Harnessing the power of artificial intelligence to advance cell therapy.

Cell therapies are powerful technologies in which human cells are reprogrammed for therapeutic applications such as killing cancer cells or replacing defective cells. The technologies underlying cell therapies are increasing in effectiveness and complexity, making rational engineering of cell therapies more difficult. Creating the next generation of cell therapies will require improved experimental approaches and predictive models. Artificial intelligence (AI) and machine learning (ML) methods have revolutionized several fields in biology including genome annotation, protein structure prediction, and enzyme design. In this review, we discuss the potential of combining experimental library screens and AI to build predictive models for the development of modular cell therapy technologies. Advances in DNA synthesis and high-throughput screening techniques enable the construction and screening of libraries of modular cell therapy constructs. AI and ML models trained on this screening data can accelerate the development of cell therapies by generating predictive models, design rules, and improved designs.

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 ~2300 synthetic costimulatory domains, built from combinations of 13 signaling motifs. These CARs promoted diverse human T 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 that 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.

T cells selectively filter oscillatory signals on the minutes timescale.

T cells experience complex temporal patterns of stimulus via receptor-ligand-binding interactions with surrounding cells. From these temporal patterns, T cells are able to pick out antigenic signals while establishing self-tolerance. Although features such as duration of antigen binding have been examined, our understanding of how T cells interpret signals with different frequencies or temporal stimulation patterns is relatively unexplored. We engineered T cells to respond to light as a stimulus by building an optogenetically controlled chimeric antigen receptor (optoCAR). We discovered that T cells respond to minute-scale oscillations of activation signal by stimulating optoCAR T cells with tunable pulse trains of light. Systematically scanning signal oscillation period from 1 to 150 min revealed that expression of CD69, a T cell activation marker, reached a local minimum at a period of ∼25 min (corresponding to 5 to 15 min pulse widths). A combination of inhibitors and genetic knockouts suggest that this frequency filtering mechanism lies downstream of the Erk signaling branch of the T cell response network and may involve a negative feedback loop that diminishes Erk activity. The timescale of CD69 filtering corresponds with the duration of T cell encounters with self-peptide-presenting APCs observed via intravital imaging in mice, indicating a potential functional role for temporal filtering in vivo. This study illustrates that the T cell signaling machinery is tuned to temporally filter and interpret time-variant input signals in discriminatory ways.

Electrostatic Energetics of Bacillus subtilis Ribonuclease P Protein Determined by Nuclear Magnetic Resonance-Based Histidine pKa Measurements.

The pKa values of ionizable groups in proteins report the free energy of site-specific proton binding and provide a direct means of studying pH-dependent stability. We measured histidine pKa values (H3, H22, and H105) in the unfolded (U), intermediate (I), and sulfate-bound folded (F) states of RNase P protein, using an efficient and accurate nuclear magnetic resonance-monitored titration approach that utilizes internal reference compounds and a parametric fitting method. The three histidines in the sulfate-bound folded protein have pKa values depressed by 0.21 ± 0.01, 0.49 ± 0.01, and 1.00 ± 0.01 units, respectively, relative to that of the model compound N-acetyl-l-histidine methylamide. In the unliganded and unfolded protein, the pKa values are depressed relative to that of the model compound by 0.73 ± 0.02, 0.45 ± 0.02, and 0.68 ± 0.02 units, respectively. Above pH 5.5, H22 displays a separate resonance, which we have assigned to I, whose apparent pKa value is depressed by 1.03 ± 0.25 units, which is ∼0.5 units more than in either U or F. The depressed pKa values we observe are consistent with repulsive interactions between protonated histidine side chains and the net positive charge of the protein. However, the pKa differences between F and U are small for all three histidines, and they have little ionic strength dependence in F. Taken together, these observations suggest that unfavorable electrostatics alone do not account for the fact that RNase P protein is intrinsically unfolded in the absence of ligand. Multiple factors encoded in the P protein sequence account for its IUP property, which may play an important role in its function.

Suppression of conformational heterogeneity at a protein-protein interface.

Staphylococcal protein A (SpA) is an important virulence factor from Staphylococcus aureus responsible for the bacterium's evasion of the host immune system. SpA includes five small three-helix-bundle domains that can each bind with high affinity to many host proteins such as antibodies. The interaction between a SpA domain and the Fc fragment of IgG was partially elucidated previously in the crystal structure 1FC2. Although informative, the previous structure was not properly folded and left many substantial questions unanswered, such as a detailed description of the tertiary structure of SpA domains in complex with Fc and the structural changes that take place upon binding. Here we report the 2.3-Å structure of a fully folded SpA domain in complex with Fc. Our structure indicates that there are extensive structural rearrangements necessary for binding Fc, including a general reduction in SpA conformational heterogeneity, freezing out of polyrotameric interfacial residues, and displacement of a SpA side chain by an Fc side chain in a molecular-recognition pocket. Such a loss of conformational heterogeneity upon formation of the protein-protein interface may occur when SpA binds its multiple binding partners. Suppression of conformational heterogeneity may be an important structural paradigm in functionally plastic proteins.

Conformational kinetics reveals affinities of protein conformational states.

Most biological reactions rely on interplay between binding and changes in both macromolecular structure and dynamics. Practical understanding of this interplay requires detection of critical intermediates and determination of their binding and conformational characteristics. However, many of these species are only transiently present and they have often been overlooked in mechanistic studies of reactions that couple binding to conformational change. We monitored the kinetics of ligand-induced conformational changes in a small protein using six different ligands. We analyzed the kinetic data to simultaneously determine both binding affinities for the conformational states and the rate constants of conformational change. The approach we used is sufficiently robust to determine the affinities of three conformational states and detect even modest differences in the protein's affinities for relatively similar ligands. Ligand binding favors higher-affinity conformational states by increasing forward conformational rate constants and/or decreasing reverse conformational rate constants. The amounts by which forward rate constants increase and reverse rate constants decrease are proportional to the ratio of affinities of the conformational states. We also show that both the affinity ratio and another parameter, which quantifies the changes in conformational rate constants upon ligand binding, are strong determinants of the mechanism (conformational selection and/or induced fit) of molecular recognition. Our results highlight the utility of analyzing the kinetics of conformational changes to determine affinities that cannot be determined from equilibrium experiments. Most importantly, they demonstrate an inextricable link between conformational dynamics and the binding affinities of conformational states.

Ligand concentration regulates the pathways of coupled protein folding and binding.

Coupled ligand binding and conformational change plays a central role in biological regulation. Ligands often regulate protein function by modulating conformational dynamics, yet the order in which binding and conformational change occurs are often hotly debated. Here we show that the "conformational selection versus induced fit" distinction on which this debate is based is a false dichotomy because the mechanism depends on ligand concentration. Using the binding of pyrophosphate (PPi) to Bacillus subtilis RNase P protein as a model, we show that coupled reactions are best understood as a change in flux between competing pathways with distinct orders of binding and conformational change. The degree of partitioning through each pathway depends strongly on PPi concentration, with ligand binding redistributing the conformational ensemble toward the folded state by both increasing folding rates and decreasing unfolding rates. These results indicate that ligand binding induces marked and varied changes in protein conformational dynamics, and that the order of binding and conformational change is ligand concentration dependent.

Thermodynamic analysis of protein-ligand interactions in complex biological mixtures using a shotgun proteomics approach.

Shotgun proteomics protocols are widely used for the identification and/or quantitation of proteins in complex biological samples. Described here is a shotgun proteomics protocol that can be used to identify the protein targets of biologically relevant ligands in complex protein mixtures. The protocol combines a quantitative proteomics platform with a covalent modification strategy, termed Stability of Proteins from Rates of Oxidation (SPROX), which utilizes the denaturant dependence of hydrogen peroxide-mediated oxidation of methionine side chains in proteins to assess the thermodynamic properties of proteins and protein-ligand complexes. The quantitative proteomics platform involves the use of isobaric mass tags and a methionine-containing peptide enhancement strategy. The protocol is evaluated in a ligand binding experiment designed to identify the proteins in a yeast cell lysate that bind the well-known enzyme cofactor, ß-nicotinamide adenine dinucleotide (NAD+). The protocol is also used to investigate the protein targets of resveratrol, a biologically active ligand with less well-understood protein targets. A known protein target of resveratrol, cytosolic aldehyde dehydrogenase, was identified in addition to six other potential new proteins targets including four that are associated with the protein translation machinery, which has previously been implicated as a target of resveratrol.

Biochemical properties and biological function of a monofunctional microbial biotin protein ligase.

Biotin protein ligases constitute a family of enzymes that catalyze the linkage of biotin to biotin-dependent carboxylases. In bacteria, these enzymes are functionally divided into two classes: the monofunctional enzymes that catalyze only biotin addition and the bifunctional enzymes that also bind to DNA to regulate transcription initiation. Biochemical and biophysical studies of the bifunctional Escherichia coli ligase suggest that several properties of the enzyme have evolved to support its additional regulatory role. Included among these properties are the order of substrate binding and linkage between the oligomeric state and ligand binding. To test this hypothesized relationship between functionality and biochemical properties in ligases, we have conducted studies of the monofunctional ligase from Pyrococcus horikoshii. Sedimentation equilibrium measurements to determine the effect of ligand binding on oligomerization indicate that the enzyme exists as a dimer regardless of liganded state. Measurements performed using isothermal titration calorimetry and fluorescence spectroscopy indicate that, in contrast to the bifunctional E. coli enzyme, substrate binding does not occur by an obligatorily ordered mechanism. Finally, thermodynamic signatures of ligand binding to the monofunctional enzyme differ significantly from those measured for the bifunctional enzyme. These results indicate a correlation between the functional complexity of biotin protein ligases and their detailed biochemical characteristics.