Exploratory mapping of tumor associated macrophage nanoparticle article abstracts using an eLDA topic modeling machine learning approach
Chloe Brown, Colette S. M. Bilynsky, Melanie Gainey, Sarah Young, John Kitchin and Elizabeth C. Wayne, “Exploratory mapping of tumor associated macrophage nanoparticle article abstracts using an eLDA topic modeling machine learning approach”, PLOS ONE 19(6), e0304505 (2024). https://doi.org/10.1371/journal.pone.0304505
This summary was written by a language model from the paper and has not been checked line by line. Rely on the published paper for details.
Question
Nanoparticles that target tumor associated macrophages (TAMs) in solid tumors have become a fast-growing research area spread over many disciplines, journals and databases, with nomenclature that differs between fields. The paper asks whether unsupervised topic modeling of abstracts can give a useful bird’s-eye view of this literature, identifying its main themes, the nanoparticle types, tumor types and therapies associated with each, and whether the same model can sort newly published papers into existing themes to keep a scoping review current (a “living review”) (Brown et al. 2024).
Approach
The corpus came from a registered scoping review protocol on macrophage polarization in nanoparticle-based cancer immunotherapy. Two information specialists built the search string, which was run in 2020 over Web of Science Core Collection, Scopus, IEEE Xplore, Medline (PubMed) and Biotechnology and BioEngineering Abstracts plus hand searching and Google Scholar, for papers from 2000 onward. Records were deduplicated in Zotero and screened at title and abstract level and then full text in Sysrev, with two independent reviewers and a third to resolve conflicts. Sysrev’s built-in classifier, trained on 7,482 manually screened records, was used to auto-exclude 3,460 records predicted below 40% inclusion; 14 of the training records (0.19%) would have been wrongly excluded at that threshold and were added back by hand (Brown et al. 2024).
The abstracts were lowercased, tokenized with an NLTK regular expression tokenizer, stripped of English stop words plus domain stop words drawn from the search string (such as cancer, nanoparticle and macrophage), single characters, and words occurring once or in 90% of abstracts, then stemmed with the Porter stemmer. Plain LDA gave noisy, incoherent topics that did not improve with hyperparameter tuning, so the authors used Gensim’s ensemble LDA (eLDA), which pools many LDA runs and clusters the resulting topics with a DBSCAN variant to separate stable from noisy topics. Models were judged quantitatively by average Cv topic coherence and qualitatively by domain experts reading the top 30 words per topic. A document was assigned to a topic when that topic’s weight was at least 0.4, and the 100 most frequent terms per topic were sorted by the authors into six word categories (immunology, nanoparticles, imaging, biologics, cancer, therapeutics) (Brown et al. 2024).
Key findings
Searches returned 28,217 records; 15,374 remained after deduplication and removal of retractions, and 854 passed full-text screening to train the model. Publication counts showed two growth inflection points, which the authors relate to the 2011 FDA approval of the checkpoint inhibitor ipilimumab and the 2017 approval of the first CAR T therapy. The median number of authors per paper was about 8, and Biomaterials, Journal of Controlled Release and ACS Nano were the most common journals (Brown et al. 2024).
The eLDA model reached a Cv coherence of 0.589, and coherence was insensitive to hyperparameters and number of topics. It found six topics: Immune and TAMs (the largest, centered on repolarizing TAMs with cytokines, microRNA delivery or small molecules acting on pathways such as NF-kB and TLR 7/8), Nanoparticles (formulation, PK/PD and cytotoxicity, dominated by polymeric carriers such as PLGA, PEG, PLA and chitosan loaded with paclitaxel, docetaxel or doxorubicin), Imaging (metallic particles such as iron and gold for MRI contrast, often mannose-conjugated, with breast, liver and brain prominent), Biologics Delivery (one of the smallest, featuring extracellular vesicles and exosomes carrying nucleic acids, and membrane-coated particles), Vaccines (antigens with adjuvants such as CpG ODNs, frequently in liposomes, with lymph nodes prominent) and Multi-modal Therapies (photothermal, photodynamic or magnetic approaches combined with chemo- or immunotherapy, with glioblastoma prominent). In documents assigned to the Nanoparticles topic, nanoparticle terms made up about 24% of occurrences of the top 100 words (Brown et al. 2024).
Topics were not fully reproducible across reruns because LDA initialization is random. Immunology and Nanoparticles topics appeared consistently, while Imaging and Multi-modal Therapies tended to merge into Nanoparticles and Vaccines and Biologics Delivery into Immunology. Applied without retraining to 95 newer abstracts, the model recovered five of the six topics (all but Biologics Delivery) with word clouds consistent with the original topics (Brown et al. 2024).
Significance
The study shows that an ensemble topic model built on a carefully screened corpus can summarize hundreds of biomedical abstracts that would be impractical to code by hand, and that the fitted model can classify new papers as a proof of concept for a living scoping review. The authors present topic modeling as a supplement to, not a replacement for, manual evidence synthesis, stress that topics reflect writing style as well as content and that different runs can organize the same data differently, and recommend working with information specialists when designing searches (Brown et al. 2024). For the Kitchin group it is an application of text mining and machine learning outside catalysis, building on the LDA abstract analysis demonstrated with pybliometrics.