E. John Wherry

Invikafusp alfa (invika), a first-in-class, bispecific, dual T cell agonistic antibody selectively targeting Vβ6/Vβ10 T cell subsets, is being evaluated in START-001, a multicenter phase 1/2 monotherapy trial in participants with mostly immune checkpoint blockade (ICB)-resistant, antigen-rich solid tumors. Primary outcomes were safety/tolerability and recommended phase 2 dose (RP2D) in phase 1 and anti-tumor activity in phase 2. As of 12 March 2025, phase 1 completed dose escalation (n = 34) through dose levels of 0.01, 0.02, 0.04, 0.08, 0.12 and 0.16 mg kg⁻¹ without reaching maximum tolerated dose, and initial phase 2 enrolled 18 participants. Two participants experienced reversible dose-limiting toxicities, one at each of the 0.08 mg kg⁻¹ and 0.16 mg kg⁻¹ dose levels. The most common treatment-related adverse event was cytokine release syndrome, mainly grades 1 and 2. No treatment-related deaths or non-laboratory grade 4 adverse events were reported. At the optimal biological doses (OBDs; 0.08 mg kg⁻¹ and 0.12 mg kg⁻¹) in combined phase 1 and phase 2 cohorts, the overall objective response rate was 14.8% (4/27), including confirmed partial responses in immunologically cold tumors, such as microsatellite stable colorectal cancer and programmed death-ligand 1 (PD-L1)-negative non-small cell lung cancer. No objective responses were seen at doses lower than 0.08 mg kg⁻¹. Invika selectively and potently expanded peripheral and intratumoral Vβ6/Vβ10 T cells, of a predominantly memory-like phenotype, dose-dependently, peaking at the OBDs. Pharmacodynamic increases in interferon-γ and soluble CD25 followed a similar pattern. Integrated analyses from safety, pharmacokinetic, pharmacodynamic and preliminary anti-tumor data supported the selection of 0.08 mg kg⁻¹ as the RP2D. In conclusion, phase 1 of START-001 met its primary outcomes by demonstrating the acceptable safety profile of invika consistent with its mechanism of action and selecting an RP2D. Initial phase 2 results further corroborated the safety of invika and expanded its preliminary evidence of clinically meaningful anti-tumor activity in ICB-resistant solid tumors, supporting further evaluation in the ongoing phase 2. ClinicalTrials.gov identifier: NCT05592626.

Long COVID (LC) affects millions of people worldwide, yet its driving mechanisms remain incompletely defined and no treatment has been approved. Persistence of SARS-CoV-2 and reactivation of herpesviruses have been implicated in LC, but direct detection of these viral antigens in blood and tissues has been challenging. T cells are exquisitely sensitive to viral antigens, survey organs throughout the body, and can be readily sampled in peripheral blood. Using high-dimensional spectral flow cytometry and single-cell sequencing approaches with multiplexed peptide/HLA tetramers, we profiled virus-specific T cells from hundreds of people with LC and recovered controls. We identified a subset of people with LC that maintained elevated activation of CD8 T cells targeting SARS-CoV-2 spike and/or lytic antigens from Epstein-Barr virus, but not other viral specificities. Spike-specific cells in LC expressed genes linked to TCR stimulation, terminal effector differentiation, and exhaustion and demonstrated reduced functionality upon cognate peptide stimulation ex vivo . Thus, T cells from a subset of people with LC are likely chronically stimulated by persistent SARS-CoV-2 antigens. These findings establish the use of T cells as biosensors of persisting antigen in LC and other disease settings and may help develop new therapeutic strategies for LC.

Persistence of SARS-CoV-2 has been proposed as a biological driver of Long COVID, a disabling chronic illness with no proven treatments. We conducted an exploratory, placebo-controlled, double-blind, 2:1 randomized mechanistic trial (NCT05877508) of the SARS-CoV-2-specific monoclonal antibody AER002 in 36 participants who met the World Health Organization case definition of Long COVID. After baseline characterization, participants received a single infusion and were followed for 360 days. The primary endpoint was the PROMIS-29 Physical Health Summary Score (PHSS) at 90 days; secondary and exploratory endpoints included patient-reported and objective measures of physical and neurocognitive function as well as blood-, imaging-, and tissue-based biomarkers. While AER002 was safe and well tolerated, no significant differences in physical health, quality of life, objective measures of physical function or cognition, or blood-based biomarkers were demonstrated between the treatment and control arms. In a post-hoc analysis, participants with a lower baseline SARS-CoV-2 antibody level and higher drug exposure were more likely to perceive treatment benefit based on the Patient Global Impression of Change scale ( p < 0.05 for anti-S, S1, and RBD). Although AER002 was not efficacious in this proof-of-concept study, our findings could inform future trials using monoclonal antibodies to target viral persistence in Long COVID.

Large-scale plasma proteomics offers opportunities to characterize disease susceptibility and improve prospective risk prediction, but transferring proteomic predictors across datasets remains challenging because measured protein sets differ across cohorts, study phases and assay configurations. Here we developed a self-supervised protein-token Transformer that maps the proteins observed in each sample to a fixed-dimensional participant representation, allowing unavailable proteins to be omitted rather than imputed. Using plasma proteomic profiles from 53,014 participants in the UK Biobank Pharma Proteomics Project, we pretrained the encoder by masked-protein reconstruction and evaluated whether disease models developed from comprehensive 2,920-protein profiles could be reused with a predefined subset of 1,460 proteins. Across 144 diseases, median AUC was 0.679 with comprehensive coverage and 0.637 when the same encoder and disease models were applied to partial-coverage representations without refitting; retraining only the disease-specific models increased median AUC to 0.673. Under partial coverage, protein-token proteomic risk scores (ProRS) exceeded coefficient-truncated LASSO ProRS by a median paired AUC difference of 0.027 and were comparable to LASSO ProRS refitted using outcome labels, with a median difference of 0.003. Performance was relatively stable across cardiovascular-kidney-metabolic diseases but more heterogeneous across autoimmune diseases, while protein-token representations improved discrimination beyond clinical covariates for 10 of 12 focused diseases under partial coverage without refitting. These results support self-supervised protein-token representations as a strategy for building proteomic prediction models that remain usable across heterogeneous measurement settings.

In cytometry, the workhorse single-cell technology of clinical immunology, every study defines its own antibody panel and cell-type vocabulary, so a classifier trained on one cannot annotate the next. Immunologists instead annotate by manual gating, splitting one parent population at a time on a two-marker plot, down an expert-defined hierarchy. We introduce CytoGate-Bench, a benchmark that reformulates this per-step procedure as a zero-shot, panel-agnostic task for large language models. It comprises 23,646 expert-annotated instances re-curated from 11 public flow- and mass-cytometry cohorts spanning eight marker panels. Across six open- and closed-weight backbones, the strongest formulation draws one rectangular gate per candidate and falls within the range of trained, panel-specialized baselines. It degrades less under distribution shift. Walking the hierarchy stepwise outperforms predicting every cell type at once. Ablations trace the signal to the data distribution shape and curated marker priors. However, adding vision or a self-verification loop systematically tightens gates.

Chronic T cell stimulation in tumors and chronic viral infections leads to T cell exhaustion, a state of dysfunction. As LAG3 marks a prominent subset of exhausted T (TEX) cells in mid-to-late stage tumors, we generated a Lag3 lineage–tracing mouse model (Lag3iCreERT2Rosa26LSL-tdTomato) to fate map and characterize tumor-reactive LAG3+CD8+ TEX cells. In tumor-bearing mice, two distinct tumor-specific tdTomato+ CD8+ T cell subsets stratified by LAG3 surface expression (LAG3+tdT+ and LAG3−tdT+) exhibited contrasting anatomical distributions, functionality and transcriptional profiles, yet shared TCR clonotypes, suggesting a common origin. While LAG3+tdT+ CD8+ TEX cells were restricted to the tumor microenvironment and predominantly terminally exhausted, LAG3−tdT+ CD8+ TEX cells were progenitors that persisted in vivo and are required for anti-tumor immunity against a secondary tumor challenge. This study highlights TEX cell functional heterogeneity and plasticity and characterizes a unique fate-flexible LAG3−tdT+ progenitor TEX subset that drives an anti-tumor memory response, supporting antibody-based therapeutic targeting of LAG3+ TEX cells to unleash anti-tumor immunity and promote durability.

For rapidly mutating viruses such as influenza viruses and severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), immune memory recalled by antigenically drifted variants primarily comprises antibodies that cross-react to the priming strain rather than de novo elicited responses, a phenomenon termed original antigenic sin or immune imprinting. The composition and functionality of de novo responses elicited by variant exposures remain unclear. Here we isolated and characterized hundreds of recall and de novo neutralizing monoclonal antibodies after sequential exposures to SARS-CoV-2 variants in ancestral-imprinted humans. De novo variant type-specific antibodies used different V(D)J genes that were closer to germline sequence, potently neutralized future variants and targeted distinct receptor binding domain epitopes compared to ancestral cross-reactive (recall) antibodies. Nevertheless, neutralizing responses to the updated 2024–2025 booster were predominantly ancestral cross-reactive. These results reveal the distinct contributions of recall and de novo antibodies to a balanced immune response and underscore the benefit of updated booster vaccines, which augment both subsets.

The human immune system is composed of ∼30-50 distinct cell types, each of which can exist in different states of activation or differentiation. Indeed, the mammalian immune system has evolved to sense and respond to infections, cancers, injuries, and changes in tissue or host homeostasis (1). Moreover, an increasingly large fraction of approved drugs target the immune system directly, and/or cause immune changes (2–4). A key feature of the immune system is to store some of this information, for example as innate or adaptive immune memory (5). In addition, rewiring of immune network architecture induced by disease, environmental exposures, drug treatments, and/or chronological age allows the immune system to store information in the pattern of connections and activity across populations of immune cells. This ensemble information storage, in addition to changes to individual cells, functions as a major way the immune system encodes aspects of immune history and future potential. Genetic information can identify inherited risk alleles, but cannot capture the continual remodeling of the immune system shaped by exposures, infection, inflammation, therapy, and aging (6, 7). To define and use such ensemble immunotypes, we developed a self-supervised deep learning framework that transforms high-dimensional immune profiles into representations of immune health. MAESTRO (MAsked Encoding Set TRansformer with self-distillatiOn) encodes a set of cells from an individual into an embedding that captures immune cell population-level organization. Pretrained on 1,792 peripheral blood samples comprising over 418 million immune cells across 13 clinical diagnoses, MAESTRO learns immune fingerprints that are stable within individuals yet diverse across populations, states of health, disease, and treatment, providing a quantitative basis for comparing immune states across individuals and over time. These fingerprints capture immune architecture beyond coarse cell type proportions, enabling patient-efficient clinical prediction using simple task specific models. MAESTRO model embeddings retain a temporal dimension of immune history and potential, reflecting signatures of past exposures and baseline features that predict future immune responses. Finally, we demonstrate a translational precision immunotherapy application by testing this approach in metastatic Pancreatic Ductal Adenocarcinoma (PDAC), where pretreatment immune landscape circuitry maps enable patient stratification and therapeutic response prediction. Overall, we developed a large, attention-based model that captures deep network architecture of immune states through self- supervised representations of immune cytometry data as a reusable foundation for precision immunology, converting immune complexity into clinically actionable embeddings for diagnosis, monitoring, and therapy selection.