ThernaBench-1.0

Predict how strongly a UTR drives expression in a given cell line. Each item is a UTR (a 50-nt 5′ UTR or a 240-nt 3′ UTR) measured by a massively parallel reporter assay (MPRA) in one cancer cell line. The answer is one of five expression bins: very low, low, medium, high, very high.

Files

file	content
data/therna_bench.parquet	the benchmark, one row per item

Usage

1. Choose a question column, question_{minimal|full}_{text|embed}.
   • minimal gives the UTR and the cell line; full adds the RBP and miRNA context.
   • text carries the UTR as nucleotides in the prompt, for text-only LLMs. embed leaves it out, for BioReason-RNA, which reads the UTR as Evo2 embeddings of utr.
2. Send the question as a single user message.
3. Read the answer from the last \boxed{}. It must contain one of the five bins, ignoring LaTeX formatting such as \boxed{\text{low}} or \boxed{very\ low}; anything else counts as wrong.
4. Score against label. Report exact accuracy, with 3-way and within-one-bin accuracy alongside. 3-way merges the bins into low (very low, low), medium, and high (high, very high). Report the minimal and full settings separately.

Questions

All four use the same prompt: a role line, a guide to reading the input, the ## Input block, and a closing instruction to put the prediction, one of the five bins, in \boxed{}. Only the ## Input block differs:

UTR:                                          <- text variants
<|RNA_START|>CCCGGCCCCTGGTCCC...<|RNA_END|>   <- text variants
UTR type: 5_utr

Cell type: MDA-MB-231
MDA-MB-231 is a triple-negative claudin-low breast adenocarcinoma line ...

RBP context:                                  <- full variants
  ...
miRNA context:                                <- full variants
  ...

Columns

column	description
id	item id, therna-0001 …
source	source test set: internal (14 cell lines) or external (5 cell lines)
source_idx	row index in the source test set
sequence_id	UTR identifier in the source library
utr_type	5_utr or 3_utr
utr	UTR sequence (DNA alphabet)
utr_length	50 or 240
cell_type	cell line
cell_description	cell-line description, as shown in the question
n_rbps, rbp_context	RBPs predicted to bind the UTR, one struct each, in question order
n_mirnas, mirna_context	miRNAs with seed matches in the UTR, one struct each, in question order
question_minimal_text	UTR as text, UTR type and cell line
question_minimal_embed	question_minimal_text without the UTR text
question_full_text	question_minimal_text plus the RBP and miRNA context
question_full_embed	question_full_text without the UTR text
expression	measured MPRA expression, the assay's center of mass; higher means higher expression
label	expression bin; edges are quintiles per UTR type and cell line of the internal library

The rbp_context and mirna_context structs hold every value the question shows. Every *_bin field is a quintile bin. Null means the question shows N/A or omits the line.

rbp_context field	description
name, protein_name, uniprot_id	RBP gene symbol, protein name, UniProt accession
binding_prob, binding_prob_bin	predicted probability that the RBP binds this UTR
spearman_corr	Spearman correlation between the RBP's protein expression and this UTR's MPRA expression across cell lines
rna_expr, rna_expr_bin	RBP RNA expression in this cell line, log2(TPM+1)
prot_expr, prot_expr_bin	RBP protein expression in this cell line, normalized
functions	functional roles
protein_function	protein function summary
mirna_crosstalk	roles in miRNA-mediated regulation
protein_domains	protein domains
subcellular_location	subcellular location
interacts_with	other RBPs in this item's panel that it interacts with

mirna_context field	description
name	miRNA name
description	canonical targets or cluster
best_match_type	strongest seed match: 8mer, 7mer-m8, 7mer-A1 or 6mer; the scan labels most canonical 8mers 7mer-m8
match_score, match_score_bin	cumulative seed match score from the seed scan, which counts nested matches: one site scores about 1.0 as a 6mer, 1.2 as a 7mer-A1, 1.8 as a 7mer-m8 and 2.8–3.0 as an 8mer, not the per-site weights the questions describe
spearman_corr	Spearman correlation between the miRNA's expression and this UTR's MPRA expression across cell lines
expr_level, expr_level_bin	miRNA expression in this cell line (CCLE, log2)

Contents

rows	500: 250 internal, 250 external, 424 distinct UTRs
UTR type	252 5′ / 248 3′
label	very low 121 · low 88 · medium 92 · high 78 · very high 121
cell lines	14

Results

Full setting, one generation per item, scored as in Usage.

model	items	5-way	3-way	within ±1
BioReason-RNA, RL	500	0.430	0.622	0.790
BioReason-RNA, SFT	500	0.362	0.562	0.740
GPT-6 Astra	499	0.279	0.517	0.665
GLM-5.3	500	0.232	0.480	0.638
Kimi K3	500	0.212	0.480	0.658
Claude Opus 5.5	440	0.207	0.445	0.668
Random	500	0.200	0.363	0.503
Nemotron-3.5-Lightning-30B, untrained	500	0.178	0.412	0.598

• GPT-6 Astra, GLM-5.3, Kimi K3 and Claude Opus 5.5 read question_full_text at maximum reasoning effort through the NVIDIA Inference Hub, with up to 128k output tokens. Opus is scored on 440 items, because the provider's content filter blocked the other 60 on every attempt. GPT-6 Astra is scored on 499, because one item timed out on every attempt. At high effort, GPT-6 Astra (0.264) and Opus (0.207) score within noise of max.
• BioReason-RNA reads the served question, which is question_full_embed without the five bins named in its closing sentence, with greedy decoding.
• Nemotron-3.5 reads the served question with the UTR as text and a one-shot example of the answer format, with greedy decoding under a 16,384-token budget. 51 of its answers have no valid boxed bin and count as wrong.
• Random is the expected score of a uniform guess over the five bins.