AutoTPD+

A living targeted protein degradation database of molecular glue and PROTAC assay records, mined from the literature by an agentic LLM workflow.

Overview

Agentic Literature Extraction Augments Targeted Protein Degradation Databases

Yaochen Rao¹, Farzaneh Jalalypour¹, N. M. Anoop Krishnan², Rocío Mercado^1,3*

¹Chalmers University of Technology and University of Gothenburg · ²Yardi School of Artificial Intelligence, IIT Delhi · ³SciLifeLab, Gothenburg · ^*Corresponding author

From papers to records, every month.

AutoTPD+ is a living database of PROTAC and molecular glue degradation assays. An LLM workflow mines open-access PubMed Central papers for each compound's target, recruiter, assay, cell line and DC₅₀/D_max values. PROTAC-DB and MolGlueDB records cover papers without open full text. A monthly search adds new papers.

Updated monthly · last update —

At a glance

What is in AutoTPD+

—

assay records

—

source publications

—

unique chemical structures

modalities: PROTAC and molecular glue

—

assay records

—

degradation targets

—

recruiters (E3 ligases and others)

—

cell lines

Where the records come from

Records and papers by source

Most frequent targets

Records per degradation target (gene symbol)

Recruiters

Records per recruiter (gene symbol)

Assay families

Records per standardized assay technique

Data format

One row, one degradation assay

Example record

AUTOTPD-MG-000155 · 10.1021/acs.jmedchem.0c01313

Compound: CC-90009
SMILES: O=C1CCC(N2Cc3cc(CNC(=O)C(F)(F)c4ccc(Cl)cc4)ccc3C2=O)C(=O)N1
Target: GSPT1 · UniProt P15170
Recruiter: CRBN · UniProt Q96SW2
Assay: immunoblot → Western blot
Cell line: MV4-11 · CVCL_0064
DC₅₀: 1.6 nM at 24 h
D_max: > 90% at 24 h

Compound

Reported name, IUPAC name, SMILES, InChIKey

Target and recruiter

Reported name, gene symbol, UniProt accession, sequence

Assay context

Assay and its standardized family; cell line with Cellosaurus accession and species

Endpoints

DC₅₀ with units and time; D_max with time and concentration; qualifiers (<, >, ~) and ranges

Provenance

Source DOI and a stable AutoTPD_ID

All columns

Column	Description	Example (AUTOTPD-MG-000155)
Identifiers
`AutoTPD_ID`	Stable row identifier, unchanged across monthly releases	AUTOTPD-MG-000155
`Record_ID`	Record ID in the source database; empty for LLM-extracted records	—
`Compound_ID`	Compound ID in the source database; empty for LLM-extracted records	—
`DOI`	Source publication	10.1021/acs.jmedchem.0c01313
Compound
`Compound_Name`	Compound name or label as reported in the paper	CC-90009
`IUPAC_Name`	IUPAC name	2-(4-chlorophenyl)-N-[[2…
`IUPAC_Name_Source`	Where the IUPAC name came from (PAPER, PUBCHEM)	PUBCHEM
`SMILES`	RDKit canonical SMILES	O=C1CCC(N2Cc3cc(CNC(=O)C(F)(F)c4ccc(Cl)cc4)ccc3C2=O)C(=O)N1
`SMILES_Source`	Where the SMILES came from (PAPER, PUBCHEM, OPSIN, source database)	PUBCHEM
`StdInChIKey`	Standard InChIKey computed by RDKit from the standardized canonical SMILES	PWBHUSLMHZLGRN-UHFFFAOYSA-N
`Connectivity_Key`	First block of the InChIKey (connectivity only, ignores stereochemistry)	PWBHUSLMHZLGRN
Degradation target
`Degradation_Target`	Target as reported in the paper	GSPT1
`Degradation_Target_Gene`	HGNC gene symbol	GSPT1
`Degradation_Target_Uniprot`	UniProt accession	P15170
`Degradation_Target_Uniprot_MutationID`	UniProt variant ID when a specific mutant is targeted	—
`Degradation_Target_Sequence`	UniProt protein sequence	MELSEPIVENGETEMSPEESWEHK…
Recruiter
`Recruiter`	Recruiter (usually an E3 ligase) as reported	CRBN
`Recruiter_Gene`	Gene symbol	CRBN
`Recruiter_Uniprot`	UniProt accession	Q96SW2
`Recruiter_Sequence`	UniProt protein sequence	MAGEGDQQDAAHNMGNHLPLLPAE…
Assay context
`Assay`	Assay as described in the paper	immunoblot
`Assay_STD`	Standardized assay technique family	Western blot
`Cell_Line`	Cell line as reported, or cell-free	MV4-11
`Cell_Line_STD`	Cellosaurus recommended name	MV4-11
`Cell_Line_STD_Accession`	Cellosaurus accession (RRID)	CVCL_0064
`Cell_Line_Modification`	Engineered construct, e.g. a HiBiT knock-in	—
`Cell_Line_Species`	Species of the cell line	Homo sapiens
DC₅₀
`DC50`	Concentration giving 50% degradation	1.6
`DC50_units`	Units of DC50	nM
`DC50_h`	Treatment time in hours	24
`DC50_operator`	Qualifier: <, >, ~, <=, >=	—
`DC50_range_min`	Range low end, or mean − error	—
`DC50_range_max`	Range high end, or mean + error	—
D_max
`Dmax`	Maximal degradation, in %	90
`Dmax_h`	Treatment time in hours	24
`Dmax_operator`	Qualifier: <, >, ~, <=, >=	>
`Dmax_range_min`	Range low end, or mean − error	—
`Dmax_range_max`	Range high end, or mean + error	—
D_max concentration
`Dmax_conc`	Concentration at which Dmax was measured	—
`Dmax_conc_units`	Units of Dmax_conc	—
`Dmax_conc_operator`	Qualifier for Dmax_conc	—
`Dmax_conc_range_min`	Range low end	—
`Dmax_conc_range_max`	Range high end	—

Data

Download

AutoTPD+ living · monthly

AutoTPD extended to all open-access MG and PROTAC papers plus PROTAC-DB and MolGlueDB, updated monthly. Use it for analysis and model training.

AutoTPD benchmark · frozen

Records extracted from the 211 papers used to benchmark extraction accuracy in the paper.

Extraction covers text and tables, not figures.

Data: CC BY 4.0 · Code: MIT. Source publications, PROTAC-DB and MolGlueDB keep their own licenses.

Citation

Cite AutoTPD+

@article{rao2026autotpd,
  title   = {Agentic Literature Extraction Augments Targeted Protein
             Degradation Databases},
  author  = {Rao, Yaochen and Jalalypour, Farzaneh and Krishnan, N. M. Anoop
             and Mercado, Rocío},
  journal = {arXiv preprint arXiv:2605.11221},
  year    = {2026}
}

Supported by the Swedish Research Council (VR) and the Wallenberg AI, Autonomous Systems and Software Program (WASP). Computations enabled by Chalmers e-Commons and NAISS.