Decoding CAR T cell phenotype using combinatorial signaling motif libraries and machine learning

Chimeric antigen receptor (CAR) costimulatory domains derived from native immune receptors steer the phenotypic output of therapeutic T cells. We constructed a library of CARs containing ~2,300 synthetic costimulatory domains, built from combinations of 13 signaling motifs. These CARs promoted diverse cell fates, which were sensitive to motif combinations and configurations. Neural networks trained to decode the combinatorial grammar of CAR signaling motifs allowed extraction of key design rules. For example, non-native combinations of motifs which bind tumor necrosis factor receptor-associated factors (TRAFs) and phospholipase C gamma 1 (PLCγ1) enhanced cytotoxicity and stemness associated with effective tumor killing. Thus, libraries built from minimal building blocks of signaling, combined with machine learning, can efficiently guide engineering of receptors with desired phenotypes.


Lentiviral Transduction and Sorting of Human T Cells
Pantropic vesicular stomatitis virus G pseudotyped lentivirus was produced via transfection of LentiX 293T cells (Clontech #11131D) with a pHR'SIN:CSW transgene expression vector and the viral packaging plasmids pCMVdR8.91 and pMD2.G using FuGENE HD (Promega, #E2312). Primary T cells were thawed the same day and after 24 h in culture, were stimulated with Dynabeads Human T-Activator CD3/CD28 (Life Technologies #11131D) at 25 μL per 1 × 10 6 T cells. At 48 h (day 2), viral supernatant was harvested via centrifugation at 500 G for 5 min, and the primary T cells were exposed to the virus for 24 h in a 6-well plate (pooled screens) or in 96-well plates (arrayed screens). Dynabeads were removed at day 5 post-T cell stimulation. For pooled screens, GFP+ T cells were sorted on day 6 post-T cell stimulation with a FACSAria II. Assays were performed 10 days after removal of Dynabeads.

Arrayed Screening
CARs were constructed as described above in Viral Vector Construction. An additional pooled CAR sub-library was constructed with enriched concentration of DNA corresponding to M1, M4, M7, M9, and M10 on the basis of their high proliferation, degranulation, and memory formation in the pooled screening assay. Pooled CAR library DNA was used to transform 5-alpha F' I q competent E. coli cells (New England BioLabs C2992H), which were then plated on LB/Carbenicillin. At 24 hours 384 colonies (288 from the unbiased library, and 96 from the high-performance sub-library) were picked and miniprepped, added to 96-well plates and sequence verified. Wells with failed sequencing results or unidentifiable sequences were removed from plates and the well contents were replaced with duplicates of nearby wells, TE buffer (for empty well controls), or standard costimulatory domain (4-1BB, CD28) controls. CARs containing 4-1BB, CD28, and ICOS costimulatory domains and M1-M13 were left in place, but excluded from the analysis.
One plate of the arrayed screen was tested in triplicate with three different T cell donors: once in the context of the four-plate arrayed screen, and twice independently to ensure that trends in the four-plate screen were representative and reproducible. Screening of neural network-inspired receptors was performed in triplicate technical replicates.
Primary human T cells transduced with CAR library constructs were mixed with Nalm 6 to reach 1×10 6 T cells per mL and 2×10 6 per mL Nalm 6 and centrifuged at 300g for 2 min. For day 3, 5, and 7 challenges with Nalm 6, 80 μL of co-cultured T cells and Nalm 6 were added to 120 μL of Nalm 6 at 2×10 6 per mL and centrifuged at 300g for 2 min.

Pooled Screening
Primary human T cells transduced with pooled virus for the CAR library were mixed with Nalm 6 to reach 1×10 6 T cells per mL and 2×10 6 per mL Nalm 6. For day 3, 5, and 7 challenges with Nalm 6, co-cultured T cells and Nalm 6 were centrifuged at 400g for 4 min and resuspended at 1×10 6 per mL in 1/3 current HTCM and 2/3 fresh HTCM.
For extracellular staining, samples were centrifuged at 500g for 5 minutes and resuspended in FACS buffer with 1:50 PE anti-human CD4 antibody, 1:50 BV421 mouse anti-human CD45RA antibody, and 1:50 AF647 mouse anti-human CCR7 antibody. After a 30-min incubation at room temperature, samples were washed twice, and resuspended in FACS buffer. Samples were sorted on a BD FACS AriaII.

Data Preparation
For the arrayed data, in addition to the positional information of the combinational motifs, the initial CAR T cell number is a variable which affects the experimental output. Both are inputs of the machine learning algorithms. We randomly split D2 and use 90% for training and 10% for test, (we repeated this splitting process until duplicate motif combinations were found either exclusively in the training sets or exclusively in the test sets) ensuring all the motif combinations in the test data are different from those in the training.
We used a one-hot encoding to input motif combinations into the machine learning algorithms. Each motif position was described by a vector of fifteen 0s, and one 0 in each vector was replaced with a 1 corresponding to the absence of a motif (replace the first 0 with 1), the presence of a motif (replace the 0 equal to the part number + 1 with 1), or the presence of CD3z (replace the 15th 0 with 1). During synthesis of the 3-motif library, several CARs with 5 motifs were created. Rather than discard the data for these CARs, we allowed up to 5 motif positions in the model, as well as CD3z, for a total of 6 vectors.
This allows for inclusion of a small number of CARs that contained more than 3 motifs.

Machine learning Framework
In this work, we used a Convolutional Neural Networks ( We also compared our methods with other widely used machine learning regression methods, such as k-nearest neighbor regression, linear regression, nearest neighbors, random forest regression, and gradient boosted regression. The CNN + LSTM neural network has the best performance and predictive power of the methods compared.

Selection of neural network Hyperparameters
We tuned the hyperparameters for layers in the neural networks to find optimal hyperparameters for the cytotoxicity and stemness datasets. The tuned hyperparameters include convolutional layers filters (10, 20, 50), kernel size (2, 3, 4, 5); LSTM layer units Hyperparameters were tuned as follows: We performed a grid search of hyperparameters and scored each parameter set by 10-fold cross validation of the training set. The best-performing 10 hyperparameter sets for each dataset (cytotoxicity or stemness) were selected using the K-fold averaging cross validation (ACV) method and used to train 10 neural networks whose outputs were then averaged ( Hyperparameters for final neural networks are available in supplementary tables.

Ensemble Method
Due to the stochastic nature of network initialization and dropout, as well as the availability of a limited training set, every neural network is unique in terms of the parameterization of the network connections (31, 32). To mitigate the potential impact of this issue, we implemented an ensemble decision method to obtain consensus prediction from ten identical neural networks.

Distribution analysis
Distribution analysis was performed in Mathematica (Wolfram). CARs were sorted by proliferation (lowest enrichment to highest enrichment), cytotoxicity (highest Nalm 6 survival to lowest Nalm 6 survival), or stemness (lowest %IL7R+/KLRG1-to highest %IL7R+/KLRG1-) and assigned percentiles from 0 to 100. Individual parts or motif analysis was performed by selecting all CARs that contain a given part of interest.
Pairs of parts or motifs analysis was performed by selecting all CARs that contain a given pair of parts. Position analysis was performed by selecting all CARs that contain a given part at a position of interest. Distributions for selected CARs were constructed using the HistogramDistribution functionality and smoothed by using the PDF (probability distribution function) functionality to calculate the probability from 2.5 th percentile to 97.5 th percentile in steps of 5. The mean and standard error of the mean for each distribution was calculated by repeating the above processing for each of 10 neural networks (for predicted array screen data) or for experimental replicates (pooled screen data).  M1 and M10 generate high cytotoxicity and stemness when combined such that M1 is distal from the membrane, but generate reduced cytotoxicity and stemness when not combined or when M1 is not distal from the membrane. F, One 96-well plate in the arrayed screen was tested for cytotoxicity using three donors to assess donor-to-donor variability.
cytotoxicity also reflects variability in T cell transduction efficiency.    Cytotoxicity (B) and IL7R and KLRG1 expression (C) were assessed on day 9 after 4 challenges with Nalm 6 target cells. D, T cells with and without CAR were pulsed three times with wildtype Nalm 6 and CD19-negative Nalm 6 to assess off-target killing. T cells were cultured for 48 hours after final Nalm 6 pulse to maximize killing. E, Tumor progression was monitored using bioluminescent imaging of Nalm 6 expressing the firefly luciferase ( Table S2. Hyperparameters for neural networks trained on stemness (%IL7Ra+/KLRG1-) data. Hyperparameters were selected by grid search as described in the methods.

Data S2. (DataS2ArrayTraining.csv)
Training data set from CAR library screening array data shown in Figure 2. Motifs are labeled 1 through 13 as shown in Figure 1C. Motif 17 corresponds to CD3z.

Data S3. (DataS3ArrayTest.csv)
Test data set from CAR library screening array data shown in Figure 2. Motifs are labeled 1 through 13 as shown in Figure 1C. Motif 17 corresponds to CD3z.