SoftGene: Protein Language Model-Enhanced Soft Prompting for Interpretable Gene Set Annotation

Drew Ross; Arya Hadizadeh Moghaddam; Dongjie Wang; Xiaoyu Zhang; Zijun Yao

cs.AIOct 2, 2026

SoftGene: Protein Language Model-Enhanced Soft Prompting for Interpretable Gene Set Annotation

Authors: Drew Ross, Arya Hadizadeh Moghaddam, Dongjie Wang, Xiaoyu Zhang, Zijun Yao

Organizations: Electrical Engineering and Computer Science, University of Kansas, USA · Cell Biology and Physiology, University of Kansas Medical Center, USA

Abstract

Gene set analysis is a cornerstone of functional genomics, yet it remains labor-intensive and heavily dependent on manual curation and expert biological interpretation. While Large Language Models (LLMs) have emerged as powerful tools for genomic reasoning and annotation, most existing approaches rely on symbolic gene names and fail to capture domain-specific biological structure, particularly protein sequence information that governs molecular activity, interactions, and downstream gene function. In this work, we propose SoftGene, a novel framework for LLM-based gene set annotation that leverages the hierarchical structure of gene sets. First, we use a hierarchical attention-based encoder built on ESM, a protein language model, to represent each gene set using protein-level amino acid sequence information. Second, we construct a hybrid prompting scheme that combines soft prompts derived from gene set embeddings with hard prompts containing auxiliary context generated by an LLM, and feed the resulting prompt into a local LLM for annotation. We evaluate our framework on two benchmark datasets: Gene Ontology (GO) and the Molecular Signatures Database (MSigDB). Our results show that integrating protein-sequence representations with textual context improves gene set annotation overall, while per-domain analyses reveal that the contribution of protein embeddings varies across biological domains.

Figures & tables

Figure 1: High-level overview of the model architecture. The proposed approach combines a sequence-level pathway that encodes amino acid information into soft prompt tokens with a textual pathway that provides high-level functional context from a large language model, enabling the generation of descriptive gene set annotations optimized via supervision against ground-truth labels.

Figure 2: Overview of the hierarchical gene set encoder. Frozen pretrained protein representations are hierarchically aggregated using trainable attention-based pooling to produce a gene set embedding.

Figure 3: To obtain the gene set context, we utilize a system prompt to define a persona and stylistic guidelines alongside a user prompt for task instructions, which are fed to a commercial model.

Figure 4: Overview of the hybrid gene set soft prompting module. A gene set embedding is converted into soft prompt tokens that, together with instruction and context prompts, condition a frozen language model to generate free-text gene set annotations.

Measure	GO	MSigDB
Number of Gene Sets	18,889	2,461
Avg. Genes Per Set	21.18	50.07
Min. Number of Genes	1	5
Max. Number of Genes	13,616	1,491
Unique Genes Covered	42,597	12,021
Avg. Annotation Length	38.51	36.25

Table 1: Dataset statistics (annotation length measured in characters).

Prompt Engineering Methods
Configuration	Gene Ontology (GO)			MSigDB
Configuration	METEOR	ROUGE-L	BERTScore-F1	METEOR	ROUGE-L	BERTScore-F1
Zero-shot	0.024 ± 0.001	0.133 ± 0.001	0.773 ± 0.00	0.087 ± 0.014	0.127 ± 0.009	0.762 ± 0.002
Few-shot	0.018 ± 0.001	0.164 ± 0.003	0.740 ± 0.001	0.100 ± 0.025	0.136 ± 0.017	0.771 ± 0.008
One-shot CoT	0.065 ± 0.001	0.096 ± 0.002	0.761 ± 0.001	0.086 ± 0.010	0.112 ± 0.003	0.762 ± 0.001
Prompt-tuning Methods

Table 2: Performance of SoftGene and baselines on the Gene Ontology (GO) and MSigDB datasets. First, second, and third-best results are indicated by boldface , double underline , and single underline , respectively.

Configuration	Gene Ontology (GO)			MSigDB
Configuration	METEOR	ROUGE-L	BERTScore	METEOR	ROUGE-L	BERTScore
w/o Soft Prompt ( $S$ )	0.020 ± 0.000	0.065 ± 0.001	0.707 ± 0.001	0.006 ± 0.004	0.045 ± 0.002	0.704 ± 0.001
w/o Context ( $H_{cont}$ )	0.182 ± 0.024	0.232 ± 0.028	0.795 ± 0.005	0.145 ± 0.007	0.202 ± 0.020	0.790 ± 0.003
w/o Protein Embeddings	0.199 ± 0.015	0.252 ± 0.017	0.803 ± 0.002	0.097 ± 0.017	0.135 ± 0.025	0.774 ± 0.012
SoftGene	0.220 ± 0.033	0.267 ± 0.029	0.806 ± 0.006	0.157 ± 0.005	0.215 ± 0.001	0.795 ± 0.003

Table 3: Ablation study.

Figure 5: Performance comparison of the SoftGene model across different soft prompt lengths using the GO dataset.

Figure 6: Performance comparison of the SoftGene model across different soft prompt lengths using the MSigDB dataset.

	Gene Ontology (GO)			MSigDB
Split	METEOR	ROUGE-L	BERTScore	METEOR	ROUGE-L	BERTScore
Gene-Disjoint	0.205 ± 0.014	0.251 ± 0.015	0.801 ± 0.009	0.136 ± 0.082	0.193 ± 0.062	0.791 ± 0.034
Random	0.220 ± 0.033	0.267 ± 0.029	0.806 ± 0.006	0.157 ± 0.005	0.215 ± 0.001	0.795 ± 0.003

Table 4: Performance comparison between random and gene-disjoint train-test splits.

	PLM			Randomized			$\Delta$
Domain	METEOR	ROUGE-L	BERTScore	METEOR	ROUGE-L	BERTScore	METEOR	ROUGE-L	BERTScore
BP	0.223	0.264	0.802	0.222	0.256	0.798	-0.001	-0.008	-0.004
CC	0.164	0.212	0.784	0.102	0.151	0.769	-0.062	-0.061	-0.015
MF	0.236	0.314	0.825	0.213	0.285	0.820	-0.023	-0.029	-0.005

Table 5: Per-domain ablation of PLM embeddings on the GO dataset. Comparison between real PLM embeddings and randomized embeddings, with differences.

Figure 7: Attention visualization for GO:0045986. The tree depicts the three-level hierarchy of the gene set encoder, with the top five protein isoforms (by attention weight) shown beneath each gene. Node size reflects learned attention weights, indicating the relative contribution of each gene and protein isoform.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8: Performance comparison of SoftGene across different commercial models using the GO dataset.

Figure 9: Performance comparison of SoftGene across different commercial models using the MSigDB dataset.

Figure 10: Performance comparison of SoftGene across different protein language models using the GO dataset.

Figure 11: Performance comparison of SoftGene across different protein language models using the MSigDB dataset.

Figure 12: BERTScore F1 performance across gene set characteristics in the GO dataset. Box plots display the distribution of BERTScore F1 scores stratified by four gene set properties: (A) description length in words, (B) GO domain (biological process, cellular component, molecular function), (C) mean gene frequency across all gene sets (binned into quintiles Q1–Q5, where Q1 represents the most unique genes and Q5 the least unique), and (D) number of genes per gene set. Sample sizes (n) are indicated at the bottom of each panel.

Figure 13: BERTScore F1 performance across gene set characteristics in the MSigDB dataset. Box plots display the distribution of BERTScore F1 scores stratified by four gene set properties: (A) description length in words, (B) MSigDB collection (BIOCARTA, KEGG LEGACY, PID, REACTOME), (C) mean gene frequency across all gene sets (binned into quintiles Q1–Q5, where Q1 represents the most unique genes and Q5 the least unique), and (D) number of genes per gene set. Sample sizes (n) are indicated at the bottom of each panel.

Figure 14: BERTScore F1 performance stratified by GO ontology depth. Box plots display the distribution of BERTScore F1 scores across gene sets binned by their depth within the Gene Ontology hierarchy (0–3, 4–5, 6–7, 8–9, and 10+).

Prompt Engineering Methods
Configuration	Gene Ontology (GO)		MSigDB
Configuration	Semantic Similarity	Hierarchical Semantic Similarity	Semantic Similarity
Zero-shot	0.861 ± 0.002	0.871 ± 0.002	0.876 ± 0.001
Few-shot	0.860 ± 0.001	0.869 ± 0.001	0.872 ± 0.002
One-shot CoT	0.863 ± 0.010	0.885 ± 0.000	0.876 ± 0.003
Prompt-tuning Methods

Appendix

Table 6: Performance of SoftGene and baselines on the Gene Ontology (GO) and MSigDB datasets using additional performance metrics. Best results are indicated by boldface .

Group	Gene Set	Score	Reference	Predicted
Top 5	GO:0004143	0.9500	ATP-dependent diacylglycerol kinase activity	diacylglycerol kinase activity
	GO:0045588	0.9500	positive regulation of gamma-delta T cell differentiation	positive regulation of T cell differentiation
	GO:0042129	0.9499	regulation of T cell proliferation	positive regulation of T cell activation
	GO:0140950	0.9496	histone H2A deubiquitinase activity	histone H2A deubiquitination
	GO:0006563	0.9493	L-serine metabolic process	serine metabolic process
Bottom 5	GO:0030431	0.5774	sleep	positive regulation of interleukin-18 production

Appendix

Table 7: Top 5 and bottom 5 scoring gene set predictions from GO by BERTScore F1 (excluding perfect scores).

Group	Gene Set	Score	Reference	Predicted
Top 5	REACTOME_SARS_COV_INFECTIONS	0.9477	SARS-CoV Infections	SARS-CoV-2 infection
	REACTOME_DOWNSTREAM_ SIGNALING_OF_ACTIVATED_ FGFR4	0.9458	Downstream signaling of activated FGFR4	Signaling by FGFR4
	REACTOME_DOWNSTREAM_ SIGNALING_OF_ACTIVATED_ FGFR2	0.9454	Downstream signaling of activated FGFR2	Signaling by FGFR2
	BIOCARTA_MTOR_PATHWAY	0.9453	mTOR Signaling Pathway	mTOR signaling
	REACTOME_METHIONINE_SALVAGE_ PATHWAY	0.9450	Methionine salvage pathway	Methionine salvage
Bottom 5	REACTOME_S_PHASE	0.6173	S Phase	Regulation of APC/C activators in early mitosis

Appendix

Table 8: Top 5 and bottom 5 scoring gene set predictions from MSigDB by BERTScore F1 (excluding perfect scores).

Metric	Value
Total Time	6,575.92s (1.83 hr)
Average Time per Protein	26.80 ms
Throughput	37.3 proteins/s
Peak GPU Memory	2.250 GB

Appendix

Table 9: Protein embedding inference runtime using ESM-2 on a single NVIDIA A6000 GPU.

Metric	Value
Total Training Time	1,709.11s (0.47 hr)
Peak GPU Memory	46.39 GB

Appendix

Table 10: SoftGene training runtime on the GO dataset using two NVIDIA A6000 GPUs.

Metric	SoftGene	Zero-Shot
Total Inference Time	584.72s	478.91s
Average Time per Sample	103.18 ms	84.51 ms
Throughput	9.69 samples/s	11.83 samples/s
Peak GPU Memory	19.84 GB	18.87 GB

Appendix

Table 11: Inference runtime comparison between SoftGene and the zero-shot baseline on the GO dataset using a single NVIDIA A6000 GPU.

Metric	Value
Total Time	2.43 s
Average Time per Sample	467.45 ms
Throughput	2.14 samples/s

Appendix

Table 12: Inference runtime of SoftGene on the five largest gene sets in the GO evaluation set.

Dataset	Ground Truth Overlap
GO	1.37%
MSigDB	3.41%

Appendix

Table 13: Lexical overlap between GPT-generated auxiliary context and ground-truth annotations.

Hierarchical Attention Encoder
Parameter	GO	MSigDB
Learning Rate $\alpha$	$1\times 10^{-3}$	$1\times 10^{-3}$
Protein Pooling Hidden Size	512	512
Gene Pooling Hidden Size	512	512
Weight Decay	$1\times 10^{-4}$	$1\times 10^{-4}$
Gradient Clipping Norm	1	1

Appendix

Table 14: Hyperparameter settings.

LLM-Assisted Annotation

Large language model-assisted annotation uses AI to label, classify, and structure data, often with human review to improve scale and consistency.

See all 201 papers

QLoRA Fine-Tuning of Ministral LLM for Sequence-to-Function Protein Annotation
Sep 21, 2026Demian Pavlyshenko, Bohdan PavlyshenkoLLM Fine-Tuning LLM-Assisted Annotation
Homo-RAG: Homology-Guided Retrieval-Augmented Generation for Cross-Species Gene Function Prediction
Aug 26, 2026Azrin SultanaRetrieval-Augmented Generation LLM-Assisted Annotation
Department of Computer Science, American International University-Bangladesh, 408/1, Kuratoli, Khilkhet, 1229, Dhaka, Bangladesh.
LLMBDC: Language Model for Biological Domains Oriented Clustering of Gene Ontology
Jul 30, 2026Ximing Ran, Jie Xu, Peng Jin +3Clustering Genomics
Emory University Atlanta, GA, USA

Sep 21, 2026J/K move · Enter open · S save

QLoRA Fine-Tuning of Ministral LLM for Sequence-to-Function Protein Annotation

Demian Pavlyshenko, Bohdan Pavlyshenko

LLM Fine-Tuning LLM-Assisted Annotation Protein Function Prediction

Open paper AskPDF arXiv

Functional annotation of newly sequenced proteins remains a bottleneck in molecular biology: the number of sequences in public repositories grows far faster than the capacity for manual curation. Most computational approaches consider annotation as multi-label classification over a fixed ontology, which constrains predictions to a predefined label set. In this work we study the the protein annotation as a sequence-to-text generation problem. We fine-tune the 3B-parameter Ministral 3 base model with QLoRA (4-bit NF4 quantization with low-rank adapters) on sequence annotation pairs. We assess predictions with an LLM-as-expert protocol: a GPT model prompted as a senior molecular-biology curator scores organism identification as binary and function annotation quality. We conclude that QLoRA-fine-tuned compact LLMs can generate curator-style annotations with genuine biological value for a substantial subset of proteins. We also discuss future directions in data quality, model scaling, and evidence grounding that are needed to make the approach sufficiently reliable for practical use.

Abstract

LLM-Assisted Annotation

Explore similar work

QLoRA Fine-Tuning of Ministral LLM for Sequence-to-Function Protein Annotation

Homo-RAG: Homology-Guided Retrieval-Augmented Generation for Cross-Species Gene Function Prediction

LLMBDC: Language Model for Biological Domains Oriented Clustering of Gene Ontology