Research map / Probabilistic and scientific modeling / Molecules and proteins
Designing Cell-Type-Specific Promoter Sequences Using Conservative Model-Based Optimization
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.
Builds a data-efficient offline model-based optimization workflow (pretrained MTLucifer fine-tuned with conservative regularization, gradient-ascent design, ensemble-based diverse selection) to design cell-type-specific 250bp promoters for Jurkat, K562, and THP1 leukemia cells, validated experimentally.
Problem
Designing short cell-type-specific promoters for closely related cell types is hard: existing methods need manual curation or large expression datasets, and optimization of learned models yields adversarial designs.
Approach
Adapts conservative objective models (COMs) to multi-cell-type promoter design by regularizing predicted differential expression during fine-tuning of pretrained sequence models, then designs via gradient ascent and selects a diverse, low-uncertainty subset using an ensemble lower confidence bound.
Key results
- Designed sequences improved over starting sequences in 72.12% of cases for Jurkat and 80.48% for K562.
- A K562 promoter reached DE of 5.17 vs 2.94 for the best fine-tuning sequence (75.85% higher).
- Motif tiling almost always produced worse sequences than starting sequences, and the workflow failed to improve most THP1 starting sequences.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use conservative model-based optimization with gradient ascent instead of motif tiling | Cell-type-specific promoter design · Jurkat and K562 | cell-type-specificity: improving 72.12% of sequences in Jurkat cells and 80.48% in K562 cells | Helped |
| Use conservative model-based optimization with gradient ascent instead of motif tiling | Cell-type-specific promoter design · K562 | cell-type-specificity (DE): 75.85% higher DE; 5.17 vs 2.94 | Helped |
| Use conservative model-based optimization with gradient ascent instead of motif tiling | Cell-type-specific promoter design · THP1 | cell-type-specificity | Hurt |
| Replace gradient ascent with DENs (ablation) | Our promoter-design workflow · Jurkat, K562, and THP1 | cell-type-specificity (DE) | Hurt |
| Tile motifs into high-DE background sequences | Motif tiling promoter-design baseline · Jurkat, K562, and THP1 | cell-type-specificity (DE) | Hurt |
| Use DENs instead of gradient ascent (ablation) | Promoter-design workflow · THP1 | mean base pair entropy: 1.96 -> 1.00 | Hurt |
Surprising findings
- Motif tiling, a traditional heuristic, almost always led to worse sequences when improving high-DE starting sequences.
- Gradient ascent generally outperformed the generative DEN optimizer, and DENs produced far less diverse THP1 designs (entropy 1.00).
Limitations
- Performance depends on design model accuracy; failed for THP1 due to noisy data and few high-DE sequences.
- Pretraining data mostly from K562, likely favoring K562 performance over Jurkat.
- Conservative regularization may reduce diversity when the fine-tuning dataset has limited sequence diversity.
Experimental setup
Datasets: Sharpr-MPRA SuRE MPRA Reddy et al. promoter expression dataset (17,104 measurements)
Benchmarks: experimentally measured promoter differential expression in Jurkat, K562, THP1
Metrics: percentile score of differential expression base pair entropy mean Hamming distance Spearman correlation Pearson r
Baselines: Motif tiling DENs
Builds on: conservative objective models Enformer deep exploration networks MTLucifer
Closest papers by meaning
- Regulatory DNA Sequence Design with Reinforcement Learning (ICLR 2025 · similarity 0.85)
- STRAND: Sequence-Conditioned Transport for Single-Cell Perturbations (NeurIPS 2026 · similarity 0.85)
- C3P: Contrastive promoter-protein pretraining yields representations capturing bacterial gene regulation (NeurIPS 2026 · similarity 0.84)
- A Simple and Comprehensive Benchmark for Single-Cell Transcriptomics (AAAI 2025 · similarity 0.83)
- Absorb & Escape: Overcoming Single Model Limitations in Generating Heterogeneous Genomic Sequences (NeurIPS 2024 · similarity 0.83)
- PRESCRIBE: Predicting Single-Cell Responses with Bayesian Estimation (NeurIPS 2025 · similarity 0.82)
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.