Research map / Label-efficient and robust learning / Recognition and long-tail
A Label Disambiguation-Based Multimodal Massive Multiple Instance Learning Approach for Immune Repertoire Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LaDM3IL, an instance-level massive MIL approach with prototype-based label disambiguation and gated tensor-fusion of gene and AA-sequence features for immune repertoire classification and receptor identification.
Problem
Immune repertoire classification with massive numbers of receptor instances per bag and very low witness rate, where bag-level labels assigned to instances are misleading and bag-level MIL is computationally expensive.
Approach
Uses instance-level MIL to cut computation, and iteratively corrects noisy instance labels via prototype similarity (partial-label-learning-inspired disambiguation), combined with a gated-attention and tensor-fusion encoder over V(D)J gene embeddings and a pre-trained AA-sequence BERT.
Key results
- On CMV, LaDM3IL reached AUC 84.88±2.20, ACC 78.80, F1 78.56, versus DeepRC AUC 83.10 and NLLIRC AUC 83.24.
- On the Cancer dataset, LaDM3IL reached AUC 90.22±0.33, outperforming NLLIRC (89.03) and DeepTCR (86.60).
- Ablation shows removing label disambiguation lowers CMV AUC from 85.44 to 82.55; receptor-level AUC of 0.767 for 164 known CMV-associated TCRs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LaDM3IL instead of DeepRC | CMV repertoire classification · CMV | AUC: 83.10±2.20 -> 84.88±2.20 | Helped |
| use LaDM3IL instead of NLLIRC | CMV repertoire classification · CMV | accuracy: 76.91±3.00 -> 78.80±2.30 | Helped |
| use LaDM3IL instead of NLLIRC | CMV repertoire classification · CMV | F1 score: 76.37±3.00 -> 78.56±2.40 | Helped |
| use LaDM3IL instead of NLLIRC | Cancer-associated receptor identification · Cancer | AUC: 89.03±0.06 -> 90.22±0.33 | Helped |
| use LaDM3IL instead of NLLIRC | Cancer-associated receptor identification · Cancer | accuracy: 80.13±0.40 -> 83.00±0.57 | Helped |
| use LaDM3IL instead of NLLIRC | Cancer-associated receptor identification · Cancer | F1 score: 78.79±0.30 -> 80.59±0.37 | Helped |
| add label disambiguation (ablation) | LaDM3IL · CMV | ACU: 82.55 -> 85.44 | Helped |
| add multimodal fusion of VDJ genes and AA sequence (ablation) | LaDM3IL · CMV | ACU: 82.94 -> 85.44 | Helped |
| use pretrained sequence encoder weights (ablation) | LaDM3IL · CMV | ACU: 81.52 -> 85.44 | Helped |
| add label disambiguation (ablation) | LaDM3IL · Cancer | ACU: 89.58 -> 89.98 | Helped |
| use pretrained sequence encoder weights (ablation) | LaDM3IL · Cancer | ACU: 89.55 -> 89.98 | Helped |
Limitations
- D gene information excluded due to absence in many receptors
- Performance decreases as receptor sequence length increases (Table 3)
- Cancer dataset lacks VDJ gene information so multimodal fusion ablation only on CMV
Experimental setup
Datasets: CMV dataset (Emerson et al. 2017) Cancer dataset (Emerson et al. 2017; Beshnova et al. 2020)
Benchmarks: CMV dataset Cancer dataset
Metrics: AUC ACC F1 score
Baselines: DeepRC DeepTCR NLLIRC DeepCAT
Builds on: multiple instance learning partial label learning (PICO) DeepRC tensor fusion gated attention
Closest papers by meaning
- Scalable Universal T-Cell Receptor Embeddings from Adaptive Immune Repertoires (ICLR 2025 · similarity 0.87)
- Immunogenicity Prediction with Dual Attention Enables Vaccine Target Selection (ICLR 2025 · similarity 0.82)
- Self Iterative Label Refinement via Robust Unlabeled Learning (NeurIPS 2025 · similarity 0.82)
- Self-Paced Unified Representation Learning for Hierarchical Multi-Label Classification (AAAI 2024 · similarity 0.82)
- Context-Aware Regularization with Markovian Integration for Attention-Based Nucleotide Analysis (NeurIPS 2025 · similarity 0.81)
- Distilling Reliable Knowledge for Instance-Dependent Partial Label Learning (AAAI 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.