Bispecific T-Cell Engager Clinical Translation Is Not Predicted by Target Antigen Density Alone: A Field-Level Empirical Analysis of Joint Binding–Effector Context

1. Introduction

The bispecific T-cell engager (TCE) modality redirects autologous T cells against tumor cells through a single molecule with two binding arms: one engaging a target antigen on tumor cells, the other engaging the CD3 complex on T cells. Since the regulatory approval of blinatumomab in 2014 the modality has produced multiple additional approved drugs across hematologic malignancies — CD19, CD20, BCMA, GPRC5D — and, more recently, DLL3-positive small-cell lung cancer, and continues to attract substantial industrial investment with new molecular architectures and target classes being explored across the major pharmaceutical research portfolios.

Despite this investment, the translation record remains uneven. Several solid-tumor TCE programs have been discontinued, including those targeting EpCAM, CEA, P-cadherin, gpA33, and PSMA [12,13,14,15,16]. Selected hematologic programs have produced disappointing readouts after promising preclinical data, most prominently AMG 330 targeting CD33 in acute myeloid leukemia. New program terminations and disappointing clinical readouts continue to be reported. The pattern of continued aggressive development alongside continued program failures is consistent with a field operating without a unifying analytical framework that predicts which programs will translate.

A specific puzzle within this broader picture concerns the relationship between target antigen density on tumor cells and clinical or preclinical outcome. The literature contains, side by side, reports of strong within-target density-potency correlations on cell-line panels (Spearman r approaching $-0.91$ in one BCMA program [7]), reports of moderate correlations ($r^2=0.31$ for one PSMA program [6]), and reports of no correlation at all between target antigen density and cytotoxic activity, observed for different drug programs targeting BCMA, CD20, CD33, ROR1, and other antigens, in cell-line panels and primary patient samples and clinical correlative analyses including the registration trial of an approved drug [1,2,3,4,5,8]. Density variables that one published account treats as predictive, another treats as non-predictive within the same target class. Density variables that predict outcome on cell-line panels do not predict outcome in clinical correlative analyses of the same drug. The published record contains the apparent contradiction without organizing it.

A second observation about the published record concerns the effector axis and is methodological. Both target antigen density and effector-side biology are necessary for bispecific TCE function — in the limiting case, target cells lacking the antigen produce no cytotoxicity at any effector-to-target ratio, and target cells expressing the antigen at any density produce no cytotoxicity in the absence of T cells. Yet the published record samples the two variables asymmetrically. Target antigen density is reported across cell-line panels, primary patient samples, and clinical correlative cohorts. Effector-side variation is structurally absent from cell-line panels: the standard cell-line cytotoxicity assay fixes effector-to-target ratios at 1:1 to 10:1 — one to two orders of magnitude above the ratios encountered in patients (bone-marrow E:T in relapsed-refractory AML is approximately 1:80 to 1:1; tumor-infiltrating lymphocyte densities in solid tumors are commonly far lower) — and uses uniform donor T cells, holding effector availability and effector functional state outside the variation regime in which they would be informative. Effector-side variables therefore appear as predictors of outcome only in primary-sample assays at endogenous effector content and in clinical correlative analyses, where they have been reported as predictive. This asymmetry in the published evidence base reflects experimental design more than underlying biology.

This apparent contradiction suggests that the field may be observing a joint-axis structure rather than genuinely inconsistent biology — specifically, that target antigen density and effector-side biology (effector availability and effector functional state) operate as joint variables in determining cytotoxic potency and clinical outcome, with the rate-limiting variable transitioning between them depending on assay conditions and disease setting. Under this account, parallel univariate analyses of either variable would produce results that depend on where each individual study sits relative to the joint structure, generating the apparent inconsistencies observable in the published record. The contradictions would not be in the data; they would be in the analytical convention that treats jointly operating variables as separable.

To investigate this systematically, we assembled a verified dataset of bispecific TCE clinical-stage programs spanning the major target antigens, with primary-source verification of clinical outcomes, dose-response data where reported, target antigen expression characterizations, and disease-setting effector biology. We then read the dataset against the published primary-source record of within-target density-outcome reports for each program. The organizing framework is mechanistically motivated but is evaluated here at the level of the published empirical record. We therefore evaluated whether the verified primary-source record is more coherently organized by target antigen density alone, by effector-side biology alone, or by the joint target–effector context. Under the joint target–effector reading, neither target antigen density nor effector-side biology is expected to serve as a universal stand-alone predictor. Rather, each becomes informative only in relation to the other axis, clinical exposure, assay conditions, and disease context.

The systematic empirical pattern that emerges from the dataset is consistent with this joint binding-effector account along multiple independent dimensions. We do not claim that this account uniquely explains the record; alternative explanations remain possible and the paper does not attempt to rule them out. The claim made here is that the joint binding-effector account provides a coherent, internally consistent explanation of the primary-source record assembled here, on the criteria appropriate to a field-level structural argument: coherence, internal consistency, and consistency with the documented empirical pattern. The account also generates testable predictions for future paired antigen-density and effector-side analyses.

We are not aware of a unifying explanation along these lines having been proposed in the field’s published literature, and in any case none appears to have been adopted, as the field’s published analytical practice — including in the most recent and rigorous correlative studies from the registration trials of approved drugs — continues to report parallel univariate analyses of target antigen density and effector-side biology without an integrated framework. The paper’s contribution is to assemble the systematic empirical record, identify the structural pattern that runs through it, and propose the joint binding-effector account as a testable organizing explanation of that record. The verified dataset and the documentation of the no-correlation pattern are deposited as supplementary material so that the field can engage with the empirical material directly, independent of how it engages with the proposed explanation.

The central claim of this paper is therefore not simply that effector-to-target ratio matters, nor that target antigen density is irrelevant. Rather, the claim is that the field’s apparently divergent findings arise because TCE activity is governed by a joint binding-effector structure that is commonly analyzed through separate univariate correlations. In this structure, antigen density can be predictive when the binding axis is rate-limiting, but loses predictive value when clinical exposure and target engagement shift limitation to effector availability or effector functional state. This account reconciles positive density-potency correlations, null density-outcome findings, and positive effector-side associations without treating any of them as anomalous.

2. Materials and Methods: Database Construction and Verification

2.1. Scope and Inclusion Criteria

We assembled a dataset of bispecific T-cell engager programs that have advanced to clinical evaluation, prioritizing programs with reported Phase 1 or later clinical data sufficient to characterize dose-response and clinical outcome. Inclusion criteria required: (i) the molecule is a bispecific T-cell engager in the strict sense, with one arm engaging the CD3 complex on T cells and the other arm engaging a tumor-associated antigen on target cells; (ii) the program has reached at least Phase 1 clinical evaluation with publicly available primary-source documentation of dose escalation, response data, or program termination; (iii) primary sources sufficient to characterize the program’s clinical trajectory and at least one readout of target antigen expression on tumor cells were locatable through verified peer-reviewed publications, regulatory filings, or sponsor disclosures.

We excluded: bispecific antibodies that are not T-cell engagers (such as anti-angiogenic bispecifics targeting two soluble or vasculature-related antigens); bispecific antibodies engaging effector cells other than CD3+ T cells (such as NK cell engagers); chimeric antigen receptor (CAR) T-cell programs, which are conceptually related but mechanistically distinct; and programs for which primary-source documentation of clinical outcome could not be verified.

Terminology.

Throughout this paper, effector-side biology is used as the umbrella term covering the full set of T-cell-related variables relevant to bispecific TCE function. We use effector availability more narrowly for variables of effector abundance — effector-to-target ratio, T-cell counts, CD3:target ratio, and related measures of how many effector cells are present. We use effector functional state (or effector fitness) for variables of effector competence — exhaustion-marker expression, regulatory T-cell frequency and activity, checkpoint biology, and immune-infiltrated phenotypes. The structural argument of this paper turns on the joint behavior of target antigen density and effector-side biology as a whole; where the underlying primary sources distinguished availability from functional state, we preserve that distinction.

2.2. Search and Inclusion Process

A structured search for clinical-stage bispecific T-cell engager programs was conducted across PubMed, ClinicalTrials.gov, regulatory databases (FDA and EMA), sponsor disclosures, and reference lists of primary reports, for records available through 2025. The search covered the modality at the level of its standard descriptors and program names, the clinical-development trajectory of each candidate program, and the translational variables relevant to this analysis (target antigen expression, cytotoxicity readouts, effector-to-target ratio, T-cell content, T-cell functional state including exhaustion and regulatory T-cell measures). Reference-list snowballing from primary reports was used to identify additional programs as the search progressed. Candidate programs were filtered against the inclusion and exclusion criteria defined in Section 2.1. Every program identified that met those criteria was retained for verification and inclusion in the dataset. The search yielded sixteen Tier-1 records spanning eleven target antigens.

We define Tier-1 records as clinical-stage bispecific TCE programs for which: (i) the molecule meets the strict TCE definition (one CD3-engaging arm and one tumor-antigen-engaging arm with the mechanism of T-cell-redirected cytotoxicity); (ii) primary-source documentation of clinical trajectory (Phase 1 or later dose escalation, response data, or program termination) is available; (iii) at least one peer-reviewed or regulatory primary source provides target antigen expression data on tumor cells in the relevant disease setting. Programs that did not meet all three criteria were excluded as not meeting the Tier-1 definition; the principal exclusion reasons were absence of primary-source clinical trajectory data (programs at preclinical-only stages or with sponsor disclosures insufficient for clinical characterization) or absence of primary-source target antigen expression data on tumor cells in the disease setting. Bispecific antibodies that are not TCEs (such as anti-angiogenic bispecifics targeting two soluble or vasculature-related antigens, or bispecifics engaging effector cells other than CD3+ T cells) were excluded as outside the scope of this analysis. Search flow documentation including programs considered and the criterion that determined any exclusion is provided in the supplementary material (Supplementary Table S1).

2.3. Records, Fields, and Verification

The dataset comprises sixteen Tier-1 records spanning eleven target antigens (CD19, CD20, BCMA, GPRC5D, CD33, CD123, CLL-1, FLT3, EpCAM, PSMA, and DLL3) and a range of disease settings including B-cell malignancies (acute lymphoblastic leukemia, non-Hodgkin lymphoma, chronic lymphocytic leukemia, multiple myeloma), acute myeloid leukemia, small cell lung cancer, and multiple solid-tumor indications. For each record, we recorded: drug name, sponsor, target antigen, indication, molecular format, clinical development status, regulatory approval status, summary of dose-response data, target antigen expression characterization, and primary references with verified DOIs.

Each record was constructed from primary-source verification: for approved drugs, regulatory filings (FDA prescribing information, EMA SmPC) and pivotal-trial publications; for active or terminated investigational programs, peer-reviewed primary publications, ClinicalTrials.gov entries, and sponsor press releases or conference presentations as available. Where prior compilations of bispecific TCE programs reported information that could not be verified against primary sources, the information was excluded from the database.

2.4. Documentation of the Within-Target Density-Outcome Literature

Independently of the program-level database, we assembled a documentation of peer-reviewed primary sources reporting within-target relationships between tumor cell antigen density (or expression) and bispecific TCE potency or clinical response. Five primary sources were verified by direct reading of the source PDF: Pillarisetti et al. 2020 (BCMA cell-line) [1]; Cortes-Selva et al. 2024 (BCMA clinical correlative, $n=165$, MajesTEC-1) [2]; Mhibik et al. 2023 (CD20 primary CLL samples) [3]; van der Horst et al. 2021 (CD20 primary B-NHL lymph node samples) [4]; Gohil et al. 2017 (ROR1 solid-tumor cell-line panels) [5]. Each entry includes the verbatim author statement of the no-correlation observation, the methodology, and where authors quantified the observation, the reported statistic.

Three additional primary sources reporting within-target correlations of varying strength on cell-line panels were also verified: Friedrich et al. 2012 (PSMA, $r^2=0.31$, $P<0.0001$) [6]; Panowski et al. 2019 (BCMA, $r=-0.91$, $P=0.011$) [7]; Friedrich et al. 2014 (CD33, AMG 330, eleven AML cell lines) [8]. The Friedrich 2014 panel was reanalyzed using formal statistical methods on the published data; the reanalysis is presented in Section 3.2.

2.5. Quality and Quantity of the Dataset

The verified dataset assembled for this paper has several features that bear on the strength of the conclusions it can support. The cross-program inventory spans eleven distinct target antigens, eleven disease settings, and a range of molecular formats including small-format BiTE constructs, full-length IgG bispecifics, half-life-extended formats, and novel architectures. The within-target documentation spans five peer-reviewed primary sources covering four target antigen classes (BCMA, CD20, CD33, ROR1), six disease settings encompassing both hematologic malignancies and solid tumors, two distinct molecular formats, and three distinct readout methodologies (cell-line cytotoxicity panels, primary patient samples at endogenous or fixed effector ratios, clinical correlative biomarker analysis). The publication dates of the verified primary sources span 2017 through 2024, including correlative analyses from the registration trials of approved drugs.

Every entry in the database and every quotation in the supporting documentation has been verified against its primary source. Where prior compilations reported information that could not be verified, or where attributions in secondary sources proved incorrect on primary-source examination, the information was corrected or excluded. The verified dataset and the documentation of the within-target density-outcome literature are provided as supplementary material to enable independent engagement with the empirical record.

3. Results

The results are organized across three sections. Section 3.1 documents the systematic empirical pattern in within-target density-outcome reports across the field, including the within-program paired analyses that constitute the strongest direct evidence in the verified record. Section 3.2 presents within-program triangulation of a clinical failure (CD33/AMG 330) across cell-line, primary-sample, and clinical readouts. Section 3.3 presents the cross-program inventory of clinical-stage TCE programs and the pattern by which approved and discontinued programs are organized. Table 1 summarizes the four evidence streams the structural argument rests on, the documented pattern in each, and the role each plays in the argument; the sections that follow walk through each stream in turn.

3.1. The Empirical Pattern Across the Field

3.1.1. Within-Target Density-Potency Null Findings Are Reported Broadly

The peer-reviewed literature contains repeated reports that target antigen density on tumor cells does not correlate with bispecific TCE cytotoxic potency or clinical response. The pattern has been observed across distinct drug programs, distinct target antigen classes, distinct disease settings, distinct molecular formats, and distinct readout methodologies.

For BCMA-targeting bispecific TCEs in multiple myeloma, the pattern has been reported in two peer-reviewed primary sources covering the cell-line preclinical readout and the clinical correlative readout of the same approved drug. Pillarisetti et al. 2020 [1] reported, for teclistamab tested against multiple myeloma cell lines spanning a range of BCMA receptor densities (3,705 to 13,173 receptors per cell), that “the BCMA receptor number did not correlate with the cytotoxic activity of teclistamab.” Cortes-Selva et al. 2024 [2], the MajesTEC-1 correlative biomarker analysis at the recommended Phase 2 dose ($n=165$), reported that “neither frequency of baseline bone marrow BCMA expression nor BCMA-receptor density was associated with clinical response to teclistamab,” with reported Wilcoxon $P=0.7375$ for BCMA-positive frequency and $P=0.8272$ for BCMA-receptor density. The same cohort showed multiple effector-side biomarkers (spanning availability and functional state) significantly associated with both response and progression-free survival.

For CD20-targeting bispecific TCEs, the pattern has been reported in two peer-reviewed primary sources covering chronic lymphocytic leukemia and non-Hodgkin lymphoma settings. Mhibik et al. 2023 [3], examining epcoritamab against CLL primary patient samples ($n=56$), reported that “cytotoxic activity was independent of CD20 expression on CLL cells” and quantified the relationship as Spearman $r=-0.088$, $P=0.5597$ between baseline CD20 mean fluorescence intensity and cytotoxicity at day 7. van der Horst et al. 2021 [4], examining the same drug against primary lymph node tumor samples from patients with diffuse large B-cell lymphoma, follicular lymphoma, and mantle cell lymphoma ($n=39$), reported that “CD20 expression levels on tumor cells in CD20-naïve patient samples did not correlate with epcoritamab-dependent cytotoxicity” (Spearman $r=0.23$, $P=0.24$ across CD20 expression levels spanning approximately two orders of magnitude).

For ROR1-targeting bispecific TCEs in solid-tumor settings, Gohil et al. 2017 [5] reported across panels of pancreatic ductal adenocarcinoma cell lines and ovarian cancer cell lines that “T-cell-mediated killing did not correlate with ROR1 expression on target cells” for the pancreatic panel, and that they “did not observe a correlation between ROR1 expression and cytotoxicity” for the ovarian panel.

For CD33-targeting bispecific TCEs, the original Friedrich et al. 2014 publication [8] on AMG 330 against eleven AML cell lines described a qualitative trend toward higher activity at higher CD33 expression. We re-applied formal statistical analysis to the published data: across the eleven cell lines spanning CD33 receptor densities of 14,400 to 56,700 sites per cell (a 3.9-fold range), Pearson correlation between log CD33 receptor density and log EC₅₀ gives $r=-0.483$, $P=0.132$; Spearman correlation gives $\rho =-0.282$, $P=0.401$. A 10,000-iteration percentile bootstrap on the Spearman coefficient yields a 95% confidence interval of $[-0.80,+0.45]$, with 24% of bootstrap replicates producing a non-negative coefficient. The panel does not statistically support a within-target correlation by formal test. The reanalysis is detailed in Section 3.2.

The breadth of this pattern is itself empirically informative. The no-correlation finding has been reported across four target antigen classes (BCMA, CD20, ROR1, CD33), in both hematologic and solid-tumor disease settings, in cell-line panels and primary patient samples and a clinical correlative analysis from a registration trial cohort of 165 patients, in both small-format BiTE constructs and full-length IgG bispecifics, and across publication dates spanning 2017 through 2024. The pattern is not specific to any individual drug, target, disease, format, or methodology.

3.1.2. Within-Target Correlation Strengths Reported in Cell-Line Panels Are Heterogeneous

The same field’s published record also contains reports of within-target density-potency correlations on cell-line panels with widely varying strengths. Friedrich et al. 2012 [6] reported $r^2=0.31$, $P<0.0001$ for a PSMA-targeting BiTE across human prostate cancer cell lines. Panowski et al. 2019 [7] reported $r=-0.91$, $P=0.011$ for a BCMA-targeting bispecific (a different drug and different panel from Pillarisetti 2020) across multiple myeloma cell lines. The Friedrich 2014 CD33 panel discussed above does not support a correlation by formal test. Pillarisetti 2020 and the no-correlation reports above similarly do not.

Reports of within-target correlation strengths in the published cell-line literature therefore span a range from no statistical support for any correlation up to coefficients approaching $r=-0.91$. The heterogeneity is not organized by target class (BCMA appears at multiple correlation strengths in different reports), by disease setting, or by molecular format. The published literature presents these reports as local observations about specific drug programs without assembling them into a unified analytical framework that would predict when within-target correlations should be present and when they should be absent.

3.1.3. Within-Program Paired Analyses Demonstrate the Pattern in Single Studies

The strongest empirical evidence in the verified record comes from studies that examined both the binding-axis variable (target antigen density on tumor cells) and at least one effector-side variable in the same study population, with both readouts available for the same patients or samples. In each such study available in the verified record, the binding-axis univariate analysis produced a null result while the effector-side univariate analysis produced a positive result.

Mhibik et al. 2023 [3] examined both axes on the same 56 CLL patient samples. The CD20 expression versus cytotoxicity correlation gave Spearman $r=-0.088$, $P=0.5597$. The effector-to-target ratio versus cytotoxicity correlation, on the same 56 samples, gave Spearman $r=0.60$, $P<0.0001$. The contrast was reported as a central finding of the paper and summarized in the abstract Key Points: “Epcoritamab-mediated killing of CLL cells by autologous T cells correlates with the effector-to-target ratio but not CD20 expression.”

Cortes-Selva et al. 2024 [2] examined both axes on the same 165 patients in the MajesTEC-1 RP2D cohort. The BCMA-receptor-density versus response analysis gave Wilcoxon $P=0.8272$. The BCMA-frequency versus response analysis gave $P=0.7375$. In the same cohort, multiple effector-side biomarkers showed significant associations with both response and progression-free survival, spanning both effector availability — peripheral CD3 T-cell frequency ($P=0.018$) and counts ($P=0.012$), CD8 T-cell counts ($P=0.024$) — and effector functional state — regulatory T-cell frequency ($P=0.029$), CD38+ regulatory T-cell frequency ($P=0.0172$), and multiple T-cell exhaustion-marker measures across peripheral and bone-marrow compartments.

van der Horst et al. 2021 [4] examined both axes across the same 39 B-NHL primary lymph node samples in two effector configurations. In the allogeneic-effector configuration with healthy-donor PBMCs added, the CD20-versus-cytotoxicity correlation gave Spearman $r=0.23$, $P=0.24$. In the autologous configuration with only lymph-node-residing T cells, the local CD3+ T-cell frequency versus cytotoxicity correlation gave Spearman $r=0.63$, $P<0.0001$.

Three peer-reviewed primary sources, examining three distinct drug programs across BCMA in multiple myeloma, CD20 in CLL, and CD20 in B-NHL, each demonstrate within-paper that the binding-axis univariate analysis produces a null while the effector-side univariate analysis produces a positive result, in the same study population. None of the three papers integrates the two variables into a joint analytical framework. Each reports the two univariate analyses in parallel and notes the contrast verbally. The joint structure is implicit in the data and the verbal contrast; the analytical integration that would name and characterize the joint structure is absent from the published treatment.

3.2. Within-Program Triangulation in a Clinical Failure: CD33/AMG 330

The within-program demonstrations in Section 3.1 cover three approved or registration-stage drug programs, all of which are clinical successes. To examine whether the same structural pattern applies in a clinical failure direction, we present a within-program triangulation of CD33-targeted bispecific TCE development across three peer-reviewed primary sources. The CD33/AMG 330 trajectory is consistent with, but does not uniquely prove, the joint-axis account: the program was advanced on cell-line preclinical data showing potent activity at high effector-to-target ratios, encountered effector-related dose-limitation in primary AML samples, and ultimately failed to produce a broadly durable approved clinical program in AML, with the available primary-sample and clinical evidence pointing to effector-side constraints as one important limiting factor. Other contributors (CD33 internalization kinetics, on-target on-myeloid toxicity that constrained dose escalation) are not excluded by the structural reading offered here.

3.2.1. The Cell-Line Preclinical Readout

Friedrich et al. 2014 [8] characterized AMG 330 against eleven AML cell lines spanning CD33 receptor densities of 14,400 to 56,700 sites per cell, reporting EC₅₀ values in the low picomolar to single-digit picomolar range. The original publication described a qualitative trend of higher activity at higher CD33 expression and reported the program as preclinically validated for clinical advancement.

Reanalysis of the published Friedrich 2014 cell-line data using formal statistical methods does not support a within-target correlation by either parametric or non-parametric tests. On log-transformed CD33 receptor density and log EC₅₀, Pearson $r=-0.483$ with parametric $P=0.132$ and two-sided permutation $P=0.136$ (10,000 iterations); Spearman $\rho =-0.282$ with parametric $P=0.401$ and two-sided permutation $P=0.399$ (10,000 iterations). A 10,000-iteration percentile bootstrap on the Spearman coefficient produces a 95% confidence interval of $[-0.80,+0.45]$, with 24% of bootstrap replicates producing a non-negative coefficient. All resampling procedures used a fixed random seed for reproducibility; the analysis code and the underlying data are provided in the supplementary material. The published panel does not statistically distinguish a correlation from no relationship.

Beyond the statistical question, the panel does not span the clinical CD33 expression distribution in AML. Jilani et al. 2002 [11] reported clinical CD33 site densities on AML blasts with median approximately 10,380 sites per cell and a substantial low-tail extending below 1,000 sites per cell. The Friedrich 2014 panel covers only the upper portion of the clinical distribution. The cell-line readout therefore does not address the binding-axis behavior of the drug across the range of expression actually encountered in patients.

3.2.2. The Primary-Sample Readout

Krupka et al. 2015 [9] examined AMG 330 in primary AML bone-marrow samples at endogenous effector-to-target ratios — that is, with the T-cell content as it actually exists in patient marrow rather than at the high E:T ratios used in cell-line co-culture assays. Bone-marrow E:T ratios in relapsed-refractory AML are typically in the range of 1:80 to 1:1, two to three orders of magnitude lower than the 1:1 to 10:1 ratios standard in cell-line cytotoxicity panels. Krupka et al. reported that AMG 330 cytotoxic activity broke down at E:T ratios below approximately 1:80 in primary AML samples, with extensive donor-to-donor variability.

The threshold breakdown documented at clinically relevant E:T conditions is on the effector-availability axis, not the binding-axis. The drug’s binding to CD33 functions; the rate-limiting step at low E:T is whether sufficient functional T-cell engagement can be achieved to produce cytotoxicity. The same drug that showed potent low-picomolar EC₅₀ on cell-line panels at high E:T showed E:T-dependent activity breakdown on primary samples at clinically realistic E:T.

3.2.3. The Clinical Readout

Ravandi et al. 2024 [10] reported the Phase 1a clinical experience with AMG 330 in relapsed-refractory AML. Among the response correlates examined in this small dose-escalation cohort, baseline effector-to-target ratio and lower PD-1 expression on T cells at baseline showed a trend toward better clinical response, while no correlation was reported between CD33 expression on AML blasts at baseline and patient responses to AMG 330. The clinical readout is consistent in direction with the primary-sample finding: at clinical exposure levels and clinical effector conditions, the binding-axis variable did not separate responders from non-responders, while effector-side variables (availability and functional state) trended in the predicted direction. The cohort is small and the trend is qualitative; the structural reading does not require statistical confirmation at this sample size.

3.2.4. The Within-Program Structural Pattern

Reading the three sources together, the CD33/AMG 330 program traces a within-program trajectory that is structurally consistent with the joint binding-effector account. The cell-line preclinical readout (high E:T) showed potent activity but does not statistically support a within-panel density correlation, and does not span the clinical density distribution. The primary-sample readout at clinically realistic E:T documented breakdown of activity on the effector axis. The clinical readout identified the effector axis as a response correlate and did not identify the binding axis. Across all three readouts, the binding axis was either uninformative or showed no clear relationship with outcome; across the two readouts conducted under clinically realistic effector conditions, the effector axis was the predictive variable.

The CD33 program was developed and advanced based on the cell-line preclinical readout, which read primarily the binding axis under high-E:T conditions. The clinical failure is consistent with the program having been advanced on a readout that did not characterize the variable that ultimately was rate-limiting in the clinical setting. The structural pattern documented across the three readouts is the same structural pattern documented in Section 3.1 for the BCMA and CD20 programs: the binding axis univariate analysis is null or uninformative, the effector axis univariate analysis is predictive, neither program’s published treatment integrates the two axes into a joint analytical framework.

3.3. The Cross-Program Inventory

The cross-program inventory of clinical-stage bispecific TCE programs spanning eleven target antigens is presented in Table 2. The inventory comprises sixteen Tier-1 records meeting the inclusion criteria stated in Section 2.1, plus two supporting records that corroborate the structural pattern but do not meet the full Tier-1 definition. Records are grouped by structural function within the cross-program analysis: hematologic clinical successes (Group A); CD19 prior-art records that established the modality (Group B); hematologic programs that failed in clinical development or remain in extended early development (Group C); the single solid-tumor approval to date (Group D); discontinued solid-tumor programs (Group E); and supporting records (Group F). The grouping reflects clinical outcome and disease-setting structure rather than chronological or alphabetical ordering, and is the latent structure that the cross-program reading below relies on. The same table is also reproduced in the supplement (Supplementary Table S1) for cross-reference with the per-record verification notes provided there.

The cross-program record is consistent with the same structural reading developed in Section 3.1 and Section 3.2. We summarize the qualitative pattern below; the per-record primary-source documentation is provided in the supplement.

Bispecific TCE programs that have produced approved drugs cluster in disease settings where the disease microenvironment supports productive effector engagement at clinical exposure: B-cell malignancies in peripheral blood and lymph nodes (CD19/blinatumomab; CD20/glofitamab, epcoritamab, mosunetuzumab), multiple myeloma in bone marrow (BCMA/teclistamab, elranatamab; GPRC5D/talquetamab), and small-cell lung cancer (DLL3/tarlatamab). In these settings, effector-side constraints are not absolute barriers at approved dosing, although response within each setting still varies with baseline immune fitness and disease context. The binding-axis variables across these approved drugs span a wide range of target antigen expression — from the relatively low expression of DLL3 on SCLC tumor cells to the relatively high expression of BCMA on multiple myeloma plasma cells — without the cross-program clinical outcome tracking the cross-program target expression.

Bispecific TCE programs that have failed in clinical development cluster, conversely, in disease settings where effector engagement is constrained, absent, or inconsistent at clinical exposure. Solid-tumor TCE programs targeting EpCAM (solitomab [12]), CEA [13], P-cadherin [14], gpA33 [15], and PSMA (pasotuxizumab [16]) have been discontinued. The relapsed-refractory AML setting has proven particularly difficult: CD33/AMG 330 encountered effector-side dose-limitation in primary samples and clinic, while CD123/flotetuzumab is better read as supporting the effector-context framework than as a simple failure case, with response appearing enriched in immune-infiltrated AML — suggesting that the relevant variable is not CD123 expression alone but whether the disease state supports productive effector engagement. Across these unsuccessful program clusters, the disease microenvironment does not reliably provide effector content adequate to produce cytotoxicity at the E:T ratios encountered in patients, regardless of how effectively the molecule’s binding arm engages its target.

Across the clinical-development record, successful and unsuccessful programs are more coherently organized by the joint binding-effector context than by target-antigen density alone. Approval status tracks where the disease microenvironment supports effector engagement at clinical exposure; failure tracks where effector engagement is constrained or inconsistent at clinical exposure. The binding-axis variables (target antigen expression, binding affinity, molecular format) do not separate the approved programs from the failed programs in a way that organizes the clinical record.

This is the cross-program counterpart to the within-program demonstrations in Section 3.1 and Section 3.2. At both scales, the record is more coherently organized by the joint target–effector context than by either axis alone. At neither scale is the joint analytical framework that would name this structure currently present in the field’s published treatment of the empirical record.

4. Discussion

4.1. What the Empirical Record Shows

The systematic empirical pattern documented in Section 3.1 through Section 3.3 has three interlocking features. First, within-target reports of target antigen density on tumor cells fail to predict bispecific TCE potency or clinical response across multiple drug programs, target classes, disease settings, molecular formats, and readout methodologies; this finding has been observed in five peer-reviewed primary sources spanning 2017 through 2024 including a clinical correlative analysis from the registration trial of an approved drug. Second, the same field’s published cell-line literature contains within-target correlation reports of widely varying strengths, ranging from coefficients approaching $r=-0.91$ to no statistical support for any correlation, with the heterogeneity unorganized by any unifying analytical framework. Third, in the studies that examined both target antigen density and a measure of effector-side biology in the same study population, the binding-axis univariate analysis produced a null result while the effector-side univariate analysis produced a positive result; this paired pattern was observed in three approved or registration-stage programs (teclistamab, epcoritamab in two distinct disease settings) and in within-program triangulation of a clinical failure (AMG 330).

These three features together describe a field whose empirical record contains a structural pattern that the field’s analytical conventions have not organized. The contradictions are real in the sense that the published reports are genuinely heterogeneous in their univariate findings; the contradictions are not real in the sense that the heterogeneity reflects a joint structure being read by analytical conventions that examine the variables in isolation. When studies measured both variables in the same population, the contradictions disappeared into a stable two-axis pattern.

4.2. The Proposed Explanation

The joint binding-effector account provides a coherent explanation of the systematic empirical record. It proposes that target antigen density and effector-side biology operate as joint variables in determining bispecific TCE cytotoxic potency and clinical outcome, with the rate-limiting variable transitioning between them depending on assay conditions and disease setting. Under typical conditions of clinical exposure for approved or registration-stage TCEs in their approved indications, the binding layer is engaged adequately by the molecule’s binding arm, shifting the rate-limiting variable to effector-side biology; this accounts for why the binding-axis univariate analysis produces null results in clinical correlative studies and primary-sample studies under clinically realistic effector conditions. Under conditions of cell-line cytotoxicity assays at high E:T ratios, effector availability is held abundantly above the rate-limiting threshold and effector functional state is uniform (donor-derived healthy T cells), shifting the rate-limiting variable to binding-axis engagement; this accounts for why some cell-line panels show within-target density correlations and why correlation strengths vary across panels depending on where each panel sits in the joint variable space. In the within-program failure case (CD33/AMG 330), the program was advanced on cell-line readouts (where binding-axis was rate-limiting) into a clinical setting (relapsed-refractory AML) where effector-side biology was the rate-limiting variable; the program’s failure to produce a broadly durable approved clinical program is consistent with the readout used to advance the program not characterizing the variable that ultimately was rate-limiting.

The joint binding-effector account has three important properties. First, it is logically coherent: the joint-axis structure is well-defined, the rate-limiting transition is well-defined, and the predictions about which axis dominates under which conditions are derivable qualitatively from the structure. Second, it is internally consistent: it does not generate predictions that contradict each other across the conditions it covers. Third, it is consistent with the systematic empirical record assembled here: across all three features of the empirical record (Section 4.1), it provides a coherent account of the published findings, including the variation in within-target cell-line correlation strengths across panels, which is the expected consequence of different panels sitting in different regions of the joint variable space.

We do not claim that this account uniquely explains the empirical record. Alternative explanations remain possible, including target-specific biology, format-specific binding properties, methodological factors specific to assay conditions, and disease-specific immunological features not captured in the published record. The account is offered as a testable organizing explanation of the systematic empirical record we have assembled, on the criteria appropriate to a field-level structural argument: coherence, internal consistency, and consistency with the documented empirical pattern. Whether the account subsequently survives finer-grained tests against alternative explanations is a question that further work will address.

4.3. The Status of the Explanation as a Contribution

The contribution of the present paper is analytical and empirical. It documents a systematic primary-source record, identifies a recurrent structural pattern within that record, and proposes a mechanistically motivated explanation that can be refined or tested in future work. This is a useful first step for a field-level structural claim because it defines the pattern that finer-grained mechanistic, quantitative, and prospective studies can subsequently test. Without an analytically grounded structural account, finer-grained empirical investigation has no organizing target.

The explanation is mechanistically motivated by a model of membrane-confined ternary binding under non-equilibrium effector kinetics, the qualitative form of which provides the rate-limiting transition described above. The structural argument in this paper does not depend on the quantitative apparatus of the underlying model; the qualitative two-axis structure is consistent with elementary considerations of saturation kinetics in ternary engagement systems and is sufficient to organize the empirical record assembled here. The quantitative apparatus, when applied, generates specific predictions about transition thresholds and dose-response shapes that this paper does not address.

4.4. Field-State Observations

We are not aware of a unifying explanation along these lines having been proposed in the field’s published literature, and in any case none appears to have been adopted, as the field’s published analytical practice continues to report parallel univariate analyses of target antigen density and effector-side biology without an integrated framework. This observation about adoption is supported by direct examination of the most recent and rigorous correlative studies in the field. Cortes-Selva et al. 2024 [2], the registration-trial correlative biomarker analysis for teclistamab published in Blood with $n=165$, reports BCMA-density and effector-side biomarkers in parallel univariate analyses without integration into a joint analytical framework. Mhibik et al. 2023 [3], the most rigorous primary-sample CD20 study in the verified record, reports the two axes in parallel and notes the contrast verbally without integrating them analytically. These are not isolated examples; they represent the current state of the field’s analytical practice in its most visible recent work on this question.

The pattern of continued parallel-univariate reporting is consistent with a field that has not adopted any unifying analytical framework for these variables, regardless of whether such a framework may have been proposed in some venue we have not encountered. The empirical observation is about field practice, not about the comprehensive existence of all possible prior proposals.

4.5. Implications for Preclinical Practice

If the joint-axis structure is correct, several implications follow for how bispecific TCE preclinical investigation should be conducted to produce findings predictive of clinical outcome. Cell-line cytotoxicity assays at high effector-to-target ratios read primarily the binding axis under conditions in which effector availability is held above its rate-limiting threshold and effector functional state is uniform; these assays do not characterize the variables that become rate-limiting in clinical settings where effector availability is constrained or effector functional state is compromised. Primary-sample cytotoxicity assays at endogenous effector content provide a more clinically relevant readout. Clinical correlative biomarker analyses that include effector-side measures (availability and functional state) alongside binding-axis measures provide the strongest test of which variable is rate-limiting in a given clinical context.

Programs intended for clinical settings where effector-side biology is likely constraining (relapsed-refractory hematologic malignancies after multiple lines of therapy, where availability is depleted and exhaustion phenotypes accumulate; solid tumors with immunosuppressive microenvironments, where functional state is compromised; settings where prior chemotherapy has depleted T-cell content) cannot be reliably advanced on cell-line readouts that hold both availability and functional-state variables outside the regime they will encounter in the clinic. The CD33/AMG 330 case is one documented example of a program advanced on such readouts that subsequently encountered effector-side constraints in the clinical setting.

Conversely, programs in disease settings where effector engagement is more reliably supported at clinical exposure (B-cell malignancies in peripheral blood and lymph nodes; multiple myeloma in bone marrow at approved dosing; selected disease settings where T-cell content is preserved) can advance on readouts characterizing primarily the binding axis, because the binding axis is more often rate-limiting in those settings. We note that even within these settings, response varies with baseline immune fitness and disease context — myeloma marrow is not uniformly effector-adequate, and approved myeloma TCEs show that effector-side constraints are not absolute barriers in this setting at approved dosing without implying that effector-side biology is irrelevant to within-setting response. The current field practice of treating cell-line preclinical readouts as standardized program-evaluation tools across all disease settings is not consistent with the joint-axis structure; the readout that is informative depends on which variable is rate-limiting in the intended clinical setting.

4.6. Limitations

The systematic empirical record assembled here has limitations that bear on the conclusions it can support. The verified within-target documentation covers five peer-reviewed primary sources; additional primary sources may exist that we did not encounter, and the breadth of the pattern may be larger or more heterogeneous than the verified record indicates. The cross-program inventory covers sixteen Tier-1 records spanning eleven target antigens; additional clinical-stage programs exist whose primary-source documentation we did not verify. The CD33/AMG 330 case is one within-program failure case; additional within-program failure cases would strengthen or modify the structural reading depending on what they show.

The joint binding-effector account is a testable explanation; alternative explanations have not been ruled out. Target-specific biology, format-specific properties, and factors not captured in the published record may contribute to the empirical pattern. The within-program demonstrations in Section 3.1 establish that when both axes are measured in the same population the binding-axis univariate result is null and the effector-axis univariate result is positive; they do not establish that the joint-axis structure is the unique mechanistic explanation for this finding.

The present paper evaluates the proposed structure at the level of the published empirical record. More detailed quantitative formulations may generate additional predictions, but those are outside the scope of the present manuscript.

5. Conclusion

The bispecific T-cell engager field has produced a substantial empirical record on the relationship between target antigen density and clinical outcome that contains apparent contradictions when read in parallel-univariate framings. We have assembled a verified systematic dataset spanning the field’s clinical-stage programs and read it against the published primary-source record. The empirical pattern that emerges — within-target density null findings reported across multiple targets, diseases, formats, and methodologies; heterogeneous within-target cell-line correlation strengths; within-program paired analyses showing null binding-axis results and positive effector-side results in the same study populations; cross-program clinical record more coherently organized by joint binding-effector context than by target expression alone — is consistent with a joint binding-effector account in which target antigen density and effector-side biology operate as joint variables with rate-limiting transitions between them.

We offer this joint binding-effector account as a testable explanation of the systematic empirical record. It is not the only possible account, but it provides a logically coherent and internally consistent explanation of the documented findings. The systematic dataset and the verified primary-source documentation are deposited as supplementary material to support independent engagement with the empirical record.