Pith. sign in

REVIEW 3 major objections 6 minor 43 references

Temporal, event-aware modeling of T-cell receptor repertoires improves cancer-status prediction beyond static sequence-bag encoders, with internally matched AUCs of 0.982 (lung) and 0.984 (thyroid).

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 18:34 UTC pith:CKLH5ECX

load-bearing objection The temporal-attribution claim is confounded by the hybrid transport loss; the paper's packaging is good but it needs an ablation that isolates transport from ODE dynamics. the 3 major comments →

arxiv 2607.17244 v1 pith:CKLH5ECX submitted 2026-07-19 cs.LG

DynImmune-BERT: Dynamic Immune Repertoire Modeling with Neural ODE Driven Continuous Transformers

Rong Fu , Yongtai Liu , Xiaowen Ma , Haoyu Zhao , Shuo Yin , Yiqing Lyu , Long Zhang , Wangyu Wu This is my paper
classification cs.LG
keywords T cell receptor repertoirelongitudinal modelingNeural ODEcontinuous-time transformercancer detectionimmune status predictionoptimal transportevent-aware time series
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to show that longitudinal T-cell receptor (TCR) repertoires carry cancer-relevant signal that static repertoire models throw away: when a clone expands, disappears, or reappears across irregularly spaced blood draws, that event pattern is informative. It introduces a continuous-time model that treats each patient as a trajectory, using a presence-gated neural ODE to evolve latent clone states between visits, restarting states when clones reappear, and supervising both dominant and rare clone mass through a hybrid transport objective. Under internally matched patient-level splits, the full model reaches AUC 0.982 on lung cancer and 0.984 on thyroid cancer, outperforming static pooling and simpler time-aware encoders, with the ODE core driving most of the gain. If these numbers hold, event-aware temporal modeling is a practical complement to strong static immune-repertoire encoders for patient-level prediction.

Core claim

The central discovery is that the presence pattern and timing of clones—not just their sequences—are learnable, useful features for immune status. DynImmune-BERT encodes each observed repertoire as depth-stabilized compositional states, propagates these states continuously between irregular sampling times through a presence-gated vector field, and treats clone reappearance as an event that resets the latent state via a low-rank meta-adapter. A hybrid transport loss aligns predicted clone mass with observed abundance for both high-frequency and rare clones. In internally controlled comparisons, the full temporal model exceeds a latent-ODE-only control (0.982 vs 0.963 lung; 0.984 vs 0.969 thyr

What carries the argument

The load-bearing mechanism is a presence-gated Neural ODE: between observations, clone states evolve by dy_c/dt = I_c(t) F_theta(y_c, e_c, N_c(t), s(t), t), where the indicator I_c(t) stops absent clones from drifting, and reappearance triggers a state restart that is recentered onto the compositional support. This is paired with depth-adaptive centered-log-ratio initialization, bounded neighborhood self-attention for local context, and a hybrid transport objective that supervises dominant clone mass with entropic transport and rare clone mass with sliced Wasserstein distance.

Load-bearing premise

The central claim stands only if the internally matched AUC gains are real differences in predictive signal rather than artifacts of the specific patient splits, threshold selection, or seed choices.

What would settle it

Re-run the internal temporal comparison with observation timestamps randomly permuted within each patient, keeping clone sets and labels unchanged; if the permuted-time model keeps an AUC near 0.982/0.984 instead of dropping toward the static-pooling level (0.913/0.927), the time-ordering claim is falsified. A second check is an independent prospective cohort with at least a few hundred patients per disease and a preregistered threshold.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Longitudinal repertoires with irregular sampling intervals can be used directly for patient-level prediction rather than being collapsed into per-sample bags.
  • Event-aware handling of clone reappearance beats purely smooth latent interpolation, so modeling presence changes matters, not just elapsed time.
  • Supervising rare clones alongside dominant ones helps preserve information that abundance-based pooling would suppress.
  • The reported internally controlled AUCs imply static repertoire encoders leave measurable signal on the table when repeated samples are available.
  • The method adds modest parameter and runtime overhead relative to static encoders, making temporal modeling practical at current sequencing turnaround scales.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • In my reading, the neural-ODE trajectory is an interpolation device, not a claim about smooth biological clone dynamics; the authors say this, and the distinction should guide interpretation of trajectory plots.
  • A natural extension would be to apply the same event-aware continuous-time recipe to other compositional longitudinal data—B-cell receptor repertoires, microbiome abundances, or cell-free DNA allele fractions—where irregular sampling and presence events are common.
  • A direct test of whether time is the carrier of signal: permute observation timestamps within each patient and re-run the matched comparison; if AUC does not fall toward the static-pooling level, the temporal gain would need a different explanation.
  • The external disease-level AUCs should be read as existence proofs until cohorts reach hundreds of patients; the paper's own uncertainty limits flag this, but the reader can carry that caution forward.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. Longitudinal T cell receptor repertoires contain clonal dynamics that static repertoire models ignore. The paper proposes DynImmune-BERT, which represents each patient's repertoire as a continuous-time trajectory: clone reads are transformed into depth-adaptive centered-log-ratio states, propagated by a presence-gated Neural ODE with bounded-neighborhood attention, interrupted by event restarts for reappearing clones, and supervised by a hybrid transport loss on dominant and rare clone mass. On internally matched patient-level splits the full model reaches AUC 0.982 (lung) and 0.984 (THCA), outperforming the latent-ODE control (0.963/0.969). The authors interpret this as evidence that event-aware temporal modeling complements static encoders, with the ODE core as the largest contributor. External cohorts (n=8–24) are reported with uncertainty.

Significance. If the central claim is properly supported, the framework would be a meaningful step for longitudinal immune-repertoire analysis, where most models still pool samples independently. The paper is careful to separate literature-reported numbers from internally controlled comparisons, uses held-out patient splits, and reports seed variability, calibration, and threshold diagnostics. The main limitation is that the key ablation isolating the contribution of temporal dynamics from the auxiliary transport loss is missing, so the paper's headline interpretation is not yet established.

major comments (3)
  1. [§4.3, Table 3; Eq. (16)–(17)] The comparison between 'Full DynImmune-BERT' and 'Latent ODE control' does not isolate temporal dynamics. The full model differs from this control in at least two respects: (i) the event-aware ODE with restarts and (ii) the hybrid transport loss in Eq. (16), which is part of the training objective L in Eq. (17). Section 4.3 states that the controls 'do not include the full event restart and hybrid transport design,' confirming that the latent-ODE control was trained without the transport loss. The AUC gap (0.982 vs 0.963 lung; 0.984 vs 0.969 THCA) could therefore be explained by the richer training signal alone, independent of temporal dynamics. Please provide ablations with λW=0 in the full model and/or with the transport loss added to the latent-ODE and static controls.
  2. [§4.3, Fig. 4] The claim that 'the ODE core drives most of the accuracy gain' is not verifiable from the manuscript. Figure 4 is described as reporting seed variability, but no numerical ablation table or standard deviations are given in the text, and it is unclear which components (ODE, event restart, pseudocount, low-rank adapter) are removed or whether each ablation is retrained with the same hyperparameters. The associated sentence about event handling and pseudocounts improving robustness is likewise unsupported. A numerical ablation table with mean±std over seeds is needed.
  3. [Table 2 and §4.2] The THCA results are cited to reference [43], which is a lung-cancer cohort (Joshi et al., Nature Medicine 2019). If the THCA data originate from a different public dataset, the reference is incorrect; if they are from [43], the cohort description does not match. Since the THCA rows are primary results, the data provenance must be corrected or clarified.
minor comments (6)
  1. [§3.4, Eq. (12)] I_c(t) is defined only at observed time points in Eq. (8), but the ODE requires I_c at all integration times. Specify how the gate is set between observations (e.g., nearest observation, forward-filling, or a latent presence model).
  2. [Table 2] Missing spaces in 'LungDynImmune-BERT' and 'THCADynImmune-BERT'.
  3. [Table 4] The 'AUC uncertainty' column mixes a bootstrap CI (THCA) with standard errors (all other rows); use one consistent measure.
  4. [§3.6, Eq. (18)] The Lipschitz inequality is a restatement of the assumption; consider clarifying what operational guarantee follows (e.g., existence/uniqueness of the ODE solution).
  5. [§3.2/§3.3] 'Retained candidate clones' in S_t is not defined; state how the candidate support is built and updated.
  6. [§4.1] The listed NumPy version, 1.7.1, is likely a typo for a much newer release; please verify the software versions.

Circularity Check

0 steps flagged

No construction-level circularity: the temporal gain is learned end-to-end and evaluated on held-out splits; the main concern is an ablation confound (transport loss vs ODE core), not a definitional reduction.

full rationale

The derivation chain is not circular. The patient-level prediction (Eq. 2) is a learned function of the dynamic encoder; the vector field, restart adapter, and classifier are trained parameters rather than constants fitted to the target labels. The hybrid transport objective (Eqs. 15-17) supervises clone-mass reconstruction during training, but clone-mass reconstruction is not the reported evaluation metric and is not presented as a held-out prediction claim; AUC is computed on held-out patient splits satisfying Eq. 3. I found no load-bearing self-citation: the cited literature is external prior work on neural ODEs, attention, transport, and TCR methods, and no uniqueness theorem from the present authors is invoked to force the architectural choices. The closest concern is Section 4.3, where the full model differs from the 'Latent ODE control' by both event restart and 'the full event restart and hybrid transport design', so the sentence 'Ablations indicate that the ODE core drives most of the accuracy gain' lacks a fully isolated ablation; this is an experimental-control gap, not a case where one equation reduces to another by construction. The paper itself flags the limitation that 'External subsets with fewer than 20 cases have wide uncertainty' (Section 5), reinforcing that small-cohort numbers are preliminary rather than circular predictions. Therefore no circular step is present.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The model introduces no new physical or biological entities. All listed free parameters are standard hyperparameters with reported values, none of which are fit to test labels. The main unstated assumptions are about data handling: split integrity, candidate-clone support definition, and transferability of the architecture across heterogeneous cohorts.

free parameters (4)
  • neighborhood bound B = 96
    Bounded neighborhood size (Eq. 10) chosen before test reporting; affects scalability and the attention context.
  • depth-adaptive pseudocount κ = 0.5
    Controls regularization of low-depth observations in Eq. 6; fixed before final test reporting.
  • transport weights ω_top, ω_tail = 0.65, 0.35
    Mixture weights in Eq. 16 balancing entropic transport on top clones and sliced Wasserstein on tail clones.
  • regularization weights λW, λtemp, λspec = 0.2, 0.001, 0.0001
    Weights in the final loss Eq. 17; cited as fixed before final test reporting.
axioms (5)
  • domain assumption Patient-level split guarantee (Eq. 3) is actually enforced in all tables
    The paper asserts Ptrain ∩ Pval ∩ Ptest = ∅, but no patient IDs, split code, or leakage audit is provided; if violated, the temporal gains would be spurious.
  • domain assumption Neural-ODE latent interpolation is a valid representation of clone dynamics
    Invoked throughout Sec. 3.4: latent smoothness between observations plus discrete restarts at presence events is assumed adequate to encode biological expansion/contraction/reappearance.
  • domain assumption Sequencing depth N(t) and read counts are comparable across time points and cohorts after the CLR transform
    Sec. 3.2 assumes the depth-adaptive pseudocount stabilizes compositional states across variable sequencing depths; technical variability could otherwise dominate the signal.
  • domain assumption Candidate support St is well-defined and stable
    Eq. 7 depends on a union of observed and retained candidate clones; the retention rule is not specified, and the centering mean depends on it.
  • domain assumption Literature baseline values in Table 2 are accurate transcriptions of the cited reports
    Comparison against literature values is offered only as context, but any transcription or protocol mismatch would distort the headline comparison.

pith-pipeline@v1.3.0-alltime-deepseek · 10582 in / 7096 out tokens · 55363 ms · 2026-08-01T18:34:40.122790+00:00 · methodology

0 comments
read the original abstract

Longitudinal T cell receptor repertoires contain signals of clonal expansion, contraction, disappearance, and reappearance after immune perturbation. Static repertoire language models usually summarize a sample as a bag of sequences, so the sampling interval, sequencing depth, and clone presence pattern are only weakly represented. This paper presents DynImmune-BERT, a continuous time repertoire model for patient level immune status prediction. The method combines depth adaptive centered log ratio initialization, clone presence gated Neural ordinary differential equation dynamics, bounded neighborhood self attention, event based state restart, and a hybrid transport objective that supervises dominant and rare clone mass. A low rank meta adapter initializes reappearing clonotypes while keeping the parameter count independent of the number of observed clones. The evaluation separates literature reported baselines from internally controlled temporal comparisons, reports uncertainty for small external cohorts, adds calibration and threshold diagnostics, and visualizes latent clone trajectories and attention neighborhoods. The results indicate that event aware temporal modeling can complement strong static encoders when longitudinal repertoire structure is available, while small external cohorts and protocol differences require cautious interpretation.

Figures

Figures reproduced from arXiv: 2607.17244 by Haoyu Zhao, Long Zhang, Rong Fu, Shuo Yin, Wangyu Wu, Xiaowen Ma, Yiqing Lyu, Yongtai Liu.

Figure 1
Figure 1. Figure 1: Overview of DynImmune-BERT. Input TCR sequences, read counts, sampling times, and patient labels are [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Latent trajectory and attention diagnostics. The trajectory panel visualizes clone evolution and restart behavior, [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: ROC consistency check for DynImmune-BERT. The displayed AUC values correspond to the DynImmune [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation with seed variability. The ODE core is the largest driver of AUC, while event restart and adaptive [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Validation threshold selection using Youden’s index with bootstrap uncertainty. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Reliability diagram for predicted cancer probability. The diagonal line indicates perfect calibration. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

43 extracted references · 6 linked inside Pith

  1. [1]

    Longitudinal analysis of t and b cell receptor repertoire transcripts reveal dynamic immune response in covid-19 patients.Frontiers in Immunology, 11:582010, 2020

    Xuefeng Niu, Song Li, Pingchao Li, Wenjing Pan, Qian Wang, Ying Feng, Xiaoneng Mo, Qihong Yan, Xianmiao Ye, Jia Luo, et al. Longitudinal analysis of t and b cell receptor repertoire transcripts reveal dynamic immune response in covid-19 patients.Frontiers in Immunology, 11:582010, 2020

    2020

  2. [2]

    Longitudinal analysis reveals age-related changes in the t cell receptor repertoire of human t cell subsets.The Journal of clinical investigation, 132(17), 2022

    Xiaoping Sun, Thomas Nguyen, Achouak Achour, Annette Ko, Jeffrey Cifello, Chen Ling, Jay Sharma, Toyoko Hiroi, Yongqing Zhang, Chee W Chia, et al. Longitudinal analysis reveals age-related changes in the t cell receptor repertoire of human t cell subsets.The Journal of clinical investigation, 132(17), 2022

    2022

  3. [3]

    Longitudinal high- throughput tcr repertoire profiling reveals the dynamics of t-cell memory formation after mild covid-19 infection

    Anastasia A Minervina, Ekaterina A Komech, Aleksei Titov, Meriem Bensouda Koraichi, Elisa Rosati, Ilgar Z Mamedov, Andre Franke, Grigory A Efimov, Dmitriy M Chudakov, Thierry Mora, et al. Longitudinal high- throughput tcr repertoire profiling reveals the dynamics of t-cell memory formation after mild covid-19 infection. Elife, 10:e63502, 2021

    2021

  4. [4]

    Temporal development of t cell receptor repertoires during childhood in health and disease.JCI insight, 7(18):e161885, 2022

    Angela M Mitchell, Erin E Baschal, Kristen A McDaniel, Kimber M Simmons, Laura Pyle, Kathleen Waugh, Andrea K Steck, Liping Yu, Peter A Gottlieb, Marian J Rewers, et al. Temporal development of t cell receptor repertoires during childhood in health and disease.JCI insight, 7(18):e161885, 2022

    2022

  5. [5]

    Yi-Tung Chen, Hung-Chih Hsu, Yun-Shien Lee, Hsuan Liu, Bertrand Chin-Ming Tan, Chia-Yin Chin, Ian Yi-Feng Chang, and Chia-Yu Yang. Longitudinal high-throughput sequencing of the t-cell receptor repertoire reveals dynamic change and prognostic significance of peripheral blood tcr diversity in metastatic colorectal cancer during chemotherapy.Frontiers in Im...

    2022

  6. [6]

    NH Servaas, Fatiha Zaaraoui-Boutahar, CGK Wichers, Andrea Ottria, Eleni Chouri, AJ Affandi, Sandra Silva- Cardoso, Maarten van der Kroef, T Carvalheiro, Femke van Wijk, et al. Longitudinal analysis of t-cell receptor repertoires reveals persistence of antigen-driven cd4+ and cd8+ t-cell clusters in systemic sclerosis.Journal of Autoimmunity, 117:102574, 2021

    2021

  7. [7]

    Low t cell diversity associates with poor outcome in bladder cancer: A comprehensive longitudinal analysis of the t cell receptor repertoire.Cell Reports Medicine, 6(5), 2025

    Asbjørn Kjær, Nanna Kristjánsdóttir, Randi Istrup Juul, Iver Nordentoft, Karin Birkenkamp-Demtröder, Johanne Ahrenfeldt, Trine Strandgaard, Deema Radif, Darren Hodgson, Christopher Abbosh, et al. Low t cell diversity associates with poor outcome in bladder cancer: A comprehensive longitudinal analysis of the t cell receptor repertoire.Cell Reports Medicin...

    2025

  8. [8]

    Longitudinal t cell repertoire analysis reveals dynamic clonal t cell populations in ulcerative colitis.bioRxiv, pages 2025–01, 2025

    Kristi C Briggs, Jessica S Lin, Lara Chaaban, Alyssa Parian, Mark Lazarev, Florin Selaru, Franck Housseau, Kellie N Smith, and Joanna MP Melia. Longitudinal t cell repertoire analysis reveals dynamic clonal t cell populations in ulcerative colitis.bioRxiv, pages 2025–01, 2025

    2025

  9. [9]

    De novo prediction of cancer-associated t cell receptors for noninvasive cancer detection.Science translational medicine, 12(557):eaaz3738, 2020

    Daria Beshnova, Jianfeng Ye, Oreoluwa Onabolu, Benjamin Moon, Wenxin Zheng, Yang-Xin Fu, James Bru- garolas, Jayanthi Lea, and Bo Li. De novo prediction of cancer-associated t cell receptors for noninvasive cancer detection.Science translational medicine, 12(557):eaaz3738, 2020

    2020

  10. [10]

    Deeplion: deep multi-instance learning improves the prediction of cancer-associated t cell receptors for accurate cancer detection.Frontiers in genetics, 13:860510, 2022

    Ying Xu, Xinyang Qian, Xuanping Zhang, Xin Lai, Yuqian Liu, and Jiayin Wang. Deeplion: deep multi-instance learning improves the prediction of cancer-associated t cell receptors for accurate cancer detection.Frontiers in genetics, 13:860510, 2022

    2022

  11. [11]

    Xinyang Qian, Guang Yang, Fan Li, Xuanping Zhang, Xiaoyan Zhu, Xin Lai, Xiao Xiao, Tao Wang, and Jiayin Wang. Deeplion2: deep multi-instance contrastive learning framework enhancing the prediction of cancer- associated t cell receptors by attention strategy on motifs.Frontiers in Immunology, 15:1345586, 2024

    2024

  12. [12]

    Deeptcr is a deep learning framework for revealing sequence concepts within t-cell repertoires.Nature communications, 12(1):1605, 2021

    John-William Sidhom, H Benjamin Larman, Drew M Pardoll, and Alexander S Baras. Deeptcr is a deep learning framework for revealing sequence concepts within t-cell repertoires.Nature communications, 12(1):1605, 2021

    2021

  13. [13]

    Berttcr: a bert-based deep learning framework for predicting cancer-related immune status based on t cell receptor repertoire.Briefings in Bioinformatics, 25(5):bbae420, 2024

    Min Zhang, Qi Cheng, Zhenyu Wei, Jiayu Xu, Shiwei Wu, Nan Xu, Chengkui Zhao, Lei Yu, and Weixing Feng. Berttcr: a bert-based deep learning framework for predicting cancer-related immune status based on t cell receptor repertoire.Briefings in Bioinformatics, 25(5):bbae420, 2024

    2024

  14. [14]

    Methpriorgcn: a deep learning tool for inferring dna methylation prior knowledge and guiding personalized medicine.Briefings in Bioinformatics, 26(2):bbaf131, 2025

    Jie Ni, Bin Li, Shumei Miao, Xinting Zhang, Donghui Yan, Shengqi Jing, Shan Lu, Zhuoying Xie, Xin Zhang, and Yun Liu. Methpriorgcn: a deep learning tool for inferring dna methylation prior knowledge and guiding personalized medicine.Briefings in Bioinformatics, 26(2):bbaf131, 2025

    2025

  15. [15]

    Learning immune receptor representations with protein language models.arXiv preprint arXiv:2402.03823, 2024

    Andreas Dounas, Tudor-Stefan Cotet, and Alexander Yermanos. Learning immune receptor representations with protein language models.arXiv preprint arXiv:2402.03823, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    Sequence-based tcr-peptide representations using cross-epitope contrastive fine-tuning of protein language models

    Chiho Im, Ryan Zhao, Scott D Boyd, and Anshul Kundaje. Sequence-based tcr-peptide representations using cross-epitope contrastive fine-tuning of protein language models. InInternational Conference on Research in Computational Molecular Biology, pages 34–48. Springer, 2025

    2025

  17. [17]

    A large language model for predicting t cell receptor-antigen binding specificity

    Xing Fang, Chenpeng Yu, Shiye Tian, and Hui Liu. A large language model for predicting t cell receptor-antigen binding specificity. In2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pages 1489–1492. IEEE, 2024

    2024

  18. [18]

    Deciphering the language of antibodies using self-supervised learning.Patterns, 3(7), 2022

    Jinwoo Leem, Laura S Mitchell, James HR Farmery, Justin Barton, and Jacob D Galson. Deciphering the language of antibodies using self-supervised learning.Patterns, 3(7), 2022

    2022

  19. [19]

    Deep learning-based prediction of the t cell receptor–antigen binding specificity

    Tianshi Lu, Ze Zhang, James Zhu, Yunguan Wang, Peixin Jiang, Xue Xiao, Chantale Bernatchez, John V Heymach, Don L Gibbons, Jun Wang, et al. Deep learning-based prediction of the t cell receptor–antigen binding specificity. Nature machine intelligence, 3(10):864–875, 2021

    2021

  20. [20]

    The deep learning framework icantcr enables early cancer detection using the t-cell receptor repertoire in peripheral blood.Cancer Research, 84(11):1915–1928, 2024

    Yideng Cai, Meng Luo, Wenyi Yang, Chang Xu, Pingping Wang, Guangfu Xue, Xiyun Jin, Rui Cheng, Jinhao Que, Wenyang Zhou, et al. The deep learning framework icantcr enables early cancer detection using the t-cell receptor repertoire in peripheral blood.Cancer Research, 84(11):1915–1928, 2024

    1915

  21. [21]

    Machine learning for cancer immunotherapies based on epitope recognition by t cell receptors.Frontiers in genetics, 10:1141, 2019

    Anja Mösch, Silke Raffegerst, Manon Weis, Dolores J Schendel, and Dmitrij Frishman. Machine learning for cancer immunotherapies based on epitope recognition by t cell receptors.Frontiers in genetics, 10:1141, 2019

    2019

  22. [22]

    Circulating t-cell receptor repertoire for cancer early detection.NPJ Precision Oncology, 9(1):245, 2025

    Yilong Li, Michelle Nahas, Dennis Stephens, Kate Froburg, Emma Hintz, Devin Champagne, Amaneet Lochab, Markus Brown, Jasper Braun, María Antonia Fortuño, et al. Circulating t-cell receptor repertoire for cancer early detection.NPJ Precision Oncology, 9(1):245, 2025. 12 DynImmune-BERT

    2025

  23. [23]

    Disease diagnostics using machine learning of b cell and t cell receptor sequences.Science, 387(6736):eadp2407, 2025

    Maxim E Zaslavsky, Erin Craig, Jackson K Michuda, Nidhi Sehgal, Nikhil Ram-Mohan, Ji-Yeun Lee, Khoa D Nguyen, Ramona A Hoh, Tho D Pham, Katharina Röltgen, et al. Disease diagnostics using machine learning of b cell and t cell receptor sequences.Science, 387(6736):eadp2407, 2025

    2025

  24. [24]

    Reading the repertoire: Progress in adaptive immune receptor analysis using machine learning.Cell Systems, 15(12):1168–1189, 2024

    Timothy J O’Donnell, Chakravarthi Kanduri, Giulio Isacchini, Julien P Limenitakis, Rebecca A Brachman, Raymond A Alvarez, Ingrid H Haff, Geir K Sandve, and Victor Greiff. Reading the repertoire: Progress in adaptive immune receptor analysis using machine learning.Cell Systems, 15(12):1168–1189, 2024

    2024

  25. [25]

    Neural ordinary differential equations.Advances in neural information processing systems, 31, 2018

    Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations.Advances in neural information processing systems, 31, 2018

    2018

  26. [26]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    2017

  27. [27]

    Continuous self-attention models with neural ode networks

    Jing Zhang, Peng Zhang, Baiwen Kong, Junqiu Wei, and Xin Jiang. Continuous self-attention models with neural ode networks. InProceedings of the AAAI conference on artificial intelligence, volume 35, pages 14393–14401, 2021

    2021

  28. [28]

    Contiformer: Continuous- time transformer for irregular time series modeling.Advances in Neural Information Processing Systems, 36: 47143–47175, 2023

    Yuqi Chen, Kan Ren, Yansen Wang, Yuchen Fang, Weiwei Sun, and Dongsheng Li. Contiformer: Continuous- time transformer for irregular time series modeling.Advances in Neural Information Processing Systems, 36: 47143–47175, 2023

    2023

  29. [29]

    Diffode: Neural ode with differentiable hidden state for irregular time series analysis

    Yudong Zhang, Xu Wang, Xuan Yu, Zhengyang Zhou, Xing Xu, Lei Bai, and Yang Wang. Diffode: Neural ode with differentiable hidden state for irregular time series analysis. In2025 IEEE 41st International Conference on Data Engineering (ICDE), pages 1–14. IEEE, 2025

    2025

  30. [30]

    T-cell repertoire analysis and metrics of diversity and clonality.Current opinion in biotechnology, 65:284–295, 2020

    Johanna Chiffelle, Raphael Genolet, Marta AS Perez, George Coukos, Vincent Zoete, and Alexandre Harari. T-cell repertoire analysis and metrics of diversity and clonality.Current opinion in biotechnology, 65:284–295, 2020

    2020

  31. [31]

    High-throughput and single-cell t cell receptor sequencing technologies

    Joy A Pai and Ansuman T Satpathy. High-throughput and single-cell t cell receptor sequencing technologies. Nature methods, 18(8):881–892, 2021

    2021

  32. [32]

    Multiple instance neural networks based on sparse attention for cancer detection using t-cell receptor sequences.BMC bioinformatics, 23(1):469, 2022

    Younghoon Kim, Tao Wang, Danyi Xiong, Xinlei Wang, and Seongoh Park. Multiple instance neural networks based on sparse attention for cancer detection using t-cell receptor sequences.BMC bioinformatics, 23(1):469, 2022

    2022

  33. [33]

    Transmil: Transformer based correlated multiple instance learning for whole slide image classification.Advances in neural information processing systems, 34:2136–2147, 2021

    Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification.Advances in neural information processing systems, 34:2136–2147, 2021

    2021

  34. [34]

    Biformer: Vision transformer with bi-level routing attention

    Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson WH Lau. Biformer: Vision transformer with bi-level routing attention. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10323–10333, 2023

    2023

  35. [35]

    Neural ode transformers: Analyzing internal dynamics and adaptive fine-tuning.arXiv preprint arXiv:2503.01329, 2025

    Anh Tong, Thanh Nguyen-Tang, Dongeun Lee, Duc Nguyen, Toan Tran, David Hall, Cheongwoong Kang, and Jaesik Choi. Neural ode transformers: Analyzing internal dynamics and adaptive fine-tuning.arXiv preprint arXiv:2503.01329, 2025

    Pith/arXiv arXiv 2025

  36. [36]

    Online sinkhorn: Optimal transport distances from sample streams.Advances in Neural Information Processing Systems, 33:1657–1667, 2020

    Arthur Mensch and Gabriel Peyré. Online sinkhorn: Optimal transport distances from sample streams.Advances in Neural Information Processing Systems, 33:1657–1667, 2020

    2020

  37. [37]

    A trainable optimal transport embedding for feature aggregation and its relationship to attention.arXiv preprint arXiv:2006.12065, 2020

    Grégoire Mialon, Dexiong Chen, Alexandre d’Aspremont, and Julien Mairal. A trainable optimal transport embedding for feature aggregation and its relationship to attention.arXiv preprint arXiv:2006.12065, 2020

    Pith/arXiv arXiv 2006

  38. [38]

    Energy-based sliced wasserstein distance.Advances in Neural Information Processing Systems, 36:18046–18075, 2023

    Khai Nguyen and Nhat Ho. Energy-based sliced wasserstein distance.Advances in Neural Information Processing Systems, 36:18046–18075, 2023

    2023

  39. [39]

    A sliced wasserstein loss for neural texture synthesis

    Eric Heitz, Kenneth Vanhoey, Thomas Chambon, and Laurent Belcour. A sliced wasserstein loss for neural texture synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9412–9420, 2021

    2021

  40. [40]

    Lora-fa: Memory-efficient low-rank adaptation for large language models fine-tuning.arXiv preprint arXiv:2308.03303, 2023

    Longteng Zhang, Lin Zhang, Shaohuai Shi, Xiaowen Chu, and Bo Li. Lora-fa: Memory-efficient low-rank adaptation for large language models fine-tuning.arXiv preprint arXiv:2308.03303, 2023. 13 DynImmune-BERT

    Pith/arXiv arXiv 2023

  41. [41]

    Autolora: Automatically tuning matrix ranks in low-rank adaptation based on meta learning.arXiv preprint arXiv:2403.09113, 2024

    Ruiyi Zhang, Rushi Qiang, Sai Ashish Somayajula, and Pengtao Xie. Autolora: Automatically tuning matrix ranks in low-rank adaptation based on meta learning.arXiv preprint arXiv:2403.09113, 2024

    Pith/arXiv arXiv 2024

  42. [42]

    Lora+: Efficient low rank adaptation of large models.arXiv preprint arXiv:2402.12354, 2024

    Soufiane Hayou, Nikhil Ghosh, and Bin Yu. Lora+: Efficient low rank adaptation of large models.arXiv preprint arXiv:2402.12354, 2024

    Pith/arXiv arXiv 2024

  43. [43]

    Spatial heterogeneity of the t cell receptor repertoire reflects the mutational landscape in lung cancer.Nature medicine, 25(10):1549–1559, 2019

    Kroopa Joshi, Marc Robert de Massy, Mazlina Ismail, James L Reading, Imran Uddin, Annemarie Woolston, Emine Hatipoglu, Theres Oakes, Rachel Rosenthal, Thomas Peacock, et al. Spatial heterogeneity of the t cell receptor repertoire reflects the mutational landscape in lung cancer.Nature medicine, 25(10):1549–1559, 2019. 14

    2019