The adaptive immune receptors in a big data world

Abstract

Antibodies and T cell receptors are the molecular basis of immune memory, and have become the foundation for generations of clinically successful biologics. Decades of research have established diverse systems to observe, identify, discover, and optimize the adaptive immune receptors, both from natural and synthetic sources. Recent advances in high-throughput display, next-generation sequencing analysis, and machine learning data mining are accelerating our capabilities to characterize immune receptor repertoires, including based on functional properties. This commentary discusses common principles in modern immune receptor studies, with an eye toward the near-term horizon in large functional dataset collection and analysis.

The initial frameworks

The discovery of monoclonal adaptive immune receptors first became possible in the 1970s when Köhler and Milstein¹ reported the hybridoma method that could isolate single monoclonal antibody and T cell receptor (TCR) variants. Hybridoma technology led to an explosion of scientific knowledge and new translational applications for adaptive immune receptors. Hybridomas, and later single-cell reverse transcription polymerase chain reaction (RT-PCR),² allowed the first molecular glimpses into the vast sets of unique specificities that comprise our immune repertoires. Studies explored immunogenetics like V-gene usage, heavy:light and alpha:beta gene pairing, and somatic hypermutation, and statistics on the scale of hundreds of clones revealed important gene associations with particular immune response features.^3,4

While hybridomas and single-cell sequencing provided an important, although statistically limited, view of immune repertoires, the advances in PCR, cloning, and especially protein display joined forces in the 1980s and 1990s to rapidly accelerate progress in antibody and TCR affinity engineering against target proteins.⁵ However, it was not until next-generation sequencing technologies of the 2000s that more comprehensive sequence analysis could be performed on adaptive immune repertoires.⁶ With high-throughput sequencing, we began to discern individual adaptive immune receptors with precise molecular clarity in the context of large adaptive immune response networks. Those important technical advances of recent decades have now brought our community into the current age of big data collection for the adaptive immune receptors.

Parallels between antibodies and TCRs

B cells and T cells work synergistically in vivo, and they both rely on an adaptive immune receptor to achieve specific molecular recognition (antibodies and B cell receptors are encoded by B cells, and TCRs are encoded by T cells). Many common features are shared across B and T cell immune repertoires, including central tolerance mechanisms, clonal selection and expansion in vivo, and the need to select immune receptors with the capacity to respond to antigens while minimizing cross-reactive and self-reactive clones. Antibodies and TCRs differ substantially in their mechanisms of action, and distinct experimental platforms are needed to study and engineer adaptive immunity in the humoral versus cellular immune compartments. Antibody molecules are secreted by B cells, and both the soluble antibody protein and the B cells (expressing B cell receptors) are effector agents. In contrast, TCR functions are intricately linked to the T cell that encodes them, and the T cell is the actual effector agent of TCR-driven responses. Affinity and multivalency dynamics of these two classes of immune receptors are profoundly different. Antibodies are expressed as various molecular classes that can present different numbers of binding arms: membrane-bound B cell receptors are multivalent, soluble IgG and IgE are bivalent, soluble IgA is bi- or tetravalent, and soluble IgM is decavalent. Somatic hypermutation plays an important role for antibodies to attain sufficient affinity for function in soluble formats, in which bivalent or even monovalent interactions can lead to a bound state. In contrast, T cells can express thousands of membrane-bound TCRs on their surface, and TCR recognition often (but not always) relies on multivalency for sufficient avidity. The low-affinity, multiavid interactions of TCRs can discriminate between different peptides presented by the same major histocompatibility complex (MHC) molecule, and very high affinity TCRs are often deleted in central tolerance due to high cross-reactivity among closely related peptide-MHC (pMHC) epitopes. In contrast, high-affinity antibody interactions are often specific and allow soluble antibodies to bind foreign antigens at diverse epitopes, and to traffic those antigens to appropriate effector cells and proteins within the body.

B cell receptors can have potent functional activity by displaying captured antigen-derived peptides to CD4+ T cells, and a soluble antibody’s binding activity can also lead to functional protection, for example, by neutralizing a virus. Soluble antibodies can bind antigen and activate immune pathways via Fc effector functions, which are modulated by Fc glycosylation state.⁷ In contrast, TCRs have no soluble expression format in vivo, and TCR-based immune protection is dependent on cellular responses that occur following activation signals from the surface-expressed TCR. T cell responses are highly dependent on the T cell state, and T cell regulatory mechanisms include numerous checkpoints and exhaustion pathways to mitigate the severe autoimmune consequences that can occur as a result of TCR cross-reactivity with self-peptides. Some of these biological, functional, and structural differences between antibodies and TCRs are summarized in Fig. 1.

Figure 1.

Unique and common features of the adaptive immune receptors. The listed properties constitute important factors in the design of molecular sequencing, functional screening, discovery, and engineering studies for antibodies and TCRs. ADCC, antibody-dependent cell cytotoxicity; ADCP, antibody-dependent cell phagocytosis; APC, antigen-presenting cell; CDC, complement-dependent cytotoxicity; NGS, next-generation sequencing.

The two-chain nature of antibodies and TCRs

The common two-chain nature of adaptive immune receptors (heavy and light chain for antibodies, alpha and beta chain for TCRs) provides a diversity generation mechanism for immune receptors to recognize a wide variety of targets. Because the heavy:light and alpha:beta chains are encoded on separate messenger RNAs and derived from distinct chromosomes inside the cell, there is no physical connection between complementary chain pairs until quaternary structures are generated at the protein levelpost-translation. Due to the difficulty of sequencing multiple genes on separate chromosomes together, the earliest next-generation sequencing technologies analyzed heavy, light, alpha, and beta chains separately.⁶ However, repertoire-scale technologies must collect information on both chains together to fully analyze and understand mammalian adaptive immune receptor repertoires.^8–10

It appears so far that most heavy chain V-genes can successfully pair with most light chain V-genes in developing antibody repertoires,.^4,11 During clonal selection, somatic hypermutation enhances antibody affinity directly for antigen interactions, and also often by refining the heavy:light interfaces for improved recognition.^12,13 Recent evidence demonstrates that antibodies often utilize the same heavy:light pairings across multiple individuals to recognize common antigens, revealing a degree of functional and structural convergence in recognition modes.¹⁴ Similar features can be observed by TCRs that target the same pMHC.¹⁵ Because T cells do not somatically hypermutate, the specific alpha:beta gene combinations likely provide an even greater role in antigen recognition than has currently been observed for B cells.¹⁶ The development of technologies to efficiently study paired heavy:light or paired alpha:beta chains are discussed in detail in the following section.

Determining the targets of antibodies and TCRs can reveal the workings of adaptive immunity and identify immune receptors or drug candidates against a disease. Immune receptors have been discovered from individuals that efficiently prevent viral infections or cancers, or that exacerbate autoimmunity. New vaccines or molecular therapeutics can then be devised to treat or cure based on the molecular mechanisms of immune receptor action. Functional studies of paired immune receptors are being used to identify target antigens, which is especially common for TCRs directed against cancer cells.¹⁷ Antibody and TCR discovery is also commonly performed to identify new drug candidates. While native chain pairing information is very helpful for antibody drug discovery, it is not strictly required, and many antibodies have been discovered by phage display of randomly paired heavy and light chains. Still, native heavy:light chain pairing is often superior to random pairings, including both for performance and stability features.¹⁸ For TCRs, native alpha:beta pairings are extremely important to reduce the risks of off-target therapeutic toxicity, which was observed in some early clinical studies of TCR-based T cell (TCR-T) therapy.¹⁹ Natively paired alpha:beta TCRs can reduce toxicity risks because the natively paired TCRs have gone through central tolerance and are less likely to be broadly cross-reactive than artificial affinity-selected alpha:beta pairs, especially for autologous TCR-T applications in which defined TCR(s) are expressed in patient T cells.

Recently developed single-cell technologies have dramatically expanded the scope of alpha:beta and heavy:light pairs that can be sequenced, cloned, and analyzed.^10,20 However, numerous challenges still remain for immune repertoire analysis. Biological sampling is always a limitation, and it is generally impossible to collect complete organism-level human tissue samples for immune repertoire analysis. Even so, the most advanced large-scale technologies to sequence paired immune receptors can accommodate a maximum of ∼10⁷ input B or T cells, which is still much less than the ∼10⁸ T cells in a single human blood draw, and far short of the ∼2 × 10¹¹ B cells and ∼4 × 10¹¹ T cells in a single individual.²¹ And even if biological sampling were addressed and a technique became available to pair all of the immune receptors in a cDNA format from every immune cell in an organism, it would still not be possible to sequence the entire repertoire because current 2 × 300 bp sequencing platforms still max out around ∼4 × 10⁸ reads per run, and long-read sequencing technologies analyze only around 2.5 × 10⁷ reads per run. Thus we can still only observe a small fraction of human immune receptors directly, despite the tremendous recent advances in single-cell processing and immune receptor sequencing platforms. Modern chain pairing approaches still capture more complete genetic and functional information than was previously possible, and enable a deeper characterization of the population of immune cells that are readily accessible in most clinical studies.

Connecting sequence to function

Several approaches have been devised to survey the functional properties of immune receptors, beginning with the hybridoma technique that remains commonly used today. Protocols for single-cell RT-PCR after fluorescence-activated cell sorting into well plates have matured over decades into efficient, streamlined workflows.² These single-cell RT-PCR approaches were transformed by advances in recombinant antigen bait design that enable precise flow cytometric identification of desired antigen-specific B cells,²² and multivalent pMHC constructs to identify TCRs.²³ More recently, antigen barcoding methods have been established to detect immune receptor sequences and antigen binding profiles concurrently via single-cell sequencing,^24–26 and these techniques also can analyze full transcriptomes to link B or T cell state with antigen specificity. One limitation of these immune cell-based approaches is that the primary cells are nonrenewable, and thus each antigen specificity measurement is subject to single-cell measurement error.

An alternative to analyzing primary immune cells is to use a display platform to study immune receptor function. Display platforms connect an RNA or DNA sequence to the function of the protein that it encodes, with many established technologies that are useful for immune repertoire screening. Display platforms that have been used to screen antibodies including phage display, messenger RNA display, bacterial display, and yeast display.²⁷ Several antibody display formats have also been established, including scFv (single-chain fragment variable), Fab (fragment antigen binding), and full IgG formats. Each of these formats have been successfully implemented in multiple display platforms, and each display platform offers unique advantages and disadvantages. Phage display has proven to be a workhorse technology for the biomedical community, and it supports remarkable library sizes with extremely high diversity and screening throughput. However, phage display can be limited by substantial replication bias across rounds, the limitations of protein expression in bacterial systems, is not compatible with flow cytometry, and can only be used for affinity analyses (not functional analyses). Yeast display has been growing in use due to its robust nature, eukaryotic expression systems, low cost and low replication bias across rounds, compatibility with flow cytometry, and ample library sizes, although yeast are generally limited to proteins that do not require mammalian post-translational modifications and are mostly used for affinity-based screens. Most recently, mammalian display has been emerging as a popular platform. Improved mammalian gene library construction techniques have accelerated growth in mammalian display, which can express mammalian proteins in their native formats while also being compatible with functional reporter screening systems.²⁸ These advantages outweigh the higher cost and complexity of implementing mammalian cell screening for many research groups, and thus mammalian screening platforms are anticipated to become even more widely used in the coming years.

TCRs can also be incorporated into display platforms and screened in varied formats, although with somewhat more difficulty than for antibodies. Antibodies evolved in secreted and soluble formats, and the structural configurations for soluble expression support efficient use as fusion protein domains across many systems and display formats. In contrast, TCRs are always membrane bound in nature, and they can be more difficult to express than antibodies but still have been successfully displayed in phage, yeast, and mammalian cells.²⁹ Because mammalian gene manipulation has advanced quickly in recent years, it can be anticipated that the use of mammalian platforms for TCR display will also continue to expand.

Our group has established renewable display library screening techniques to circumvent the limitations of functionally mapping immune repertoires using primary cells. Across several studies, we generate paired chain gene libraries and clone them into display platforms to enable repeated screening experiments.^9,11,30–35 An advantage of renewable libraries is that they provide the ability to collect accurate functional data by analyzing repertoires with robust statistical coverage (e.g. 10-fold more tests than the known library size), which cannot be achieved with primary cells. These display-based assays can bin immune repertoires based on performance metrics like affinity, which requires testing different antigen concentrations across an entire affinity range for accurate measurements, and likewise cannot be performed efficiently with primary cells. Renewable screening platforms can evaluate cross-reactivity to antigen panels,^32,36–38 which is complicated in primary cells because the multiple antigens can compete with each other for cell surface binding. A renewable library screening approach also enables interrogation of cell populations that do not express high levels of immune receptors. Important examples are the plasma cells in spleen and bone marrow, which secrete the majority of antibodies in circulation but do not express surface B cell receptors and are therefore not compatible with standard fluorescent antigen-bait approaches.³⁴ Finally, a critical advantage of renewable immune receptor display formats is that the immune receptors can be screened for functional activity (e.g. pathogen neutralization, activation against a cancer cell, etc.), which cannot be achieved at a library scale with primary cells and recombinant antigen-bait probes.³⁹ Collecting paired immune receptor sequence libraries and cloning into display libraries thus enables direct analysis of functional immune receptor performance en masse—bringing us potentially one step closer to a more complete, large-scale understanding of immune receptor repertoires.

Challenges and opportunities in functional antibody repertoire analysis

While antigen affinity is an important and commonly used screening parameter, an antibody’s affinity is rarely the critical functional drug property. For example in antiviral antibodies, in vitro virus neutralization is often considered the most important functional property for passive antibody prophylaxis and to evaluate vaccine-induced protection. Other antibody effector functions like antibody-dependent cellular cytotoxicity, antibody-dependent cellular phagocytosis, and complement deposition can be important to protect against infections, especially for viruses in which breadth and potency trade-offs exist and non-neutralizing antibodies make subtantial contributions to immune protection (e.g. influenza).⁴⁰ Autoimmune antibodies can recruit complement as one aspect of disease pathogenesis,⁴¹ or even have receptor agonist or antagonist properties that cause disease.⁴¹ Receptor agonism or antagonism is the critical active function for many antibody drugs, and in some cases, the desired activity is anticorrelated with affinity.⁴² The trade-offs between screening for affinity and screening for functional performance have become even more important in multispecific and fusion proteins that use antibody domains, such as T cell engagers and affinity-tuned constructs. In multispecifics, each binding partner can have many possible variants with precisely known affinities, and yet still the ultimate combination of binding arms, linkers, and formats to obtain the desired functional activity is very difficult to predict.⁴³ Due to the reduced importance of affinity in drug discovery and protection studies, more large scale functional data collection is needed to accurately explore antibody repertoire performance and obtain the necessary data to support artificial intelligence (AI) and machine learning (ML) model training.

Functional screening efforts have been pursued for many years, but substantial progress is still needed. Most platforms are insufficiently flexible, have limited throughput, or require complex instrumentation that raises experimental costs such that repertoire-scale analysis is not feasible for most groups or cannot support large-scale data collection (ideally, >10⁵ data points). For example, the Beacon single-cell platform performs function-first screening, but it requires multiple days of on-chip culture and sequencing is limited to a few thousand antibody-secreting cells. Functional yeast reporter systems have been established as purpose-built systems, but are limited to yeast-based reporter mechanisms.⁴⁴ Fluorescence-activated droplet sorting (FADS) enables multicell capture of antibody-secreting cells and reporter cells, with commercial instrumentation recently becoming available.^45–47 However, FADS is restricted by the throughput of the on-chip droplet sorter, and by the statistical constraints of limiting dilution (governed by Poisson statistics) which require many empty droplets for each occurrence of a single antibody-secreting cell with a single reporter cell in the same droplet. To address these challenges, our group recently reported an approach to combine antibody secretion and functional reporter systems into the same cell, allowing for efficient screening of functional read-outs using soluble drug-like molecules secreted by mammalian cells.³⁹ Because the reporter cell and the secreting cell are one and the same, it becomes simple to break the droplets and sort secreting cells directly based on reporter gene expression. This approach easily enables the functional analysis of >20 million single antibody-secreting cells within a single afternoon. These and other functional antibody activity screening efforts have focused on receptor agonism/antagonism, T cell engager antibodies, and virus-neutralizing antibodies. Continued advances in large-scale functional screening will accelerate antibody repertoire analysis and directed evolution for drug discovery.

Challenges and opportunities in functional TCR repertoire analysis

TCRs present numerous complications for high-throughput display screening. TCRs generally have low-affinity, low-binding-energy interactions, with cross-reactivity often observed among common MHC protein contacts and closely related bound peptides.²³ TCR discovery thus presents unique challenges because low-affinity TCRs are prone to rapid dissociation as monomers, while substantial cross-reactivity also reduces the difference between on- and off-target affinities. TCRs are selected by the immune system to sensitively distinguish between different low-affinity contacts, and the activation dynamics are further regulated within each cell. T cells dynamically modulate surface expression of the TCR-CD3 complex and the co-receptors CD4 and CD8, which affects valency, avidity, and activation sensitivity. The membrane proteins CD5, CD6, and CD45 can also tune TCR activation, and the T cell activation state is controlled both by internal cellular feedback mechanisms and external immune signals.^48–51 This dynamic tuning helps to avoid damaging and systemic autoimmune responses, while preserving local T cell potency against inflammatory threats. The unique combination of avidity-based TCR selection and cell-based TCR signal modulation makes T cells impressively able to distinguish between self and non-self in vivo. Still, it remains unclear how to collect the necessary data to understand TCR-pMHC molecular interaction dynamics at a large scale, with an ultimate goal of accounting for both cell state and TCR function to predict T cell behavior in response to a given stimulus.

It is important to express TCRs multivalently (e.g. on mammalian cells) for display technologies because their low-affinity interactions often require multivalency to attain suitable avidity to bind and/or activate. Multimers such as streptavidin tetramers or dextramers are often used to enhance avidity and prolong dissociation times in cell-based sorting, although these affinity-based screens do not directly collect information in functional TCR signaling.²³ Not all binding TCR:pMHC interactions are able to activate T cells. TCR activation is often observed from low-affinity TCR interactions, and appears to be mediated by energy versus distance dynamics of the TCR-pMHC bond. The most potent TCR interactions have been observed for TCR:pMHC pairs where the bond energy increases with distance (termed “catch bonds”) and where CD45 is locally excluded from the TCR complex.^52–55 Mammalian T cell lines such as Jurkat and SKW3 have been established as display platforms to analyze both TCR affinity and activation, often through activation markers like CD69 expression, or NFAT or NF-κB reporters.^33,35,56,57 These cell lines express CD3 as part of the TCR complex, and can also express CD8 or CD4 coreceptors to study TCR activation against cells presenting MHC class I or MHC class II, respectively.

One common technique to identify antigen-specific TCRs uses primary T cell proliferation or cytokine expression after pulsing peripheral blood mononuclear cells with antigen, and analyzing the activated T cells after culture.⁵⁸ These coculture assays identify certain populations of antigen-specific T cells, but are limited by the responsive capacity of primary T cells initially placed in culture. Some primary T cells are less sensitive and in an exhaustion state, whereas others are prone to sympathetically expand in a nonspecific manner when placed in a stimulatory environment. In many cases, the T cells relevant to specific disease states cannot be detected with peptide pulse expansion assays. T cell exhaustion is especially prevalent in cancer-specific T cells, for which discovery technologies that do not rely on antigen-specific primary cell proliferation are urgently needed. Another important challenge is to identify the pMHC antigens presented in disease-relevant contexts—for example, certain autoimmune peptides, or cancer antigens, which can include neoantigens, tumor-associated antigens, splice variants, and/or peptides containing post-translational modifications. Improved methods for antigen-specific TCR analysis and for disease-relevant peptide identification are needed to improve upon the limitations of current TCR-pMHC identification assays.

To help address these challenges, our group has established a series of assays for renewable display and screening of TCR repertoires.^33,35 These platforms support unbiased functional screening against desired pMHC tetramers or peptide-pulsed antigen-presenting cells. We first capture paired alpha:beta TCR chains from T cell repertoire, and clone them en masse into mammalian display and reporter cell lines. By transferring the TCR libraries into an activation-capable cell line for expression, we ensure that every T cell in the library is on “equal footing” with respect to activation state. We can screen libraries by both affinity³³ and activation³⁵ and analyze TCR reactivities against cancer and autoimmune model cell lines. One long-term goal for this work is to reveal the molecular characteristics of activating TCRs against pMHC antigens that drive human diseases, including for autoimmunity and cancers for which the antigen-specific TCRs and the pMHCs that they recognize are both insufficiently described.

Catalyzing new progress with big data

ML and AI are transforming the way we understand, analyze, and interact with our world. A few exciting areas that AI/ML could assist researchers and clinicians include (1) enhanceour basic sequence and structural understanding of adaptive immune interaction landscapes, (2) accelerate the design and discovery of new antibody- and TCR-based therapeutics, and (3) establish clinical biomarkers from immune receptor repertoire analyses to evaluate health, perform disease diagnosis and prognosis, and predict future disease risks. Successful AI/ML implementations nearly always occur in data-rich environments. Thus the key question for any AI/ML application is: do we have enough data already? And if not, how do we collect sufficient data for these algorithms to be successful?

Deeply mining immunized animal models to identify potentially optimal antibodies and mutations can be a rich data source suitable for ML,^60,61 and new library:library interaction data can also provide ample data points for effective discovery.⁶² However, ML’s greatest impact for immune receptors to date has been in the area of antibody developability. Several studies have compiled large datasets on various features that affect developability, which then supported model development with a reasonable degree of accuracy for developability prediction.⁵⁹

De novo immune receptor design against a desired target remains low efficiency (a small number of designed antibodies are actual binders), but is advancing rapidly. The best current antibody design technologies still require experimental screening due to relatively low hit rates, and often also require follow-up optimization. Most immune receptor design pipelines are built around large language models, diffusion models, and/or structural and protein:protein docking predictions. The high diversity and untemplated nature of the antibody CDR3 regions have always posed a challenge for structure prediction and docking algorithms, although with important recent progress.⁶³ The high affinity of antibody:antigen interactions provide a major boost for antibody computational design, due to the high binding energies of productive docking interactions. Still, the greatest remaining challenges are to design functional antibodies (agonists, allosteric modulators, and neutralizers), which can require dynamics that are not neatly captured in most structural studies or in structural training datasets. Functional antibodies can require binding to precise regions of an antigen that are difficult to target specifically by immunization or library panning, especially if a targeted epitope is conformation-dependent. Multispecifics present another major design challenge, and rely heavily on dynamics while having many more degrees of freedom for design than monoclonal antibodies. The field needs more richly annotated datasets to address these antibody design and evaluation challenges, with a focus on functional data combined with affinity data.

For TCR antigens, ML has been used extensively to mine immunoproteomic peptide elution data and refine peptide display predictions. Several reports have also explored AI/ML for TCR-pMHC interaction prediction,⁶⁴ and prediction accuracy is improving quickly. Computational TCR:pMHC interaction prediction may be more tractable than many antibody:antigen predictions due to the common binding epitope and restricted angles of approach for TCR docking, and the lack of somatic hypermutation that makes the TCR sequence space much smaller than antibody sequence space. However, TCR:pMHC structure prediction has 3 major challenges that hinder progress: (1) TCR:pMHC complexes have low affinity, which is more difficult to detect in energy-based docking models, (2) cross-reactivity is common and thus must be accounted for in designs, and (3) TCR activation potency is not only affinity driven, but also a function of bond energy versus distance. Due to the lower interaction energies and the importance of bond energy versus distance dynamics, extensive functional TCR training data may be required to design antigen-specific activating TCRs with minimal cross-reactivity. Known TCR structures are also biased toward higher-affinity TCRs, and additional structures of low-affinity, activating TCRs are likely to accelerate progress. These unique challenges for TCR structure and docking prediction are counterbalanced by the relatively lower diversity of the TCR and their common recognition modes (and thus more limited degrees of freedom for TCR:pMHC interactions) that can help support sufficient dataset collection and accelerate progress in TCR:pMHC prediction.

Certainly, more high-quality data are needed to address all these challenges. A relatively rapid way to collect large-scale affinity data is the use of library:library technologies,⁶⁵ which promise to fill out large sets of affinity information and are very useful for AI/ML training. The ability to obtain large-scale functional information on antibody or TCR interactions, especially when combined with large-scale affinity data in parallel, will help researchers more finely discriminate between affinity/avidity interactions and functional interactions. In pursuit of these aims, our group is collecting large-scale data on immune receptor performance for both affinity^{30,32,36–38} and functional activity.^35,39 We hope that these rich datasets will catalyze new progress in AI/ML performance by combining functional data along with structure predictions. Pairing large-scale functional data with structure prediction will unite advances in experimental and computational techniques, and unlock future opportunities in functional immune receptor design and prediction.

The future of immune receptor technologies

The study and use of immune receptors will expand in coming decades as our scientific tools mature to control immune responses and design new drugs, and the costs of immune drug distribution and production continue to fall. For antibody therapies, treatment costs will reduce based on a combination of higher-potency drugs resulting from refined discovery pipelines, the use of extended half-life Fc modifications, and high-concentration formulations that enable infrequent and at-home subcutaneous delivery. Future improvements in nucleic acid drug delivery could also greatly simplify manufacturing and distribution of protein drugs. Extrapolating from the ∼200 currently approved monoclonal antibodies, the large pipeline of drugs in clinical trials, and the current rate of annual Food and Drug Administration antibody approvals, it is conceivable that we could have 1,000 or more antibodies approved by 2040. At some point in the coming decades, reductions in active product dose costs and formulation/distribution advances will shift antibodies from a specialty drug class into more of a commodity, priced perhaps like the upper tier of today’s over-the-counter drugs. AI/ML will eventually become broadly useful to design and predict antibody drug performance. Experiments will not be replaced or eliminated, but instead will focus on (1) large-scale data collection and model training to enhance problem-specific AI/ML accuracy; (2) efficient large-scale validation of computationally designed molecules; and (3) engineering and optimization of initially designed candidates. Immune repertoire analyses to measure broad clinical biomarkers will also be broadly adopted, but will advance more slowly than drug discovery due to the high prediction accuracy required and the tighter restrictions on clinical data collection, analysis, and interpretation compared with bench research. Still, immune repertoire analysis for diagnostics, prognosis, and treatment will build on its currently established base in minimal residual disease detection, and will develop fastest in therapeutic settings with strong patient-to-patient presentation differences and variable disease courses with high disease burdens, such as cancer immunotherapy and severe autoimmune diseases.

TCRs are ideally suited for personalized medicines, and are already thought to be an important in approved autologous ex vivo expanded T cell therapies. TCR-T cancer gene therapies can also express defined TCR(s) in patient T cells for effective treatments. However, the timelines are slow (often ≥6 weeks) and the cost of personalized T cell therapies remains high. With sufficient data, improved TCR design and selection will accelerate TCR-T treatment timelines and reduce costs for broader therapeutic use. TCR-T therapies are somewhat scalable on an individualized nature, as the TCR molecule can be delivered using manufacturing-friendly nucleic acids and the resulting active drug product (a TCR-T cell) can multiply within the patient, which reduces the need for repeated drug administration. Thus if the TCR-pMHC matching problem can be conclusively resolved with sufficient accuracy for the clinic, then cancer therapy and autoimmune disease therapy will be radically transformed with personalized, designer TCR therapeutics. At some point in the foreseeable future, TCR drugs will be computationally designed based on immune repertoire and disease markers (e.g. based on tumor biopsy data), and a nucleic acid formulation used to deliver TCRs could be available for a quick subcutaneous injection at a local pharmacy. With robust and reliable computational TCR-pMHC matching, TCR immune repertoire analysis will become a biomarker for autoimmune monitoring, cancer suppression or progression, viral disease susceptibility, and as a general marker of health state.

Continued scientific innovation will further blend antibody and TCR drug modalities. For example, soluble TCR domains will be used like antibody binding domains as components of multispecifics against pMHC antigens (e.g. T cell engagers). Antibodies will also be used like TCRs, for example in pMHC-specific antibodies, and of course in the various currently approved formats of chimeric antigen receptor (CAR) T cells. With sufficient functional and safety/developability prediction capabilities, personalized antibody therapies could become a reality. An important first application for personalized antibody design would be pMHC-specific antibodies, which may be a tractable problem due to the restricted degrees of freedom and common MHC binding contacts for antibody-pMHC interactions. Such antibodies could have a common framework for each MHC protein and vary only in the specific peptide contacts, and would have numerous high-value applications to support their initial development in autoimmunity and cancer treatment. In some ways, pMHC-specific antibodies could be simpler to design and control in the clinic than TCR-T because antibodies are transiently dosed as protein, rather than as cell-based TCR-T. Progress is occurring quickly in de novo designs for proteins that bind pMHC,⁶⁶ and designed antibodies and TCRs to precisely target pMHCs seem right around the corner.

As treatment costs continue to fall we may see expanded prophylactic uses for antibodies and TCRs. Studies of antiviral immune protection over the past 15 years have revealed extremely potent and broad antiviral antibody responses in a subset of patients; research into other diseases like cancers may reveal similar outstanding protective adaptive immune features in certain individuals. As costs for antibody and TCR drug delivery become competitive with traditional vaccines, we may begin augmenting our immune systems with some of these exceptional adaptive immune proteins or immune constructs, rather than vaccinate and rely on our own immune memory to control infectious diseases. For example, highly potent anti-flu antibodies with extended half-life modifications may be given in small annual doses prior to flu season, or even genetically programmed to be produced by our cells throughout the year. Anticancer antibody cocktails could be given to patients with a suitable risk profile on a prophylactic basis. Such expanded drug uses of immune receptors will only be possible by first addressing the safety barriers for interventions in healthy individuals, which have already been addressed in the several monoclonal antibodies currently approved for infectious disease prevention.

The coming years will see continuous expansion in our abilities to manipulate immune protectionand achieve fine-grained, molecular control of adaptive immune receptor functions. Personalized, large-data tools will guide patients and their doctors to support optimal immune health. Immune systems will be read and analyzed en masse, adaptive immune biomarkers will be tracked for diagnosis and prognosis, and new personalized drugs will be designed, all in close partnership with guided computational assistance. The tight union of wet lab data, clinical data, and computational analyses represents the most important theme of the coming decades as we accelerate explorations of the adaptive immune receptors.

Author contributions

B.J.D.: Conceptualization, Funding acquisition, Visualization, Writing—original draft, and Writing—review & editing.

Funding

This work was delivered as part of the MATCHMAKERS team, supported by the Cancer Grand Challenges partnership financed by CRUK (CGCATF-2023/100009), the National Cancer Institute (OT2CA297514), and the Mark Foundation for Cancer Research. This work was also supported by National Institutes of Health grants 1R01AI181684, 1R01AI192975, 1U01AI169587, R21AI166396, R21AI144408, R21AI143407, R21AI178021, and DP5OD023118, and by the Mark and Lisa Schwartz AI/ML Initiative, the American Cancer Society, the Gates Foundation, the Koch Institute for Cancer Research, the MIT Research Support Committee, the MIT Department of Chemical Engineering, and the Ragon Institute.

Conflicts of interest

B.J.D. is a co-inventor of patents related to immune receptor discovery technologies currently assigned to the University of Texas at Austin, the University of Kansas, Massachusetts General Hospital, and Massachusetts Institute of Technology.

Data availability

No new data are reported.