Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
DecoderTCR: Compositional Pretraining and Entropy-Guided Decoding for TCR-pMHC Interactions
Boqiao Lai, Melissa Englund, Ramit Bharanikumar, Isabel Nocedal, Ali Davariashtiyani, Jason Perera, Aly A Khan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Stage 1 domain-specific continual pretraining on pMHC and TCR data | ESM-2 (3B) · Zero-shot pMHC binding, Class I | AUROC: 0.45 -> 0.93 | Improved |
| add Stage 1 domain-specific pretraining to ESM-2 embeddings | ESM-2 (3B) · ESCAPE-seq oncogene pMHC | AUROC: 0.63 -> 0.77 | Improved |
| scale DecoderTCR from 650M to 3B | DecoderTCR Stage 1 · Zero-shot pMHC Class I | AUROC: .96±.04 -> .93±.08 | No clear effect |
| add Stage 2 paired pretraining after Stage 1 (ablation) | DecoderTCR (3B) · YLQ TCR recognition | AUROC: 0.18 -> 0.76 | Improved |
| add Stage 1 unpaired pretraining before Stage 2 (ablation) | DecoderTCR (3B) · YLQ TCR recognition | AUROC: 0.56 -> 0.76 | Improved |
| add Stage 2 paired training on top of Stage 1 (knowledge retention) | DecoderTCR · Held-out TCR sequences | inverse perplexity: 0.72 -> 0.79 | Improved |
| continue Stage 2 training for 3500 steps | DecoderTCR · Zero-shot pMHC held-out alleles | AUROC: 0.925-0.93 | No clear effect |
| use IEGR entropy-guided decoding instead of Greedy (ablation) | DecoderTCR designs · YLQ in silico design | salt-bridge enrichment: 17.2% -> 34.2% | Improved |
| use IEGR entropy-guided decoding instead of OneShot (ablation) | DecoderTCR designs · YLQ in silico design | salt-bridge enrichment: 21.9% -> 34.2% | Improved |
| use IEGR instead of single-pass decoding (ablation) | DecoderTCR designs · YLQ AlphaFold3 ipTM | ipTM significance: p < 0.001 | Improved |
| add classifier filtering on DecoderTCR embeddings (lab-in-the-loop round 2) | IEGR designs · YLQ wet-lab hit rate | hit rate: 3% -> 12% | Improved |
| use zero-shot DecoderTCR vs supervised NetTCR-2.2 | DecoderTCR (3B) · YLQ TCR recognition | AUROC: 0.88 -> 0.76 | Hurt |
Closest papers by meaning
- A Label Disambiguation-Based Multimodal Massive Multiple Instance Learning Approach for Immune Repertoire Classification (AAAI 2024 · similarity 0.84)
- ProSST: Protein Language Modeling with Quantized Structure and Disentangled Attention (NeurIPS 2024 · similarity 0.81)
- Relation-Aware Equivariant Graph Networks for Epitope-Unknown Antibody Design and Specificity Optimization (AAAI 2025 · similarity 0.80)
- mRNA2vec: mRNA Embedding with Language Model in the 5'UTR-CDS for mRNA Design (AAAI 2025 · similarity 0.80)
- From Mechanistic Interpretability to Mechanistic Biology: Training, Evaluating, and Interpreting Sparse Autoencoders on Protein Language Models (ICML 2025 · similarity 0.80)
- Entropy-Reinforced Planning with Large Language Models for Drug Discovery (ICML 2024 · similarity 0.80)
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.
