mRNABench: A curated benchmark for mature mRNA property and function prediction

What this source contributed

Records in this release that cite this source. A record can cite several sources.

• 3 methods
• 19 configurations
• 60 pipelines
• 1 benchmark
• 11 tasks
• 8 protocols
• 10 dataset subsets
• 610 evaluations
• 610 results
• 103 evidence claims
• 1 use case

Results

Values as the source printed them. Compare results only under compatible protocols, datasets and metrics.

Claims

Statements this source supports, each with where in the source it was checked.

• aido-rna-1b600m-cds: description: Description Linear probe, trained per mRNABench task, on frozen embeddings averaged over nucleotides from aido-rna-1b600m-cds, an AIDO.RNA checkpoint domain-adapted for mRNA on coding sequences; AIDO.RNA is a Transformer RNA model trained on 42 million non-coding RNAs from RNAcentral. (expansion-p3-mrnabench-2025: Appendix A, Overview of Evaluated Models (family and variant rows; naive baselines paragraph); Section 4.3, Linear Probing ('training a linear classifier on frozen embeddings'; 'averaging over per-nucleotide embeddings'))
• aido-rna-1b600m: description: Description Linear probe, trained per mRNABench task, on frozen embeddings averaged over nucleotides from aido-rna-1b600m, an AIDO.RNA checkpoint; AIDO.RNA is a Transformer RNA model trained on 42 million non-coding RNAs from RNAcentral. (expansion-p3-mrnabench-2025: Appendix A, Overview of Evaluated Models (family and variant rows; naive baselines paragraph); Section 4.3, Linear Probing ('training a linear classifier on frozen embeddings'; 'averaging over per-nucleotide embeddings'))
• aido-rna-650m-cds: description: Description Linear probe, trained per mRNABench task, on frozen embeddings averaged over nucleotides from aido-rna-650m-cds, an AIDO.RNA checkpoint domain-adapted for mRNA on coding sequences; AIDO.RNA is a Transformer RNA model trained on 42 million non-coding RNAs from RNAcentral. (expansion-p3-mrnabench-2025: Appendix A, Overview of Evaluated Models (family and variant rows; naive baselines paragraph); Section 4.3, Linear Probing ('training a linear classifier on frozen embeddings'; 'averaging over per-nucleotide embeddings'))
• aido-rna-650m: description: Description Linear probe, trained per mRNABench task, on frozen embeddings averaged over nucleotides from aido-rna-650m, an AIDO.RNA checkpoint; AIDO.RNA is a Transformer RNA model trained on 42 million non-coding RNAs from RNAcentral. (expansion-p3-mrnabench-2025: Appendix A, Overview of Evaluated Models (family and variant rows; naive baselines paragraph); Section 4.3, Linear Probing ('training a linear classifier on frozen embeddings'; 'averaging over per-nucleotide embeddings'))
• AIDO-RNA: description: Description mRNABench evaluation of mature mRNA embeddings from AIDO.RNA, an RNA foundation model pretrained on 42 million non-coding RNA sequences at single-nucleotide resolution, with task scores quoted from prior work in MIMIC Table S11. (coverage-source-mimic-2026-v1-html: Appendix D.3 and Table S11; expansion-p3-mrnabench-2025: abstract; qa-descriptions-20261010-source-aido-rna: abstract)
• Dilated ResNet: description: Description Supervised dilated convolutional network baseline, built from three or four DilatedResNet blocks and trained from scratch on each task's training split. (expansion-p3-mrnabench-2025: Appendix A, Overview of Evaluated Models (baselines paragraph); coverage-source-mimic-2026-v1-html: Table S11 and Appendix D.3)
• dnabert-s: description: Description Linear probe, trained per mRNABench task, on frozen embeddings averaged over nucleotides from dnabert-s, a DNABERT-S checkpoint; DNABERT-S is a Transformer DNA model trained by contrastive learning on prokaryotic genomes. (expansion-p3-mrnabench-2025: Appendix A, Overview of Evaluated Models (family and variant rows; naive baselines paragraph); Section 4.3, Linear Probing ('training a linear classifier on frozen embeddings'; 'averaging over per-nucleotide embeddings'))
• DNABERT-S: description: Description mRNABench evaluation of mature mRNA embeddings from DNABERT-S, a genome model built on DNABERT-2 and trained with contrastive objectives to give species-aware DNA embeddings, with task scores quoted from prior work in MIMIC Table S11. (coverage-source-mimic-2026-v1-html: Appendix D.3 and Table S11; expansion-p3-mrnabench-2025: abstract; qa-descriptions-20261010-source-dnabert-s: abstract)
• dnabert2: description: Description Linear probe, trained per mRNABench task, on frozen embeddings averaged over nucleotides from dnabert2, a DNABERT2 checkpoint; DNABERT2 is a Transformer DNA model trained on a multispecies dataset with byte-pair-encoding tokens. (expansion-p3-mrnabench-2025: Appendix A, Overview of Evaluated Models (family and variant rows; naive baselines paragraph); Section 4.3, Linear Probing ('training a linear classifier on frozen embeddings'; 'averaging over per-nucleotide embeddings'))
• DNABERT2: description: Description mRNABench evaluation of mature mRNA embeddings from DNABERT-2, a genome foundation model that tokenises DNA with byte pair encoding instead of k-mers, with task scores quoted from prior work in MIMIC Table S11. (coverage-source-mimic-2026-v1-html: Appendix D.3 and Table S11; expansion-p3-mrnabench-2025: abstract; qa-descriptions-20261010-source-dnabert2: abstract)
• DNABERT2: family DNABERT-2: Family DNABERT-2 (mRNABench Table 2 DNABERT2 row; Appendix model inventory T1 DNABERT2 description; source-labelled configuration DNABERT2 | Existing reviewed locator: Table 2, row(DNABERT2))
• dnabert2: uses model catalog-model-dnabert-2: Uses model DNABERT-2 (Tables 5/6, row dnabert2; Appendix A (APP1), C (APP3), D (APP4))
• dnabert2: uses model discovery-model-dnabert-2: Uses model DNABERT-2 (Tables 5/6, row dnabert2; Appendix A (APP1), C (APP3), D (APP4))
• ERNIE-RNA: description: Description mRNABench evaluation of mature mRNA embeddings from ERNIE-RNA, an RNA language model based on a modified BERT with a base-pairing restriction, with task scores quoted from prior work in MIMIC Table S11. (coverage-source-mimic-2026-v1-html: Appendix D.3 and Table S11; expansion-p3-mrnabench-2025: abstract; qa-descriptions-20261010-source-ernie-rna: abstract)
• ernierna-ss: description: Description Linear probe, trained per mRNABench task, on frozen embeddings averaged over nucleotides from ernierna-ss, an ERNIE-RNA checkpoint fine-tuned on secondary structure; ERNIE-RNA is a Transformer RNA model pretrained on 20 million non-coding RNAs with structural information added as attention-mask biases. (expansion-p3-mrnabench-2025: Appendix A, Overview of Evaluated Models (family and variant rows; naive baselines paragraph); Section 4.3, Linear Probing ('training a linear classifier on frozen embeddings'; 'averaging over per-nucleotide embeddings'))
• ernierna: description: Description Linear probe, trained per mRNABench task, on frozen embeddings averaged over nucleotides from ernierna, an ERNIE-RNA checkpoint; ERNIE-RNA is a Transformer RNA model pretrained on 20 million non-coding RNAs with structural information added as attention-mask biases. (expansion-p3-mrnabench-2025: Appendix A, Overview of Evaluated Models (family and variant rows; naive baselines paragraph); Section 4.3, Linear Probing ('training a linear classifier on frozen embeddings'; 'averaging over per-nucleotide embeddings'))
• evo-1–8k-base: description: Description Linear probe, trained per mRNABench task, on frozen embeddings averaged over nucleotides from evo-1-8k-base, an Evo checkpoint that is the base model; Evo is a StripedHyena DNA model trained autoregressively on the OpenGenome dataset. (expansion-p3-mrnabench-2025: Appendix A, Overview of Evaluated Models (family and variant rows; naive baselines paragraph); Section 4.3, Linear Probing ('training a linear classifier on frozen embeddings'; 'averaging over per-nucleotide embeddings'))
• evo-1–131k-base: description: Description Linear probe, trained per mRNABench task, on frozen embeddings averaged over nucleotides from evo-1-131k-base, an Evo checkpoint trained with extended context; Evo is a StripedHyena DNA model trained autoregressively on the OpenGenome dataset. (expansion-p3-mrnabench-2025: Appendix A, Overview of Evaluated Models (family and variant rows; naive baselines paragraph); Section 4.3, Linear Probing ('training a linear classifier on frozen embeddings'; 'averaging over per-nucleotide embeddings'))
• evo-1.5–8k-base: description: Description Linear probe, trained per mRNABench task, on frozen embeddings averaged over nucleotides from evo-1.5-8k-base, an Evo checkpoint trained on an Evo 1 dataset enlarged by 50%; Evo is a StripedHyena DNA model trained autoregressively on the OpenGenome dataset. (expansion-p3-mrnabench-2025: Appendix A, Overview of Evaluated Models (family and variant rows; naive baselines paragraph); Section 4.3, Linear Probing ('training a linear classifier on frozen embeddings'; 'averaging over per-nucleotide embeddings'))
• evo2–1b-base: description: Description Linear probe, trained per mRNABench task, on frozen embeddings averaged over nucleotides from evo2-1b-base, an Evo 2 checkpoint trained as a base model on a mixture of genomic regions including mRNA; Evo 2 is a StripedHyena2 DNA model trained autoregressively on OpenGenome2, 9 trillion nucleotides across all domains of life. (expansion-p3-mrnabench-2025: Appendix A, Overview of Evaluated Models (family and variant rows; naive baselines paragraph); Section 4.3, Linear Probing ('training a linear classifier on frozen embeddings'; 'averaging over per-nucleotide embeddings'))
• evo2–1b-base: uses model discovery-model-evo-2: Uses model Evo 2 (Tables 5/6, row evo2–1b-base; Appendix A (APP1), C (APP3), D (APP4))
• evo2–7b-base: description: Description Linear probe, trained per mRNABench task, on frozen embeddings averaged over nucleotides from evo2-7b-base, an Evo 2 checkpoint trained as a base model on a mixture of genomic regions including mRNA; Evo 2 is a StripedHyena2 DNA model trained autoregressively on OpenGenome2, 9 trillion nucleotides across all domains of life. (expansion-p3-mrnabench-2025: Appendix A, Overview of Evaluated Models (family and variant rows; naive baselines paragraph); Section 4.3, Linear Probing ('training a linear classifier on frozen embeddings'; 'averaging over per-nucleotide embeddings'))
• evo2–7b-base: uses model catalog-model-evo-2: Uses model Evo 2 (Tables 5/6, row evo2–7b-base; Appendix A (APP1), C (APP3), D (APP4))
• evo2–7b-base: uses model discovery-model-evo-2: Uses model Evo 2 (Tables 5/6, row evo2–7b-base; Appendix A (APP1), C (APP3), D (APP4))
• evo2–7b: description: Description Linear probe, trained per mRNABench task, on frozen embeddings averaged over nucleotides from evo2-7b, an Evo 2 checkpoint with extended context, trained on long genomic sequences; Evo 2 is a StripedHyena2 DNA model trained autoregressively on OpenGenome2, 9 trillion nucleotides across all domains of life. (expansion-p3-mrnabench-2025: Appendix A, Overview of Evaluated Models (family and variant rows; naive baselines paragraph); Section 4.3, Linear Probing ('training a linear classifier on frozen embeddings'; 'averaging over per-nucleotide embeddings'))

Models, benchmarks and other records

Evidence

Source checking verifies the cited claim or transcription. It does not establish independent reproduction.