@inproceedings{xu-etal-2026-protsyngen,
title = "{P}rot{S}yn{G}en: Knowledge Graph-Guided Open-Ended Data Synthesis for Protein {LLM}s",
author = "Xu, Yifan and
Lv, Ying and
Xu, Gang and
Ma, Jianpeng",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.826/",
pages = "17498--17524",
ISBN = "979-8-89176-462-0",
abstract = "The performance of large language models (LLMs) depends critically on the quality of their instruction-tuning data. For protein LLMs, existing instruction-tuning datasets are typically derived from structured protein annotations through template-based or label-guided generation. Although these datasets capture important biological facts, their construction paradigms often lead to two limitations: instruction homogenization and reasoning superficiality. The former refers to questions concentrating around repetitive forms and narrow topics; the latter to answers relying on shallow annotation lookup rather than evidence-rich biological reasoning. To address these limitations, we propose ProtSynGen, a knowledge graph-guided open-ended data synthesis framework for protein LLMs. ProtSynGen integrates structured protein databases with unstructured scientific literature into a protein-centric knowledge graph, and uses hierarchical graph sampling to guide LLMs in generating open-ended QAs across atomic, aggregated, and complex reasoning scenarios. Experiments show that ProtSynGen produces instruction-tuning data with broader topic coverage, richer reasoning depth, and stronger quality metrics than existing protein LLM instruction datasets. Models trained on ProtSynGen data further achieve superior performance on open-ended protein QA benchmarks, consistently outperforming baseline methods across atomic, aggregated, and complex reasoning scenarios."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="xu-etal-2026-protsyngen">
<titleInfo>
<title>ProtSynGen: Knowledge Graph-Guided Open-Ended Data Synthesis for Protein LLMs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yifan</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ying</namePart>
<namePart type="family">Lv</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Gang</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jianpeng</namePart>
<namePart type="family">Ma</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>The performance of large language models (LLMs) depends critically on the quality of their instruction-tuning data. For protein LLMs, existing instruction-tuning datasets are typically derived from structured protein annotations through template-based or label-guided generation. Although these datasets capture important biological facts, their construction paradigms often lead to two limitations: instruction homogenization and reasoning superficiality. The former refers to questions concentrating around repetitive forms and narrow topics; the latter to answers relying on shallow annotation lookup rather than evidence-rich biological reasoning. To address these limitations, we propose ProtSynGen, a knowledge graph-guided open-ended data synthesis framework for protein LLMs. ProtSynGen integrates structured protein databases with unstructured scientific literature into a protein-centric knowledge graph, and uses hierarchical graph sampling to guide LLMs in generating open-ended QAs across atomic, aggregated, and complex reasoning scenarios. Experiments show that ProtSynGen produces instruction-tuning data with broader topic coverage, richer reasoning depth, and stronger quality metrics than existing protein LLM instruction datasets. Models trained on ProtSynGen data further achieve superior performance on open-ended protein QA benchmarks, consistently outperforming baseline methods across atomic, aggregated, and complex reasoning scenarios.</abstract>
<identifier type="citekey">xu-etal-2026-protsyngen</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.826/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>17498</start>
<end>17524</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T ProtSynGen: Knowledge Graph-Guided Open-Ended Data Synthesis for Protein LLMs
%A Xu, Yifan
%A Lv, Ying
%A Xu, Gang
%A Ma, Jianpeng
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F xu-etal-2026-protsyngen
%X The performance of large language models (LLMs) depends critically on the quality of their instruction-tuning data. For protein LLMs, existing instruction-tuning datasets are typically derived from structured protein annotations through template-based or label-guided generation. Although these datasets capture important biological facts, their construction paradigms often lead to two limitations: instruction homogenization and reasoning superficiality. The former refers to questions concentrating around repetitive forms and narrow topics; the latter to answers relying on shallow annotation lookup rather than evidence-rich biological reasoning. To address these limitations, we propose ProtSynGen, a knowledge graph-guided open-ended data synthesis framework for protein LLMs. ProtSynGen integrates structured protein databases with unstructured scientific literature into a protein-centric knowledge graph, and uses hierarchical graph sampling to guide LLMs in generating open-ended QAs across atomic, aggregated, and complex reasoning scenarios. Experiments show that ProtSynGen produces instruction-tuning data with broader topic coverage, richer reasoning depth, and stronger quality metrics than existing protein LLM instruction datasets. Models trained on ProtSynGen data further achieve superior performance on open-ended protein QA benchmarks, consistently outperforming baseline methods across atomic, aggregated, and complex reasoning scenarios.
%U https://aclanthology.org/2026.findings-emnlp.826/
%P 17498-17524
Markdown (Informal)
[ProtSynGen: Knowledge Graph-Guided Open-Ended Data Synthesis for Protein LLMs](https://aclanthology.org/2026.findings-emnlp.826/) (Xu et al., Findings 2026)
ACL