Data curation

Despite the vast economic and time resources devoted over the last three decades in actively investigating treatments for neurodegenerative amyloidosis, their success remains scarce. While none of the 14 different compounds that targetting amyloid-β that made it through phase III clinical trials made it through, two anti-amyloid antibodies have received regulatory approval in Europe (lecanemab and donanemab), and a third one in the United States (aducanumab), for treating Alzheimer’s disease (AD). Lecanemab and donanemab have shown to slow cognitive decline by around 30%. Long-term clinical evidence support the therapeutic concept of antibody-removal of amyloid deposits, as demonstrated in post-mortem evaluations of Alzheimer’s disease patients 15 years post-immunization”. Additionally, mAbs are in clinical trials for various amyloidoses with no current disease-modifying therapies, including Parkinson’s Disease (PD), Transthyretin amyloidosis or Amyloid light chain (AL) amyloidosis. Instead, multiple mAbs have shown insufficient results or are only useful for certain disease stages, genotypes or phenotypes. Understanding the molecular mechanisms of why certain antibody therapeutics succeed and others do not could allow for increasing the success rate and decreasing economic burden for these incurable conditions.

The AmyloGraph Antibody Database collects peer-reviewed experimental information on the impact of amyloid-targeting antibodies have on the amyloid aggregation.

We establish controlled-vocabulary to describe the interaction, considering:

Amyloid proteins

Amyloid proteins were obtained according to the latest update of the International Society of Amyloidosis (ISA) Nomenclature Committee, from the amyloid fibrol ptoreins and precursors in human (Aβ, AαSyn, ATau, APrP, ATTR, etc), as well as the Pathological intracellular protein aggregates which display at least some typical amyloid fibril properties (TARDBP, TIA1, SOD1, C9orf72, etc). These were supplemented with bacterial and yeast amyloids (CsgA, Sup35, etc). Finally, mammalian homologues of amyloid proteins (mice Aβ, APrP, ATau) were accepted as amyloid proteins if proof of amyloid aggregation was shown on the respective publication. The basic amyloid proteins initially considered for the AmyloGraph Antibody Database are reported in the following tables.

Human amyloid proteins

Amyloid Species Canonical Protein Entry UniProt Link DisGeNET Link
A𝛽-42 P05067 UniProt Entry DisGeNET Entry
A𝛽-40 P05067 UniProt Entry DisGeNET Entry
pE3-Aβ P05067 UniProt Entry DisGeNET Entry
ɑ-Synuclein P37840 UniProt Entry DisGeNET Entry
Tau P10636 UniProt Entry
Immunoglobulin light chain Multi-entry Variable dependent on clonal sequence Variable dependent on clonal sequence
Immunoglobulin heavy chain Multi-entry Variable dependent on clonal sequence Variable dependent on clonal sequence
(Apo) Serum amyloid A P0DJI8 UniProt Entry DisGeNET Entry
Transthyretin P02766 UniProt Entry DisGeNET Entry
𝛽2-microglobulin P61769 UniProt Entry DisGeNET Entry
Apolipoprotein A-I P02647 UniProt Entry DisGeNET Entry
Apolipoprotein A-II P02652 UniProt Entry DisGeNET Entry
Apolipoprotein A-IV P06727 UniProt Entry DisGeNET Entry
Apolipoprotein C-II P02655 UniProt Entry DisGeNET Entry
Apolipoprotein C-III P02656 UniProt Entry DisGeNET Entry
Gelsolin P06396 UniProt Entry DisGeNET Entry
Lysozyme P61626 UniProt Entry DisGeNET Entry
Leukocyte chemotactic factor 2 O14960 UniProt Entry DisGeNET Entry
Fibrinogen ɑ P02671 UniProt Entry DisGeNET Entry
Cystatin C P01034 UniProt Entry DisGeNET Entry
ABriPP Q9Y287 UniProt Entry DisGeNET Entry
ADanPP Q9Y287 UniProt Entry DisGeNET Entry
Prion protein P04156 UniProt Entry DisGeNET Entry
Transmembrane protein 106B Q9NUM4 UniProt Entry DisGeNET Entry
(Pro) Calcitonin P01258 UniProt Entry DisGeNET Entry
Islet amyloid polypeptide P10997 UniProt Entry DisGeNET Entry
Atrial natriuretic peptide P01160 UniProt Entry DisGeNET Entry
Prolactin P01236 UniProt Entry DisGeNET Entry
(Pro) Somatostatin P61278 UniProt Entry DisGeNET Entry
Glucagon P01275 UniProt Entry DisGeNET Entry
Parathyroid hormone P01270 UniProt Entry DisGeNET Entry
Lung surfactant protein C P11686 UniProt Entry DisGeNET Entry
Corneodesmosin Q15517 UniProt Entry DisGeNET Entry
Lactadherin (Medin) Q08431 UniProt Entry DisGeNET Entry
Kerato-epithelin (TGFBI) Q15582 UniProt Entry DisGeNET Entry
Lactoferrin P02788 UniProt Entry DisGeNET Entry
Odontogenic ameloblast-associated protein A1E959 UniProt Entry DisGeNET Entry
Semenogelin-1 P04279 UniProt Entry DisGeNET Entry
Cathepsin K P43235 UniProt Entry DisGeNET Entry
EGF-containing fibulin-like extracellular matrix protein 1 Q12805 UniProt Entry DisGeNET Entry
Cytokeratin P02533 UniProt Entry DisGeNET Entry

Human proteins forming inclusions with amloid properties

Amyloid-like Species Canonical Protein Entry UniProt Link DisGeNET Link
p53 P04637 UniProt Entry DisGeNET Entry
Desmin P17661 UniProt Entry DisGeNET Entry
Galectin 7 P47929 UniProt Entry DisGeNET Entry
Superoxide dismutase [SOD1] P00441 UniProt Entry DisGeNET Entry
TAR DNA-binding protein 43 [TDP-43] Q13148 UniProt Entry DisGeNET Entry
Huntingtin P42858 UniProt Entry DisGeNET Entry
Androgen receptor P10275 UniProt Entry DisGeNET Entry

Microbial Amyloid Proteins

Amyloid Species Specie - Genus UniProt Link
CsgA E. coli / Aeromonas UniProt Entry
CsgB E. coli / Aeromonas UniProt Entry
FapB Pseudomonas UniProt Entry
FapC Pseudomonas UniProt Entry
SSB Campylobacter hominis UniProt Entry
Transcription termination factor Rho Clostridium botulinum UniProt Entry
Sup35 Saccharomyces cerevisiae UniProt Entry
Ure2 Saccharomyces cerevisiae UniProt Entry
HET-s Podospora anserina UniProt Entry

Note: different microbial species and strains have shown to form amyloid aggregates. Therefore, beyond this list, specific proteins of these families are accepted.

Antibodies and antibody-based designs

The database considers the effect of a single monoclonal antibody on the amyloid aggregation of an amyloid protein. Different species IgGs, orthodox, IgG inspired, and heterodox, novel multimeric, antibody formats were accepted and named according to the original publication, or according to the following classification.

Antibody-conjugated nanoparticles and nanoparticles engineered with an antibody derivative, on the other hand, were consdidered separately. Following the World Health Organization (WHO) International Nonproprietary Names (INN) naming convention, immunoglobulin fusions with peptides were as well considered an antibody if both domains have immunoglobulin derived variable domains. Conjugates with small molecules or radioisotopes as well. Larger nanoparticles than 50 nm, approximately 5 times a typical IgG, were discarded.

Therapeutic monoclonal antibodies evaluated in clinical trials

Tha AmyloGraph Antibody Database compiles data for most monoclonal antibodies evaluated in clinical trials. These species are catalogues in the table below.

INN or development name IMGT DB PubChem Thera-SAbDab YABS DB Other IDs Parental/Heterologous Antibody ID
Aducanumab IMGT Entry PubChem SAbDab YABS DB BART, BIIB-037, BIIB037, aducanumab-avwa, Aduhelm None
Gantenerumab IMGT Entry PubChem SAbDab YABS DB R1450, RG-1450, RG1450, Ro-4909832, RO4909832 None
Donanemab IMGT Entry PubChem SAbDab YABS DB LY3002813, donanemab-azbt mE8
Crenezumab IMGT Entry PubChem SAbDab N/A MABT5102A, RG7412, MABT-5102A, RG-7412 mC2
Solanezumab IMGT Entry PubChem SAbDab YABS DB LY2062430 m266
Lecanemab IMGT Entry PubChem SAbDab YABS DB BAN-2401, BAN2401, lecanemab-irmb, Leqembi, Leqembi™ mAb158
Bapineuzumab IMGT Entry PubChem SAbDab YABS DB AAB-001 3D6
Cliramitug IMGT Entry PubChem SAbDab N/A ALXN-2220, ALXN2220 NI006
Anselamimab IMGT Entry PubChem SAbDab YABS DB CAEL-101, CAEL101 11-1F4
Birtamimab IMGT Entry PubChem SAbDab YABS DB NEOD001, NEOD-001, ELT1-01 2A4
Etalanetug IMGT Entry PubChem SAbDab YABS DB E2814 7G6
SAR228810 N/A N/A N/A N/A 13C3, SAR255952 N/A
Prasinezumab IMGT Entry PubChem SAbDab N/A PRX002, RG-7935, RG7935 9E4
Cinpanemab IMGT Entry PubChem SAbDab N/A BIIB054 12F4
Amlenetug IMGT Entry PubChem SAbDab YABS DB Lu AF82422, GM37 GM37
Exidavnemab IMGT Entry PubChem SAbDab N/A BAN0805 Ab47
Indenebart IMGT Entry PubChem SAbDab N/A MEDI1341, TAK-341 None
Ponezumab IMGT Entry PubChem SAbDab N/A PF-04360365 2H6
Gosuranemab IMGT Entry PubChem SAbDab N/A BIIB092 IPN002

The AmyloGraph Antidoby Database Curation Pipeline

Data acquisition

We designed keyword-based queries to obtain peer-reviewed publications reporting the effect of antibody and antibodies or nanobody and nanobodies, on amyloid aggregation.


- "amyloid"[Title/Abstract] AND "antibod*"[Title/Abstract]

- "amyloid"[Title/Abstract] AND "nanobod*"[Title/Abstract]

This query was executed on October 24, 2024. After removing duplicates, a total of 4023 unique publications were identified. An additional 58 publications were identified through manual screening of reference lists and included in the final analysis.

Inclusion-Exclusion Criteria & Curation Tags

Manuscripts that failed to demostrate experimentally the effect on amyloid aggregation on antibody-amyloid protein interaction were excluded during this initial curation phase. We used a multi-tagging system that enabled documenting multiple reasons for a single article’s rejection. The definition of the exclusion criteria are as follow:

1. Missing or Insufficient Experimental Information
  • Not Enough Experimental Data: The publication does not directly capture or quantify the antibody’s specific effect on amyloid formation. (For example when an antibody was used only for detection).
  • (Pre)Clinical Trials (No Interaction/Amyloid Data): Clinical trial reports that lack amyloid aggregation data.
  • No Interactions Described: The text provides insufficient data describing the functional effect of the antibody on the target amyloid system.
  • The Interactor is Not an Amyloid Protein: The target protein falls outside our designated amyloid-forming criteria.
  • The Interactor is Not an Antibody: The targeting molecule does not qualify as an antibody, a valid antibody derivative. (For example publications describing Amyloid light-chain amyloidosis, where immunoglobulin chains are not participating as an antibody but an amyloid, and no antibody directed to them exists)
  • Unknown Antibody Type: The molecule features a highly unconventional architecture or excessive modifications, preventing it from being accurately classified under the established antibody definitions.
2. Publication Formats & Methodology Filters
  • Review article: The paper summarizes existing literature without presenting novel, primary experimental data. Despite PubMed tags many reviews automatically, we manually audit the remaining ~50% to ensure no secondary sources enter the pipeline. This category also included editorial, perspectives, viewpoints and other non-article publications, statements and similar that do not provide new information.
  • Pre-print: The manuscript has not yet undergone formal peer review. To maintain maximum data integrity, pre-prints are systematically excluded.
  • In silico information Only: We prioritize experimental evidence. Studies relying solely on Molecular Dynamics (MD) simulations or in silico modeling are discarded, unless the computational data complements in vitro or in vivo experiments.
  • Non-English pulication: To ensure flawless comprehension, we discard publications not published in English.
  • Other: Anomalous edge cases that do not map to standard exclusionary categories.

Initial curation

After first assessment on the query and non-query publications, 6.81% (209) publications were deemed as “Useful”, 167 corresponding to the query and 42 to the non-query bibliographic addition.

After the secondary curation, 8 (3.82%) of the initially taged as “Useful” publications were re-avaluated as “Not Useful”, while 1 publicaiton was retracted during the curation process.

Development of a Machine Learning method to improve curation queues

The large body of publications to initially review presented a significant bottleneck in terms of human and time resources. In parallel, we developed a Machine Learning methodology which could help prioritize curation queues. Once 25% of the publications were assessed, we used the Title, Abstracts and references of the positive “Useful” and negative “Rejected” publications to train this algorithm. The subsequent iterations of it were fed with the increasing number of reviewed publications in a continuous loop. Additionally, the revisions were tagged with “Reason for Rejection” (as presented in the Inclusion-Exclusion Criteria & Curation Tags section), to improve the ML curation queues development.

The curation of the AmyloGraph Antibody DB, and development of a ML curation queue algorithm, were considered as a topic for the German ELIXIR BioHackathon, 2025.

Two-step indepedent data curation

To ensure maintaining curation consistency across time and different curators, we implemented various overlapping strategies. We scheduled regular meetings between the Leading Curators and the Junior Curators, a one-on-one introduction meeting with each Junior Curator, provided a Curator handbook with definitions, FAQs and curation examples, and set-up a Slack channel for curation-related discussions.

Curator handbook with examples

Curators were provided with a reference guide featuring clear definitions FAQs, and real world curation examples. Definitions include what is considered an antibody, an amyloid, the biophysical, biomedical and biochemical basis of the most employed techniques to report changes in amyloid aggregation. FAQs include how to report non-human versions of the amyloid proteins, how to consider ambiguous entities (such as publications reporting “amyloid plaques” instead of a unique Aβ₄₂ or Aβ₄₀ species), how to annotate Immunoglobulin light chain (AL) amyloidosis when no unnequivocal sequential data is provided, etc. This trianing resource ensures maximum reproducibility and consistency across all annotators of the AmyloGraph Antibody Database.

First step: standardising annotation and description of the data.

To ensure maximum granularity and data integrity, our curation process divides each publication across three interconnected modules: Antibody Properties, Amyloid Targets, and Interaction Profiles. To maximize data standardization, these modules are pre-populated (in the forms of Google Forms), with acceptable answers to publication doi, the amyloid protein, the antibody, antibody type, impact of the antibody on the aggregation, etc.

  • If a publication reports the effect on multiple (>1) amyloids, the curator will fill in one “Amyloid form” for each species.
  • If a publication reports several antibodies, the curator will fill in one “Antibody form” for each species.

Finally, the curator will fill one “Interaction form” for each antibody-amyloid pair.

Second, independent, machine-assisted curation

Secondary curation was assisted by means of two python libraries: MarkItDown and LangExtract to accelerate curation and minimize data leakage.

  • MarkItDown, developed by Microsoft, converts different document formats—including into clean, structured Markdown text.

MarkItDown was implemented for PDF-only publications, to generate Markdown files that could be processed by LangExtract.

  • LangExtract, developed by Google, uses LLMs to read text-based publications, highlighting the rellevant information and develop human-friendly htmls with the rellevant text highlighted.

Text data from the publications were processed by LangExtract coupled with Ollama running the Gemma3:4b model.

Further considerations

Amyloid plaques

Amyloid plaques are composed of different proteotypes of the amyloid βeta peptide, such as Aβ₄₂, Aβ₄₀, Aβ₃₈, and Pyroglutamate-Aβ (pE3-Aβ). Despite this, the overwhelingly major constituent is the Aβ₄₂ peptide (42 amino acids). Therefore, for in vivo experiments reporting changes in amyloid plaques (plaque number, plaque density, Positron Emission Tomography uptake values…) we annotate Aβ₄₂ as the amyloid interactor.

Considerations on experimental amyloid reporters

The ISA recommends the use of Serum Free Light Chains (sFLC) as a the biomarker to follow Immunoglobulin light chain (AL) amyloidosis disease progression and therapeutic response number. sFLC reports the disease progression, but is an indirect reporter of the amyloid state, the amyloid deposits. Therefore, we have not included it as evidence of changes in amyloid aggregation. In a similar fashion, we have considered “indirect” other reporters such as clinical dementia rating, cellular cytotoxicity, Morris-water maze performance, brain soluble Aβ₄₀ or Aβ₄₂ levels, microglia uptake, or % truncated α-Synuclein.

:::