Agentic Literature Extraction Augments Targeted Protein Degradation Databases
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: I'm Ines.
Marcus: And I'm Marcus, here with Yuki.
Ines: We're looking at the paper: "Agentic Literature Extraction Augments Targeted Protein Degradation Databases".
Marcus: Here's the gist of the paper. Predictive models in biomedicine depend on structured assay data locked in the text, tables, and supplements of primary publications This bottleneck is especially acute in targeted protein degradation (TPD), where each assay record must combine compound identity, degradation target, recruiter, assay context, and endpoint values reported across sections, tables, and supplementary files.
Ines: The workflow achieves record-level F1 “ 0.98 and transferred to PROTACs by terminology substitution alone, maintaining record-level F1 ą 0.93.
Marcus: The Problem Addressed The bottleneck in data-driven drug discovery is the difficulty of turning published results into structured, machine-readable data.
Ines: This problem is particularly evident in targeted protein degradation (TPD), where each assay record must combine compound identity, degradation target, recruiter, assay context, and endpoint values reported across sections, tables, and supplementary files.
Title and authors: Ines: So we're talking about this paper titled "Agentic Literature Extraction Augments Targeted Protein Degradation Databases". It seems like they are tackling the huge problem of turning all that scattered, unstructured text and tables from scientific journals into something actual computer systems can use.
Marcus: Yeah, it focuses on targeted protein degradation, or TPD. The authors are showing how they built a workflow to extract data from those publications—like compound identity and the target they're trying to degrade—and structure it properly across all the different parts of a paper, not just one spot in it.
Yuki: From my side, I think this is important because TPD is really growing right now. With new modalities like PROTACs and molecular glues popping up, we need these databases to keep up with the literature. If we don't structure this data, the models can't learn anything useful about how these drugs work in a biological context.
Ines: Exactly. The paper describes an expert-in-the-loop workflow that uses multiple agents—an extractor, an analyst, and a refiner—plus this thing called CAPO to constantly improve the instructions they give the AI for extraction. It’s basically an automated way to go beyond just scraping text and actually pull out all those necessary details.
Marcus: And what they found is pretty solid. They achieved a record-level F1 score of zero point nine eight under both experimental and mechanistic modes when extracting molecular glue data, which is a big jump compared to the existing databases they’re augmenting <ref:2605.11221#pg1>.
Yuki: The fact that it recovered kinetic information and assay context—things like timepoints and cell lines—is what really stands out for me. That context-aware modeling is essential because degradation activity isn't just a property of the compound itself; it depends entirely on the environment where you test it.
Ines: Right, because if you don't know the assay context, you can't accurately compare results between different studies or even different experimental conditions. The method uses a triangular comparison framework to check their predictions against baseline data and expert annotations to make sure they aren't just guessing.
Title and authors: Marcus: And they showed that this approach significantly expanded the molecular glue database by eighty-one percent and the PROTAC database by ninety-two percent, with those record-level F1 scores of zero point nine three for PROTACs as well <ref:2605.11221#pg1>. That shows a real practical impact on data curation.
Yuki: It's interesting to see how this moves away from needing massive amounts of manual transcription toward AI assisting the experts. The implication is that we can keep these structured resources more aligned with how fast papers are being published, which is crucial for drug discovery pipelines.
Ines: So, what they’re suggesting here is a shift toward AI-assisted curation where domain experts provide targeted annotations and then validate the pipeline outputs instead of doing all the tedious transcription themselves. It’s about using AI to handle the heavy lifting of combining information from text, tables, and supplementary files.
Marcus: The limitation they mention is that their current workflow is restricted to text and table modalities because many of the baseline records they used were supported by figures. They acknowledge that vision-language extraction from plots or blots could be a more direct way to improve coverage later on.
Yuki: That makes sense, because getting data out of a complex figure requires a completely different kind of AI capability, which is still an open area for improvement in this space. It also sounds like the workflow isn't meant to replace experts yet; it’s designed to support them by handling the record-level structure.
Ines: That’s the core idea, right? The system handles the complexity of combining disparate data sources, and those recovered fields like concentrations and assay types are what make a real difference for downstream modeling. It shows that even with messy literature, we can build structured resources.
Marcus: And they found that this method works well under different evaluation modes, distinguishing between an experimental mode using five elements and a mechanistic mode using three elements for the key. This lets them measure field-level accuracy very precisely, seeing things like precision of zero point nine one nine with a recall of zero point nine three three under the mechanistic mode for molecular glue records <ref:2605.11221#pg1>.
Title and authors: Yuki: It’s good that they are being specific about those modes because it shows they aren't just getting a general score; they are checking the quality at different levels of biological detail, which is what we need to trust when we use this data.
Ines: So, to wrap up on the "Agentic Literature Extraction Augments Targeted Protein Degradation Databases" paper, it shows that sophisticated AI workflows can effectively handle the multi-source complexity inherent in TPD literature by combining structured extraction with constant prompt refinement. The result is a significant boost in database size and the recovery of crucial contextual information like timepoints and assay types, which are vital for building accurate predictive models.
Marcus: It moves us closer to having resources that can keep pace with the rapid growth of therapeutic modalities like PROTACs, by making the process of turning published results into structured data much more scalable through this agentic approach.
Yuki: For us in population genetics, seeing how these databases are built and enriched is key because it helps us understand what kind of biological questions we can actually ask when we have high-quality, context-aware TPD data available.
Ines: It’s a practical way to keep the structured resources aligned with the pace of publication without requiring exhaustive manual transcription from every single paper. We’ll keep an eye on how they plan to integrate vision extraction next.
Marcus: Exactly. The work released, including the prompts and evaluation code, is a resource for anyone looking to improve TPD data curation using AI assistance more broadly.
Yuki: That sounds like a solid step forward in making complex biomedical literature useful for the next generation of computational biology tools we develop.
Ines: Right. That’s what this paper does by improving how we capture the necessary context from messy publications on TPD, and it opens up a new way to build better predictive models for drug discovery.
The paper's summary: Ines: So, this paper is about taking all those messy research papers and turning them into structured data that computers can actually use for drug discovery.
Marcus: Right, they’ve developed this workflow where AI agents work together to pull out information from those papers—like the compound name and what it targets—and organize it across the whole document, not just one spot.
Ines: It’s about combining reading the text with looking at tables and supplementary files to get a complete picture of an experiment.
Marcus: They used these agents, an extractor, an analyst, and a refiner, plus this prompt module called CAPO that lets the system improve its instructions based on performance metrics.
Ines: The big result they showed is that they got a record-level F1 score of point nine eight when extracting molecular glue data under both experimental and mechanistic modes.
Marcus: That’s solid, because it means these augmented databases are much more accurate than what we have now, like boosting the PROTAC database by ninety-two percent.
Ines: The key thing they recovered was context—things like timepoints and assay types—which is essential because degradation activity isn't just a property of the compound itself.
Marcus: That’s what I care about from a cohort perspective; having that context means we can compare results across different studies, which helps us see if the effect is consistent or if it depends on the specific experimental conditions.
Ines: The methodology they use for checking their work involves a triangular comparison against existing databases and expert annotations to validate the AI’s predictions.
Marcus: It’s a way to quantify exactly how much new data this system finds compared to what was already there, so you know if it’s actually adding value or just repeating old information.
Ines: They found that this approach helps expand molecular glue and PROTAC databases by eighty-one percent and ninety-two percent respectively.
Marcus: And the implication for the field is a shift away from needing massive manual transcription toward AI helping experts provide targeted annotations, which keeps pace with how fast new papers are coming out.
Ines: The authors pointed out that these recovered fields like concentrations and cell lines are critical because degradation activity is not context-free; it really depends on where you test it.
Marcus: They also mentioned a limitation, though—the current setup only works well with text and tables because they can't easily extract information from figures, like blots or plots.
Ines: They suggest that vision-language extraction from those images is the next big step for improving coverage of the data.
Marcus: It’s a practical route to keeping these structured resources updated without demanding exhaustive manual transcription from every single paper in circulation.
Ines: So, this work shows how AI can handle the multi-source complexity of literature by combining extraction and prompt refinement to recover crucial context for modeling.
The paper's improvements: Ines: So, we’re talking about how these authors are suggesting ways to make this extraction workflow even better for future use.
Marcus: They aren't just stopping at getting a good F1 score; they’re looking at how to handle the limitations of the current text and table-only data collection.
Ines: They’re focusing on expanding the input modalities, meaning moving beyond just reading text and tables into analyzing figures, like blots or plots.
Marcus: That makes sense because if you can pull information out of a figure directly, you’re bypassing the bottleneck where they currently have to rely on vision-language extraction.
Ines: They see vision-language extraction as the most direct path to getting more data from those complex experimental images that are currently just missing.
Marcus: I agree, because if you can get structured data from a figure, you’re covering a huge chunk of the experimental results that manual transcription would never catch.
Ines: They also talk about refining the prompt system further, using CAPO to adapt the extraction instructions based on real performance metrics.
Marcus: That means the AI isn't just following a static set of rules; it’s learning from its mistakes and improving how it asks the question for that next paper.
Ines: And they are also looking at how to make this work even more reliably across different types of degradation assays, moving past just molecular glues.
Marcus: That’s important because if you can transfer the extraction logic across different assay types, you make the database much more useful for researchers working on diverse targets.
Ines: They also emphasize that this system isn't meant to replace expert curation entirely; it’s meant to augment it, allowing domain experts to focus their effort on validating the AI’s outputs.
Marcus: That sounds like a realistic goal; AI handling the heavy lifting of structuring data while the human expert provides targeted validation is a solid way forward.
Ines: So, they’re looking at enhancing both the input types—text and figures—and making sure the extraction instructions are dynamically optimized for performance.
Marcus: It sounds like a plan to make this system more robust and scalable for handling the sheer volume of literature coming out every day in targeted protein degradation.
Ines: This whole idea points toward building a more versatile tool that doesn't just work well on one type of publication but can handle the messy reality of scientific writing.
Conclusion: Tom: So we’re wrapping up this look at "Agentic Literature Extraction Augments Targeted Protein Degradation Databases". Basically, they showed how this AI system can take messy publication data and turn it into structured, usable information for drug discovery models.
Ines: Yeah, the main point is that by using these agents to combine text and supplementary files, we can recover context—like assay types and timepoints—that was previously buried in the papers.
Marcus: And those recovered fields are huge because they make the data useful for condition-aware modeling; you can't just look at a compound in isolation anymore.
Yuki: From a population genetics standpoint, I see this as improving our ability to map out how different experimental conditions affect biological outcomes across various studies.
Ines: The implication is that we can keep those structured resources aligned with the pace of publication without needing exhaustive manual transcription from every single paper.
Marcus: That shifts the burden away from pure transcription toward AI-assisted curation, where domain experts just validate what the AI pulls out and corrects if it misses something important.
Yuki: It means that for us studying species evolution, we get better data because the underlying experimental context is preserved rather than lost in a simple text dump.
Ines: The limitation they flagged is that right now, it’s restricted to text and tables, so vision-language extraction from figures is what’s next for expanding the coverage.
Marcus: So, while it’s a big step for database augmentation—boosting those molecular glue and PROTAC counts by eighty-one and ninety-two percent—it still isn't fully visual yet.
Yuki: It’s a good step, though; it shows we can make these complex biological resources more accessible to the wider research community faster than ever before.
Ines: Exactly, because having this kind of structured data means we can finally build predictive models that truly understand how proteins degrade in different environments.
Marcus: So next time you see a paper on arXiv, think about how these agentic workflows might be the engine behind the data preparation that makes those findings actually useful.
Department of Computer Science and Engineering Chalmers University of Technology and University of Gothenburg · Yardi School of Artificial Intelligence Indian Institute of Technology Delhi
q-bio.QM, cs.LG
Submitted: 2026-05-11
Updated: 2026-11-04
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 85/100
The gist: Predictive models in biomedicine depend on structured assay data locked in the text, tables, and supplements of primary publications This bottleneck is especially acute in targeted protein…
Key concepts
- Expert-in-the-loop LLM workflow
- This is a system where a Large Language Model (LLM) performs data extraction, but it is guided and improved by human experts. The system uses three specialized AI agents—Extractor, Analyst, and Refiner—to work together. A separate refinement module (CAPO) also allows the AI to learn from expert feedback without needing extensive model retraining.
- Source-grounded extraction
- This method involves having the AI directly read and extract information from the original source documents, including text sections, tables, and supplementary files. This ensures the extracted data is directly tied to its original context in the publication, which is vital for accurately capturing complex experimental details in TPD studies.
- CAPO module
- CAPO stands for a prompt-refinement module that operates as a dual-loop system. The inner loop extracts information from the paper, and the outer loop uses an Assess–Diagnose–Refine cycle to automatically adjust the extraction instructions based on performance metrics. This allows the AI to continuously optimize its ability to extract accurate data.
- Triangular comparison framework
- This evaluation method validates AI predictions by comparing them against three sources: LLM outputs, standardized baseline records, and human expert-annotated ground truth. This comprehensive comparison helps quantify exactly how much new or previously missed valid data the AI pipeline has successfully recovered.
Terminology
Summary
Predictive models in biomedicine depend on structured assay data locked in the text, tables, and supplements of primary publications This bottleneck is especially acute in targeted protein degradation (TPD), where each assay record must combine compound identity, degradation target, recruiter, assay context, and endpoint values reported across sections, tables, and supplementary files. p. 0 se
Abstract
Predictive models in biomedicine depend on structured assay data locked in the
text, tables, and supplements of primary publications. This bottleneck is especially
acute in targeted protein degradation (TPD), where each assay record must combine
compound identity, degradation target, recruiter, assay context, and endpoint values
reported across sections, tables, and supplementary files. Second, we build an end-to-end
expert-in-the-loop workflow that combines source-grounded extraction, supplementary-material
integration, domain-specific post-processing, and a lightweight prompt-refinement module, CAPO,
for adapting extraction instructions from scarce annotations. Third, we show that the workflow
achieves high record-level accuracy across MGs and PROTACs, expands existing databases by 81%
and 92%, and recovers contextual fields such as timepoints, concentrations, and cell lines that are
essential for condition-aware modeling. With only seven annotated molecular glue publications, the workflow
achieved record-level F1 “ 0. 98 and transferred to PROTACs by terminology substitution alone, maintaining record-level F1 ą 0.Where the paper says it
Where the paper says it
Where the paper says it
The Problem Addressed
The bottleneck in data-driven drug discovery is the difficulty of turning published results into structured, machine-readable data. This problem is particularly evident in targeted protein degradation (TPD), where each assay record must combine compound identity, degradation target, recruiter, assay context, and endpoint values reported across sections, tables, and supplementary files. p. 0 se
Abstract
Predictive models in biomedicine depend on structured assay data locked in the
text, tables, and supplements of primary publications. This bottleneck is especially
acute in targeted protein degradation (TPD), where each assay record must combine
compound identity, degradation target, recruiter, assay context, and endpoint values
reported across sections, tables, and supplementary files. Inconsistent compound
identifiers and incomplete or implicit assay context further demand domain-specific
logic that generic LLM pipelines do not provide. Existing molecular glue and
PROTAC databases are manually curated and often lack the experimental context
required for downstream modeling.Where the paper says it
Where the paper says it
The Workflow Architecture
The workflow is an expert-in-the-loop LLM workflow that combines source-grounded extraction, supplementary-material integration, domain-specific post-processing, and a lightweight prompt-refinement module, CAPO. This module operates as an automated dual-loop system where the inner loop performs paper-level extraction and the outer loop iteratively refines the extraction prompt through an Assess–Diagnose–Refine cycle. p. 3 The
inner loop performs paper-level extraction, where the Extractor agent processes each article through
multi-stage reasoning (detailed in Sec. The outer loop iteratively refines the extraction prompt
through an Assess–Diagnose–Refine cycle. 2 Agentic architecture configuration
The workflow uses three GPT-5 agents (OpenRouter identifier: openai/gpt-5): an Extractor, an
Analyst, and a Refiner, together with a separate semantic matching module. The Extractor runs at
temperature 0.Where the paper says it
Where the paper says it
Evaluation Framework
The evaluation framework employs a triangular comparison among LLM predictions, standardized baseline records, and expert-annotated ground truth. The two primary modes are Experimental mode, which uses the full 5-element key R “ xC, T, E, A, Ly where C denotes the compound, and Mechanistic mode, which uses the 3-element key R “ xC, T, Ey. p. 4 3 Evaluation framework and metrics for TPD extraction
Each predicted record is compared against the reference using evaluation modes that differ in the
granularity at which a record R is defined. The two primary modes are Experimental mode, which
uses the full 5-element key R “ xC, T, E, A, Ly, where C denotes the compound, T the degradation
target, E the recruiter, A the assay type, and L the cell line; and Mechanistic mode, which uses
the 3-element key R “ xC, T, Ey and treats A and L as field-level quantities. Record-level evaluation determines whether the pipeline
correctly identifies a degradation assay, using compound-name matching with a Standard InChIKey
fallback and two-round conflict resolution (App. Field-level evaluation, applied only to recordlevel true positives (TPs), assesses whether the quantitative columns within a matched row are correct
under operator-aware comparisons that handle unit equivalences (e.Where the paper says it
Where the paper says it
Key Results and Augmentation
The workflow achieved record-level F1 “ 0.98 under both Experimental and Mechanistic modes for MGs, with field-level F1 of 0. p. 4 3 Evaluation framework and metrics for TPD extraction
Each predicted record is compared against the reference using evaluation modes that differ in the
granularity at which a record R is defined. The two primary modes are Experimental mode, which
uses the full 5-element key R “ xC, T, E, A, Ly, where C denotes the compound, T the degradation
target, E the recruiter, A the assay type, and L the cell line; and Mechanistic mode, which uses
the 3-element key R “ xC, T, Ey and treats A and L as field-level quantities. 98 and transferred to PROTACs by terminology substitution alone, maintaining record-level F1 ą 0. Applied at scale, it expanded
molecular glue and PROTAC databases by 81% and 92% records, respectively, with
92% and 82. With only seven annotated molecular glue publications, the workflow
achieved record-level F1 “ 0. 98 and transferred to PROTACs by terminology substitution alone, maintaining record-level F1 ą 0. 5–92% of newly recovered
records validated as correct on manual review. The largest gains came from recovering kinetic and
assay-context information that is underrepresented in manual curation but essential for cross-study
comparison and condition-aware modeling.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Implications for Curation
The results support a shift from exhaustive manual transcription toward AI-assisted curation, where domain experts provide targeted annotations and validate pipeline outputs. The workflow requires limited expert annotation and transfers across TPD modalities, offering a practical route for keeping structured TPD resources aligned with the pace of publication. p. 8 The workflow also recovers compounds more exhaustively
than manual curation, which tends to prioritize representative or highly active examples; extracting
all reported compounds regardless of activity provides the balanced active/inactive examples needed
to train generalizable degradation models. Because the workflow requires limited expert annotation
and transfers across TPD modalities, it offers a practical route for keeping structured TPD resources
aligned with the pace of publication and for extending similar curation workflows to other domains
where structured data remain scattered across the literature. The value of LLM-assisted data extraction for
dataset creation is not only their increased size (+81% MG records, +92% PROTAC records) but the
richer assay context they carry: timepoints, concentrations, assay type, and cell line, frequently missing from manually-extracted databases, are recovered alongside the core DC50 and Dmax endpoints. These fields matter because degradation activity is not a context-free compound property; comparing
potency across studies and training condition-aware models both depend on the assay configuration
under which each value was measured.Where the paper says it
Where the paper says it
Limitations
The main limitation of the workflow is its restriction to text and table modalities because many baseline-only records were figure-supported. Vision-language extraction from plots, blots, and other figures is the most direct path to improving coverage. p. 8 The main limitation of the workflow is its restriction to text and
table modalities. Because many baseline-only records were figure-supported (93% for MGs, 65% for
PROTACs), the expansion reflects augmentation of text/table-accessible evidence rather than complete
literature coverage; vision-language extraction from plots, blots, and other figures is the most direct
path to improving coverage. Because many baseline-only records were figure-supported (93% for MGs, 65% for
PROTACs), the expansion reflects augmentation of text/table-accessible evidence rather than complete
literature coverage; vision-language extraction from plots, blots, and other figures is the most direct
path to improving coverage. Additional scope constraints remain: evaluation with a single foundation
model (GPT-5), a small but audited ground-truth set (7 MG and 15 PROTAC papers), and restriction
to open-access PMC full text and degradation endpoints. Future work should test workflow portability
across models, expand expert annotations to tighten confidence intervals, and extend extraction to
ubiquitination, permeability, metabolism, and TPD-relevant measurements. The workflow is intended
to support, not replace, expert curation, since unverified extractions may propagate downstream errors.Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Domain-specific extraction workflow development allows AI systems to handle complex, multi-source data requirements like targeted protein degradation (TPD) database curation by combining source-grounded extraction with supplementary material integration and domain-specific post-processing. This enables systems to recover contextual fields such as
timepoints, concentrations, and cell lines that are essential for condition-aware modeling.
-
Implementation of an expert-in-the-loop prompt refinement module CAPO enables AI systems to adapt extraction instructions from scarce expert annotations without requiring model fine-tuning. This process ensures that the extraction logic is optimized based on empirical performance metrics like F1 scores, leading to record-level F1 scores of
0.98
for molecular glue records. -
The triangular comparison framework allows AI systems to perform robust validation by comparing LLM predictions against standardized baseline records and expert-annotated ground truth simultaneously. This enables users to quantify how much curated data the pipeline
rediscovers and how many additional valid records it recovers that were missed in existing databases.
-
The hierarchical evaluation framework, distinguishing between experimental mode (5-element key) and mechanistic mode (3-element key), allows AI systems to assess extraction quality at different levels of biological specificity. This enables precise measurement of field-level accuracy, such as
Precision 0.919 (0.142)
andRecall 0.933 (0.124)
for the MG cohort under Mechanistic mode matching. -
The semantic matching module provides a mechanism for AI systems to resolve entity synonymy across diverse reporting styles, such as different protein names or cell-line spellings. This allows the system to achieve consistency by checking non-identical pairs against a
symmetric semantic cache containing previously resolved comparisons.
Abstract
Predictive models in medicinal chemistry depend on structured assay data locked in primary publications. This bottleneck is acute in targeted protein degradation (TPD), where each assay record must combine compound identity, degradation target, recruiter, assay context, and endpoint values reported across sections, tables, and supplementary files. Existing TPD databases are manually curated and often lack assay context. We present an expert-in-the-loop large language model workflow in which a cross-validated prompt-refinement module adapts extraction instructions from scarce expert annotations. With only seven annotated molecular glue publications, the workflow achieved record-level F 1 = 0.98 and transferred to proteolysis-targeting chimeras (PROTACs) by terminology substitution alone (F 1 at least 0.93). The resulting database, AutoTPD, contained 81% and 92% more molecular glue and PROTAC records than standardized reference datasets; expert audits found 92/100 and 33/40 new records to be correct. Applying the workflow to eligible PubMed Central open-access literature added 4,483 records from 598 publications, yielding AutoTPD+, a public, versioned database of 7,622 TPD records, including degradation assay measurements, designed for monthly updates. In a controlled surrogate model comparison, PROTAC degradation activity models trained on AutoTPD+ saw a 8.9-18.2% error reduction relative to models trained on the original PROTAC-DB data.
Sources
- LeMat-Synth: a multi-modal toolbox to curate broad synthesis procedure databases from scientific literature
- An autonomous living database for perovskite photovoltaics
Related papers
- ViroGym: Realistic Large-Scale Benchmarks for Evaluating Viral Proteins
- CellCanon: An Agentic Cellular World Model for Prediction, Explanation, and Observation-Driven Updating
- Probing and steering biology across Boltz-1s trunk-diffusion boundary
- Trajectory Inference of Human Aging from Cross-Sectional DNA Methylation Data
- Bridging the Modality Bottleneck in Pathology MIL through Molecularly Guided Feature Adaptation
- Molecular notation can act as a shortcut in AI benchmark evaluation