T-Detect: A Polysemous Detection Tool

• T-Detect is a multifaceted term that denotes different, context-specific methods, such as T-cell diagnostics and adversarial text detection.
• In immunology, T-Detect uses cohort-trained machine learning to classify disease exposure, contrasting with methods like ALICE that detect active immune responses.
• In NLP, T-Detect applies a heavy-tailed Student’s t-distribution normalization to improve zero-shot adversarial text detection under challenging perturbations.

T-Detect is a polysemous designation rather than a single universally defined method. In contemporary research, it appears in several unrelated technical contexts, most notably as a label for T-cell receptor repertoire diagnostics in immunology and as the name of a tail-aware zero-shot detector for adversarial machine-generated text in natural language processing. The same designation also appears in computer vision, astronomy, and pharmacovigilance, where it denotes distinct operational ideas and statistical procedures. Precise domain qualification is therefore necessary whenever the term is used (Pogorelyy et al., 2018, Pogorelyy et al., 2018, West et al., 31 Jul 2025).

1. Polysemy and domain-specific meanings

The term has been used for multiple, non-interoperable constructs. This suggests that “T-Detect” functions more as a local naming convention than as a stable cross-domain technical standard.

Usage	Domain	Defining description
T-Detect-style supervised diagnostic classifiers	TCR immunodiagnostics	Large labeled cohorts and machine learning map TCR sequences to prior disease exposure or infection status (Pogorelyy et al., 2018)
T-Detect	Adversarial text detection	Replaces Gaussian normalization with a Student’s t-distribution normalization (West et al., 31 Jul 2025)
Track-to-Detect	Video object detection	Learned tracklets improve video-level detection (Feichtenhofer et al., 2017)
T-Detect	T-dwarf companion surveys	Spitzer/IRAC imaging plus multi-epoch astrometry detects or rules out wide, cool companions (Carson et al., 2011)
T-Detect	Pharmacovigilance	Student’s t-test–driven feature selection over pre/post drug-exposure feature matrices (Liu et al., 2013)

In immunology, the designation is tied to cohort-trained TCR diagnostics and is discussed primarily through contrast with unsupervised repertoire methods such as ALICE and with longitudinal clone-tracking frameworks. In natural language processing, by contrast, T-Detect names a concrete statistical normalization procedure embedded in curvature-based detection of machine-generated text. These two usages share neither input representation nor statistical objective, and their apparent lexical overlap is accidental rather than methodological (Pogorelyy et al., 2018, Pogorelyy et al., 2018, West et al., 31 Jul 2025).

2. T-Detect in T-cell receptor repertoire diagnostics

In the immunodiagnostic sense, T-Detect and related products use large labeled cohorts and machine learning to map TCR sequences, usually exact or motif-level matches, to prior disease exposure or infection status, producing disease-specific classifiers with defined sensitivity and specificity. Within this formulation, the principal output is a disease-specific call from a single repertoire snapshot. The crucial limitation, stated explicitly in the comparative literature, is that such classifiers do not intrinsically indicate whether a response is ongoing at the time of sampling (Pogorelyy et al., 2018).

This cohort-trained paradigm is closely related to efforts to discover antigen-responding clonotypes from longitudinal data. In yellow fever vaccination data from three pairs of identical twins, a Bayesian expansion/contraction framework identified 500–1500 responding TCRs in each donor and validated them using three independent assays. The same study defined a sequence-similarity classifier by the minimal CDR3 amino-acid Hamming distance to expanded responders, showing that convergent sequence structure could generalize across donors. A plausible implication is that T-Detect-style systems can be strengthened when public cohort signatures are supplemented by longitudinally discovered private and convergent responders (Pogorelyy et al., 2018).

The conceptual distinction is therefore between trait-like and state-like inference. T-Detect-style classifiers estimate exposure or disease association from repertoire content, whereas longitudinal discovery frameworks estimate recent biological activity from temporal clone dynamics. The literature treats these as complementary rather than competing regimes (Pogorelyy et al., 2018).

3. Relationship to ALICE and single-snapshot immune-response detection

The sharpest immunological contrast to T-Detect-style cohort classifiers is ALICE, “Antigen-specific Lymphocyte Identification by Clustering of Expanded sequences,” which identifies TCR clonotypes actively involved in an ongoing immune response from a single RepSeq sample. ALICE operates per VJ combination on productive, in-frame clonotypes, removes singleton nucleotide clonotypes, defines neighbors as CDR3 amino acid sequences within Hamming distance $\leq 1$, and tests whether the observed number of neighboring clonotypes exceeds the expectation under a generative V(D)J recombination baseline corrected by a thymic selection factor $Q = 9.41$. Its null model is

$$P(d \mid \sigma)=e^{-\lambda}\frac{\lambda^d}{d!}, \qquad \lambda=n\sum_{\sigma'\sim \sigma}Q P_{\rm gen}(\sigma'),$$

with significance determined by Benjamini–Hochberg correction and BH-adjusted $p<0.001$ for reported ALICE hits. An abundance-aware variant replaces the neighbor count by $s=\sum_{i=1}^{d} f(c_i)$ with $f(c)=\log(c)$ (Pogorelyy et al., 2018).

ALICE was validated across several biological settings. In naive versus effector-memory TCR$\beta$ repertoires, it identified multiple hits in memory but virtually none in naive subsets; normalized ALICE hits differed with Wilcoxon rank-sum two-tailed $p=4.9\times10^{-16}$, and ROC-based classification achieved AUROC $0.92$ for ALICE hits and $0.96$ for the combined statistic. In mixed lymphocyte reaction cultures, ALICE found significantly more hits than in unstimulated controls with $Q = 9.41$0. In yellow fever vaccination, 40–70% of day-15 ALICE hits were highly similar to independently identified YF-17D–reactive clonotypes. In anti-CTLA4 cohorts, ALICE hits increased post-treatment, with AUROC $Q = 9.41$1 versus $Q = 9.41$2 for richness in one cohort. In ankylosing spondylitis synovial fluid, it recovered clusters containing public clonotypes shared among all three HLA-B27-positive patients (Pogorelyy et al., 2018).

This comparison matters because the literature assigns different diagnostic roles to the two approaches. ALICE is unsupervised, requires no reference panel, and flags expansions in real time but does not assign disease labels. T-Detect-style supervised classifiers assign disease labels but do not intrinsically resolve whether the response is ongoing. The proposed integration is explicit: ALICE can enrich for currently expanding clonotypes, prioritize candidate sequences for supervised training sets, and act as a “response activity flag” alongside disease-specific calls from supervised models (Pogorelyy et al., 2018).

4. T-Detect as tail-aware normalization for adversarial machine-generated text

A separate and explicitly named use of T-Detect appears in zero-shot detection of adversarial machine-generated text. In this setting, T-Detect redesigns the statistical core of curvature-based detectors by replacing Gaussian normalization with a heavy-tailed discrepancy score derived from the Student’s t-distribution. The motivation is empirical: on the RAID benchmark for adversarial text, detector scores show positive excess kurtosis of $Q = 9.41$3, and model selection by AIC favors the t-distribution over the Gaussian by $Q = 9.41$4; on HART, the score distribution is closer to Gaussian, with kurtosis $Q = 9.41$5 (West et al., 31 Jul 2025).

The method keeps the discrepancy computation of Fast-DetectGPT but changes the normalization. With

$Q = 9.41$6

T-Detect computes

$Q = 9.41$7

where $Q = 9.41$8 and the default is $Q = 9.41$9. The method is zero-shot, uses the same efficient reference-moment estimation as the baseline, and adds no extra model calls; the only alteration is the denominator $$P(d \mid \sigma)=e^{-\lambda}\frac{\lambda^d}{d!}, \qquad \lambda=n\sum_{\sigma'\sim \sigma}Q P_{\rm gen}(\sigma'),$$0 (West et al., 31 Jul 2025).

The reported gains are strongest under adversarial conditions. In the two-dimensional CT framework, CT(T-Detect) achieved AUROC $$P(d \mid \sigma)=e^{-\lambda}\frac{\lambda^d}{d!}, \qquad \lambda=n\sum_{\sigma'\sim \sigma}Q P_{\rm gen}(\sigma'),$$1, F1 $$P(d \mid \sigma)=e^{-\lambda}\frac{\lambda^d}{d!}, \qquad \lambda=n\sum_{\sigma'\sim \sigma}Q P_{\rm gen}(\sigma'),$$2, and TPR@5%FPR $$P(d \mid \sigma)=e^{-\lambda}\frac{\lambda^d}{d!}, \qquad \lambda=n\sum_{\sigma'\sim \sigma}Q P_{\rm gen}(\sigma'),$$3 on RAID ALL, with AUROC $$P(d \mid \sigma)=e^{-\lambda}\frac{\lambda^d}{d!}, \qquad \lambda=n\sum_{\sigma'\sim \sigma}Q P_{\rm gen}(\sigma'),$$4 on the Books domain and $$P(d \mid \sigma)=e^{-\lambda}\frac{\lambda^d}{d!}, \qquad \lambda=n\sum_{\sigma'\sim \sigma}Q P_{\rm gen}(\sigma'),$$5 on Poetry. In single-dimension evaluation on RAID ALL, T-Detect reached AUROC $$P(d \mid \sigma)=e^{-\lambda}\frac{\lambda^d}{d!}, \qquad \lambda=n\sum_{\sigma'\sim \sigma}Q P_{\rm gen}(\sigma'),$$6 on the T dimension and $$P(d \mid \sigma)=e^{-\lambda}\frac{\lambda^d}{d!}, \qquad \lambda=n\sum_{\sigma'\sim \sigma}Q P_{\rm gen}(\sigma'),$$7 on the C dimension. On HART Level 3 ALL, it obtained AUROC $$P(d \mid \sigma)=e^{-\lambda}\frac{\lambda^d}{d!}, \qquad \lambda=n\sum_{\sigma'\sim \sigma}Q P_{\rm gen}(\sigma'),$$8, F1 $$P(d \mid \sigma)=e^{-\lambda}\frac{\lambda^d}{d!}, \qquad \lambda=n\sum_{\sigma'\sim \sigma}Q P_{\rm gen}(\sigma'),$$9, and TPR@5%FPR $p<0.001$0. The ablation isolated the source of improvement: baseline Gaussian normalization yielded AUROC $p<0.001$1, whereas t-normalization alone yielded $p<0.001$2. Throughput was reported as $p<0.001$3 texts/s, compared with $p<0.001$4 for Fast-DetectGPT, and timing stability improved from standard deviation $p<0.001$5 to $p<0.001$6 (West et al., 31 Jul 2025).

The method’s limitations are also explicit. Zero-width spaces had a failure rate of $p<0.001$7, homoglyph attacks $p<0.001$8, paraphrase $p<0.001$9, and synonym substitution $s=\sum_{i=1}^{d} f(c_i)$0. The paper therefore identifies Unicode and character-level normalization as necessary deployment components. It also states that heavy-tail normalization reduces outlier sensitivity for non-native English text but does not fully solve fairness problems (West et al., 31 Jul 2025).

5. Additional uses of the designation

In video understanding, “Track-to-Detect” denotes the use of learned tracking signals to improve object detection at the video level. In “Detect to Track and Track to Detect,” a ConvNet jointly performs frame-based detection and across-frame track regression, then links detections with a Viterbi-style association objective and rescoring. On ImageNet VID, the D+T model achieved $s=\sum_{i=1}^{d} f(c_i)$1 mAP with ResNet-101 at temporal stride $s=\sum_{i=1}^{d} f(c_i)$2, and $s=\sum_{i=1}^{d} f(c_i)$3 at $s=\sum_{i=1}^{d} f(c_i)$4; adding tracking loss and correlation features improved single-frame detection from $s=\sum_{i=1}^{d} f(c_i)$5 to $s=\sum_{i=1}^{d} f(c_i)$6 mAP (Feichtenhofer et al., 2017).

In astronomy, “T-Detect” refers to the use of Spitzer/IRAC imaging, especially at $s=\sum_{i=1}^{d} f(c_i)$7m, together with multi-epoch astrometry to search for wide, cool T-dwarf companions around nearby low-mass primaries. The survey covered 117 nearby M, L, and T dwarf systems, found no new common-proper-motion substellar companions, and translated the null result into population constraints through Monte Carlo orbital simulations. The resulting upper limits were $s=\sum_{i=1}^{d} f(c_i)$8 for $s=\sum_{i=1}^{d} f(c_i)$9–$f(c)=\log(c)$0 K companions over $f(c)=\log(c)$1–$f(c)=\log(c)$2 AU and $f(c)=\log(c)$3 for $f(c)=\log(c)$4–$f(c)=\log(c)$5 K companions over $f(c)=\log(c)$6–$f(c)=\log(c)$7 AU at $f(c)=\log(c)$8 confidence (Carson et al., 2011).

In pharmacovigilance, T-Detect denotes a Student’s t-test–driven feature selection framework over pre/post drug-exposure feature matrices. Applied to Pioglitazone in THIN, the method used 9093 exposed patients, 60-day pre-exposure and post-exposure windows, Read Code features at levels 1–5 and 1–3, and groups of approximately 100 patients to form paired matrices $f(c)=\log(c)$9 and $\beta$0. Feature-wise paired t-tests with $\beta$1 were used, and features with $\beta$2 were treated as candidate ADR signals, without multiple testing correction (Liu et al., 2013).

These additional examples show that the same label may denote a tracking-assisted detection strategy, an observational companion-search protocol, or a t-test screening framework. No common algorithmic core connects these usages (Feichtenhofer et al., 2017, Carson et al., 2011, Liu et al., 2013).

6. Conceptual boundaries, misconceptions, and comparative interpretation

A recurrent misconception is that T-Detect names a single transferable algorithm. The literature supports the opposite conclusion. In immunology, T-Detect-style systems are supervised repertoire classifiers for prior disease exposure, whereas ALICE is an unsupervised single-snapshot detector of ongoing clonal expansion. The distinction is operationally important: a cohort-trained classifier can yield a disease-specific label without revealing whether the response is active, and ALICE can reveal active expansion without identifying the disease or antigen (Pogorelyy et al., 2018).

A second misconception is that the heavy-tailed text-detection formulation provides a universal defense against adversarially perturbed language. The reported results do not support that interpretation. T-Detect improves robustness on RAID and integrates effectively with CT fusion, but zero-width spaces and homoglyphs remain major failure modes, and the paper explicitly recommends Unicode or character-level normalization before detection. Similarly, performance on Gaussian-like regimes is described as largely convergent with standard normalization as $\beta$3, not as categorically superior in every setting (West et al., 31 Jul 2025).

The broader interpretive lesson is that “T-Detect” should be treated as a context-bound technical term. In one setting it denotes exposure classification from TCR repertoires; in another it denotes Student’s t-based normalization of log-likelihood discrepancies; elsewhere it denotes tracking-enhanced video detection, T-dwarf companion searches, or t-test feature screening in EHRs. Accurate use therefore depends less on the label itself than on the associated data modality, statistical model, and inferential target.