pMHC ipTM, ranging from 0-1, is an assessment of topological accuracy measurement of interface between the TCR and pMHC
• model confidence, ranging from 0-1, is a weighted summation of pTM and ipTM scores

For all of the confidence measurements displayed, higher values correspond to higher confidence.

To provide guidance on model confidence interpretation, we show a scatter plot on the right hand side, displaying the relationship between model confidence (y-axis) and model quality assessed by DockQ score (x-axis), from our benchmarking test on 48 non-redundant TCR-pMHC complexes. Five models were generated for each TCR-pMHC complex test case, and are shown as datapoints on the scatter plot. These datapoints are colored by CAPRI accuracy criteria. Based on our benchmarking, we find that:

• models with model confidence >= 0.85 are likely good - the majority of models with model confidence at or above 0.85 (see blue dashed line on the scatter plot) are Medium or High accuracy by CAPRI criteria.
• models with model confidence <= 0.49 are likely not good - most models with model confidence below 0.49 (see green dashed line on the scatter plot) are Incorrect by CAPRI criteria.

The expandable templates info tab shows the TCR templates with corresponding PDB (with chain IDs), CDR3 sequences and Vα and Vβ germlines used for the modeling. The CDR3 and V germlines of the input is also presented.

If the output model presents an unusual TCR binding mode, a warning icon will appear on the right side of the model confidence together with a message informing the unusual binding mode to the user. To detect the unusual binding we estimate the docking and incident angles (see TCR3d help page for details) of each model and compare those angles with the ones of known TCR:pMHC structures. In a nonredundant dataset of 81 (MHC I) and 39 (MHC II) TCR:pMHC complexes, we found docking angles ranging from 16.1 to 127.8 and 25.5 to 136.9 and incident angles ranging from 1.3 to 48.8 and 0.7 to 43.2 for MHC I and MHC II, respectively. The docking and incident angles are also provided in the result JSON files.

Click "View all models" on the result page to view all five predictions in an interactive viewer port.