CHGGM-Aachen/genolator-v1-qa

Genolator V1 — Multimodal Gene Function QA

Question–answer pairs about human gene function, paired with precomputed embeddings of three modalities per gene: the coding DNA sequence, the amino acid sequence, and the predicted 3D protein structure. It is the dataset used to train and evaluate Genolator V1, a model that projects those embeddings into the token embedding space of a biomedical Llama-3 and answers questions about a gene without ever seeing its name or its raw sequence.

The defining property of this dataset is what the questions do not contain: no gene names, no gene symbols, no DNA and no amino acid sequences appear in any prompt. Every question refers only to "the given protein" or "the given sequence". A model can therefore only answer by reasoning over the multimodal embeddings supplied alongside the prompt, not by recalling what it already knows about a named gene.

At a glance

Split	Rows	Genes	File size
train	345,927	14,186	6.7 GiB
validation	36,521	1,773	0.71 GiB
test	35,387	1,773	0.68 GiB
total	417,835	17,732	8.1 GiB

The splits are gene-disjoint: no gene name occurs in more than one split. The gene-level division is exactly 80% / 10% / 10%.

Columns

Column	Type	Description
gene_name	string	HGNC gene symbol. Metadata only — never shown to the model.
prompt	string	The question. Contains no gene name, symbol or raw sequence.
response	string	The target answer based on GO slim ontology
kind	string	confirmation, denial or generic (see below).
go_aspect	string	molecular_function, cellular_component or biological_process.
dna_seq	string	Coding sequence (CDS) of the MANE Select transcript, ACGT only.
cdna_seq	string	Full transcript sequence including UTRs; contains dna_seq as a substring.
aa_seq	string	Amino acid sequence of the encoded protein.
cdna_seq_embedding	list<float>[4096]	Evo2 7B embedding of cdna_seq, mean-pooled.
aa_seq_embedding	list<float>[2560]	ESM-2 embedding of aa_seq, mean-pooled.
pst_embedding	list<float>[1280]	PST structure-infused embedding, mean-pooled.
group	int64	train only. Epoch rotation subset, 1–3 (see below).

All three embedding columns are complete and well-formed in all three splits: every row has an embedding of exactly the stated dimensionality, with no nulls and no non-finite values (checked over all 417,835 rows). The sequence columns are repeated across all QA rows of the same gene, which is what makes the files large relative to the row count.

Note that dna_seq is retained for reference but is not the sequence that was embedded: cdna_seq_embedding is computed from the full transcript in cdna_seq.

The three kinds of question

Nine sub-datasets were generated — three question kinds × three Gene Ontology aspects — and are combined here, distinguished by the kind and go_aspect columns.

confirmation — a question about a GO term that is annotated for the gene, so the correct answer is affirmative.

Q: Could the given protein participate in signal transduction initiated by ligand binding at a cell membrane receptor?

A: The amino acid sequence contains regions that are commonly associated with proteins mediating signal transmission from membrane-bound receptors, suggesting potential involvement in signal transduction pathways.

denial — the same construction over a GO term that is not among the gene's annotations, so the correct answer is negative. Together with confirmation this forms a balanced binary task.

Q: Does the given protein structure show repeated domains commonly involved in structural support outside the cell?

A: No data indicate that the given protein structure contains repeated domains typically associated with extracellular structural support.

generic — an open-ended question inviting a description of plausible functions, processes or localisations. The generating prompt deliberately excluded gene names, protein names, pathway names and nucleic acid terminology, so answers are phrased descriptively rather than by naming the entity.

Q: What types of cellular environments or associations may be possible for the given protein?

A: The given protein may participate in highly regulated assemblies that coordinate the cell division cycle, including complexes responsible for recognizing degradation signals on specific cell cycle proteins. […]

Composition

Split	confirmation	denial	generic	molecular_function	cellular_component	biological_process
train	115,113	115,113	115,701	106,586	124,712	114,629
validation	15,965	15,417	5,139	11,832	12,494	12,195
test	15,463	14,943	4,981	11,363	12,480	11,544

In validation and test there is one generic item per gene and GO aspect, i.e. roughly one third as many as confirmation or denial, which have three each. The training split is class-balanced instead — see group below for how that was achieved.

How the dataset was built

Genes and annotations. The gene set is MANE (Matched Annotation from NCBI and EMBL-EBI) Select release 1.4 on GRCh38, which provides one representative transcript per canonical protein-coding gene. For each gene the nucleotide sequence was retrieved with BioPython along with the corresponding amino acid sequence, and the gene was annotated with Gene Ontology terms across all three aspects.

Question generation. For each GO aspect, dedicated system prompts guided GPT-4.1 (hosted on Azure AI Foundry) using at least six manually crafted few-shot examples drawn from two unrelated genes. Three confirmation samples were generated per gene and aspect from the gene's own annotations; three denial samples per gene and aspect from GO terms absent from its ground truth annotations; and one generic, open-ended sample per gene and aspect. Throughout, gene names, gene symbols, raw DNA sequences and amino acid sequences were withheld from the generation process, so that no such identifier could leak into a question or answer.

Quality assessment. QA samples for a random selection of 56 genes were independently and blindly reviewed by two researchers, each sample annotated as accepted, declined or review. Generation was traced and annotated with a self-hosted Arize Phoenix instance. The full system prompts and few-shot examples are in the paper's supplementary table "Prompts".

DNA embeddings (4096-d). Evo2 7B, a genomic foundation model trained on 9.3 trillion DNA base pairs. Per-token embeddings were taken from layer blocks.28.mlp.l3, following the model authors' recommendation, and mean-pooled along the sequence dimension.

Amino acid embeddings (2560-d). ESM-2, accessed through the ESMFold 3B model on Hugging Face. Because of the model's cubic complexity, sequences longer than 1,000 residues were split into 1,000-residue segments and inferred separately. Per-token vectors from the last hidden layer of the ESM-2 block were concatenated and mean-pooled.

Structure embeddings (1280-d). Protein Structure Transformer (PST), which infuses structural knowledge into pretrained ESM-2 sequence embeddings. Predicted structures for the MANE transcripts were taken from the AlphaFold Protein Structure Database as PDB files, processed with PST to give a 1280-d embedding per residue, then mean-pooled.

Similarity-aware splitting. A purely gene-name-based split does not prevent leakage between homologous genes with near-identical sequences or structures. Instead, the three embeddings of each gene were concatenated into one multimodal representation, normalised and clustered with KMeans (k=3) under cosine similarity, so that structurally and functionally similar genes land in the same cluster. Clusters were mapped to the train, validation and test splits with sizes adjusted to an approximate 80/10/10 division, and assignments refined by distance to the cluster centroids, prioritising the samples most central to each cluster and keeping samples exclusive to one split.

The group column. For training efficiency the training split is divided into three equal-sized subsets, and each epoch draws on the subsets in rotation so that every training point is seen regularly. The subsets are class-balanced by construction: group 1 holds exactly one confirmation, one denial and one generic item for each of its 38,371 gene–aspect pairs. Since only one generic item exists per gene and aspect — against three each of the other kinds — the three confirmation and denial items are distributed one per group, while each generic item is repeated in every group. That is why the training split reports 115,701 generic rows built from 38,571 distinct items.

Two minor irregularities are worth knowing about if you use group as an epoch subset: 200 generic items are absent from group 1 and appear only in groups 2 and 3, and groups 2 and 3 contain a few within-group duplicate generic rows (176 and 12 respectively, up to 5 copies of one item). Group 1 is exactly balanced and duplicate-free.

Loading

from datasets import load_dataset

ds = load_dataset("CHGGM-Aachen/genolator-v1-qa")           # all three splits
test = load_dataset("CHGGM-Aachen/genolator-v1-qa", split="test")

The training split is 6.7 GiB, so stream it if you do not want it on disk:

ds = load_dataset("CHGGM-Aachen/genolator-v1-qa", split="train", streaming=True)
row = next(iter(ds))
print(row["prompt"], len(row["cdna_seq_embedding"]))

To read only the columns you need — e.g. the DNA and structure embeddings for the Evo2+PST model, skipping the bulky raw sequences — go through the parquet files directly:

import pyarrow.parquet as pq

cols = ["prompt", "response", "kind", "cdna_seq_embedding", "pst_embedding"]
for batch in pq.ParquetFile("data/test.parquet").iter_batches(batch_size=256, columns=cols):
    ...

train.parquet is written in row groups of 20,000 rows (18 groups, at most 0.45 GiB of decompressed data each), so batched reads stay bounded. validation.parquet and test.parquet are each a single row group, so reading them materialises the whole split regardless of batch size — 0.81 GiB and 0.77 GiB of decompressed data respectively.

Intended use and limitations

This dataset is built for one specific setting: teaching and evaluating a language model that receives gene information only as projected embeddings. The prompts are deliberately stripped of identifiers, which makes the dataset unsuitable for tasks where the model is supposed to know which gene it is talking about.

• The answers are model-generated. GPT-4.1 wrote every question and answer from GO annotations. They were spot-checked by two blind reviewers over 56 genes, not exhaustively curated, and they inherit both the incompleteness of GO annotations and the phrasing habits of the generating model.
• denial items rest on absence of evidence. A GO term missing from a gene's annotations does not mean the association is biologically false, only that it is not annotated. Negative answers should be read as "not annotated", not "does not occur".
• Coverage is limited to MANE Select, one canonical transcript per protein-coding gene. Isoforms, non-coding genes and non-canonical transcripts are absent.
• Structures are predicted, not experimental: PST embeddings derive from AlphaFold models and carry that method's uncertainty.
• generic answers are long and descriptive. Evaluating them needs a semantic measure; exact-match scoring is meaningless for this kind.

Provenance and licensing

Derived from: MANE Select 1.4 (NCBI/EMBL-EBI), Gene Ontology annotations, the AlphaFold Protein Structure Database, and question–answer text generated with GPT-4.1. Embeddings were computed with Evo2 7B, ESM-2 and PST.

Citation

T.b.d

Code

The training and inference code for Genolator V1, including the column names this dataset is wired into, is published alongside the paper.

Contact

Martin Danner — dataset creator and person responsible for this dataset.

• Center for Human Genetics and Genomic Medicine — mdanner@ukaachen.de
• scieneers GmbH — martin.danner@scieneers.de

Questions about the data, corrections and reports of problematic content are welcome at either address.