Motivation: Biomedical text mining is becoming increasingly important as the number of biomedical documents rapidly grows. With the progress in natural language processing (NLP), extracting valuable information from biomedical literature has gained popularity among researchers, and deep learning has boosted the development of effective biomedical text mining models. However, directly applying the advancements in NLP to biomedical text mining often yields unsatisfactory results due to a word distribution shift from general domain corpora to biomedical corpora. In this article, we investigate how the recently introduced pre-trained language model BERT can be adapted for biomedical corpora. Results: We introduce BioBERT (Bidirectional Encoder Representations from Transformers for Biomedical Text Mining), which is a domain-specific language representation model pre-trained on large-scale biomedical corpora. With almost the same architecture across tasks, BioBERT largely outperforms BERT and previous state-of-the-art models in a variety of biomedical text mining tasks when pre-trained on biomedical corpora. While BERT obtains performance comparable to that of previous state-of-the-art models, BioBERT significantly outperforms them on the following three representative biomedical text mining tasks: biomedical named entity recognition (0.62% F1 score improvement), biomedical relation extraction (2.80% F1 score improvement) and biomedical question answering (12.24% MRR improvement). Our analysis results show that pre-training BERT on biomedical corpora helps it to understand complex biomedical texts. Availability and implementation: We make the pre-trained weights of BioBERT freely available at https://github.com/naver/biobert-pretrained, and the source code for fine-tuning BioBERT available at https://github.com/dmis-lab/biobert.

As the volume of publications rapidly increases, searching for relevant information from the literature becomes more challenging. To complement standard search engines such as PubMed, it is desirable to have an advanced search tool that directly returns relevant biomedical entities such as targets, drugs, and mutations rather than a long list of articles. Some existing tools submit a query to PubMed and process retrieved abstracts to extract information at query time, resulting in a slow response time and limited coverage of only a fraction of the PubMed corpus. Other tools preprocess the PubMed corpus to speed up the response time; however, they are not constantly updated, and thus produce outdated results. Further, most existing tools cannot process sophisticated queries such as searches for mutations that co-occur with query terms in the literature. To address these problems, we introduce BEST, a biomedical entity search tool. BEST returns, as a result, a list of 10 different types of biomedical entities including genes, diseases, drugs, targets, transcription factors, miRNAs, and mutations that are relevant to a user’s query. To the best of our knowledge, BEST is the only system that processes free text queries and returns up-to-date results in real time including mutation information in the results. BEST is freely accessible at http://best.korea.ac.kr.

Fusion genes represent an important class of biomarkers and therapeutic targets in cancer. ChimerDB is a comprehensive database of fusion genes encompassing analysis of deep sequencing data (ChimerSeq) and text mining of publications (ChimerPub) with extensive manual annotations (ChimerKB). In this update, we present all three modules substantially enhanced by incorporating the recent flood of deep sequencing data and related publications. ChimerSeq now covers all 10 565 patients in the TCGA project, with compilation of computational results from two reliable programs of STAR-Fusion and FusionScan with several public resources. In sum, ChimerSeq includes 65 945 fusion candidates, 21 106 of which were predicted by multiple programs (ChimerSeq-Plus). ChimerPub has been upgraded by applying a deep learning method for text mining followed by extensive manual curation, which yielded 1257 fusion genes including 777 cases with experimental supports (ChimerPub-Plus). ChimerKB includes 1597 fusion genes with publication support, experimental evidences and breakpoint information. Importantly, we implemented several new features to aid estimation of functional significance, including the fusion structure viewer with domain information, gene expression plot of fusion positive versus negative patients and a STRING network viewer. The user interface also was greatly enhanced by applying responsive web design. ChimerDB 4.0 is available at http://www.kobic.re.kr/chimerdb/.

Kim S, et al. Korean J Radiol. 2024 Feb;25(2):126-133. https://doi.org/10.3348/kjr.2023.0997

Drug combination therapy, which is considered as an alternative to single drug therapy, can potentially reduce resistance and toxicity, and have synergistic efficacy. As drug combination therapies are widely used in the clinic for hypertension, asthma, and AIDS, they have also been proposed for the treatment of cancer. However, it is difficult to select and experimentally evaluate effective combinations because not only is the number of cancer drug combinations extremely large but also the effectiveness of drug combinations varies depending on the genetic variation of cancer patients. A computational approach that prioritizes the best drug combinations considering the genetic information of a cancer patient is necessary to reduce the search space. We propose an in-silico method for personalized drug combination therapy discovery. We predict the synergy between two drugs and a cell line using genomic information, targets of drugs, and pharmacological information. We calculate and predict the synergy scores of 583 drug combinations for 31 cancer cell lines. For feature dimension reduction, we select the mutations or expression levels of the genes in cancer-related pathways. We also used various machine learning models. Extremely Randomized Trees (ERT), a tree-based ensemble model, achieved the best performance in the synergy score prediction regression task. The correlation coefficient between the synergy scores predicted by ERT and the actual observations is 0.738. To compare with an existing drug combination synergy classification model, we reformulate the problem as a binary classification problem by thresholding the synergy scores. ERT achieved an F1 score of 0.954 when synergy scores of 20 and -20 were used as the threshold, which is 8.7% higher than that obtained by the state-of-the-art baseline model. Moreover, the model correctly predicts the most synergistic combination, from approximately 100 candidate drug combinations, as the top choice for 15 out of the 31 cell lines. For 28 out of the 31 cell lines, the model predicts the most synergistic combination in the top 10 of approximately 100 candidate drug combinations. Finally, we analyze the results, generate synergistic rules using the features, and validate the rules through the literature survey. Using various types of genomic information of cancer cell lines, targets of drugs, and pharmacological information, a drug combination synergy prediction pipeline is proposed. The pipeline regresses the synergy level between two drugs and a cell line as well as classifies if there exists synergy or antagonism between them. Discovering new drug combinations by our pipeline may improve personalized cancer therapy.

Motivation: Traditional drug discovery approaches identify a target for a disease and find a compound that binds to the target. In this approach, structures of compounds are considered as the most important features because it is assumed that similar structures will bind to the same target. Therefore, structural analogs of the drugs that bind to the target are selected as drug candidates. However, even though compounds are not structural analogs, they may achieve the desired response. A new drug discovery method based on drug response, which can complement the structure-based methods, is needed. Results: We implemented Siamese neural networks called ReSimNet that take as input two chemical compounds and predicts the CMap score of the two compounds, which we use to measure the transcriptional response similarity of the two compounds. ReSimNet learns the embedding vector of a chemical compound in a transcriptional response space. ReSimNet is trained to minimize the difference between the cosine similarity of the embedding vectors of the two compounds and the CMap score of the two compounds. ReSimNet can find pairs of compounds that are similar in response even though they may have dissimilar structures. In our quantitative evaluation, ReSimNet outperformed the baseline machine learning models. The ReSimNet ensemble model achieves a Pearson correlation of 0.518 and a precision@1% of 0.989. In addition, in the qualitative analysis, we tested ReSimNet on the ZINC15 database and showed that ReSimNet successfully identifies chemical compounds that are relevant to a prototype drug whose mechanism of action is known. Availability and implementation: The source code and the pre-trained weights of ReSimNet are available at https://github.com/dmis-lab/ReSimNet. Supplementary information: Supplementary data are available at Bioinformatics online.