Research map / Vision-language and multimodal / CLIP and zero-shot

C3P: Contrastive promoter-protein pretraining yields representations capturing bacterial gene regulation

NeurIPS 2026 · Vision-language and multimodal › CLIP and zero-shot · approach: Recaptioning and text-supervised pretraining · 0 citations (OpenAlex)

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.

Introduces C3P, a CLIP-style contrastive model that aligns bacterial promoter sequences to frozen ESM2 protein embeddings, trained on 88M promoter-protein pairs, and shows it outperforms genome language models on E. coli regulon/sigmulon classification and zero-shot co-regulated gene retrieval.

Problem

Genome language models trained with reconstruction objectives appear to learn little about bacterial gene regulation, and there is no scalable unsupervised way to infer co-regulation from a genome alone.

Approach

Align promoter representations to the embeddings of their downstream proteins from a frozen pretrained protein language model using symmetric InfoNCE, so promoter similarity is driven by protein functional similarity rather than sequence reconstruction.

Key results

• C3P (100M) reaches 0.212 1-nn precision on E. coli regulons versus 0.088 for ESM2 and 0.037 for Evo2 (7B).
• On zero-shot co-regulated gene retrieval across 25 iModulonDB datasets, C3P (100M) achieves 0.242 top-1 full match accuracy versus 0.066 for Evo2 and 0.125 for ESM2.
• C3P significantly beats its random initialization on all 25 datasets, whereas gLM gains are inconsistent (e.g., Evo2 improves on 18/25 for full match).

Surprising findings

• Evo (7B) shows no improvement over its random initialization on any of the 25 datasets for either metric.
• ESM2 protein embeddings and C3P promoter embeddings produce qualitatively different clustering on B. subtilis, with no obvious co-regulatory pattern in ESM2 UMAP.

Limitations

• No interpretability analysis of learned promoter features has been performed.
• Comparison with gLMs is confounded by C3P using a pretrained protein encoder while gLMs are trained from scratch.
• Scaling analysis is limited to varying promoter encoder size, and C3P cannot generate sequences.

Experimental setup

Datasets: RefSeq bacterial genomes (40,000 assemblies) RegulonDB iModulonDB (25 datasets) promoter strength datasets

Benchmarks: RegulonDB regulon classification (E. coli, n=102) RegulonDB sigmulon classification (n=6) iModulonDB zero-shot co-regulated gene retrieval (25 datasets)

Metrics: 1-nn precision 1-nn MCC mAP@R top-1 full match accuracy top-1 any match accuracy Spearman correlation

Baselines: ESM2 (150M) gLM2 (650M) Evo2 (7B) Evo (7B) PromoGen2 (149M) Promoter Calculator randomly initialized baselines

Builds on: CLIP ESM2 InfoNCE Evo gLM2

Closest papers by meaning

• Training Compute-Optimal Protein Language Models (NeurIPS 2024 · similarity 0.86)
• Feature Reuse and Scaling: Understanding Transfer Learning with Protein Language Models (ICML 2024 · similarity 0.85)
• What DNA Foundation Models Learn Beyond Sequence Composition (NeurIPS 2026 · similarity 0.85)
• STRAND: Sequence-Conditioned Transport for Single-Cell Perturbations (NeurIPS 2026 · similarity 0.85)
• LEMON-ZEST: Evolution-Informed Tokenization for Efficient Protein Language Modeling (NeurIPS 2026 · similarity 0.85)
• Prot2Text-V2: Protein Function Prediction with Multimodal Contrastive Alignment (NeurIPS 2025 · similarity 0.84)