PEDL: extracting protein-protein associations using deep language models and distant supervision

HunFlair2 in a cross-corpus evaluation of biomedical named entity recognition and normalization tools

MOTIVATION

With the exponential growth of the life sciences literature, biomedical text mining (BTM) has become an essential technology for accelerating the extraction of insights from publications. The identification of entities in texts, such as diseases or genes, and their normalization, i.e. grounding them in knowledge base, are crucial steps in any BTM pipeline to enable information aggregation from multiple documents. However, tools for these two steps are rarely applied in the same context in which they were developed. Instead, they are applied "in the wild," i.e. on application-dependent text collections from moderately to extremely different from those used for training, varying, e.g. in focus, genre or text type. This raises the question whether the reported performance, usually obtained by training and evaluating on different partitions of the same corpus, can be trusted for downstream applications.

RESULTS

Here, we report on the results of a carefully designed cross-corpus benchmark for entity recognition and normalization, where tools were applied systematically to corpora not used during their training. Based on a survey of 28 published systems, we selected five, based on predefined criteria like feature richness and availability, for an in-depth analysis on three publicly available corpora covering four entity types. Our results present a mixed picture and show that cross-corpus performance is significantly lower than the in-corpus performance. HunFlair2, the redesigned and extended successor of the HunFlair tool, showed the best performance on average, being closely followed by PubTator Central. Our results indicate that users of BTM tools should expect a lower performance than the original published one when applying tools in "the wild" and show that further research is necessary for more robust BTM tools.

AVAILABILITY AND IMPLEMENTATION

All our models are integrated into the Natural Language Processing (NLP) framework flair: https://github.com/flairNLP/flair. Code to reproduce our results is available at: https://github.com/hu-ner/hunflair2-experiments.

PEDL+: protein-centered relation extraction from PubMed at your fingertip

SUMMARY

Relation extraction (RE) from large text collections is an important tool for database curation, pathway reconstruction, or functional omics data analysis. In practice, RE often is part of a complex data analysis pipeline requiring specific adaptations like restricting the types of relations or the set of proteins to be considered. However, current systems are either non-programmable web sites or research code with fixed functionality. We present PEDL+, a user-friendly tool for extracting protein-protein and protein-chemical associations from PubMed articles. PEDL+ combines state-of-the-art NLP technology with adaptable ranking and filtering options and can easily be integrated into analysis pipelines. We evaluated PEDL+ in two pathway curation projects and found that 59% to 80% of its extractions were helpful.

AVAILABILITY AND IMPLEMENTATION

PEDL+ is freely available at https://github.com/leonweber/pedl.

Deep learning-enabled natural language processing to identify directional pharmacokinetic drug-drug interactions

BACKGROUND

During drug development, it is essential to gather information about the change of clinical exposure of a drug (object) due to the pharmacokinetic (PK) drug-drug interactions (DDIs) with another drug (precipitant). While many natural language processing (NLP) methods for DDI have been published, most were designed to evaluate if (and what kind of) DDI relationships exist in the text, without identifying the direction of DDI (object vs. precipitant drug). Here we present a method for the automatic identification of the directionality of a PK DDI from literature or drug labels.

METHODS

We reannotated the Text Analysis Conference (TAC) DDI track 2019 corpus for identifying the direction of a PK DDI and evaluated the performance of a fine-tuned BioBERT model on this task by following the training and validation steps prespecified by TAC.

RESULTS

This initial attempt showed the model achieved an F-score of 0.82 in identifying sentences as containing PK DDI and an F-score of 0.97 in identifying object versus precipitant drugs in those sentences.

DISCUSSION AND CONCLUSION

Despite a growing list of NLP methods for DDI extraction, most of them use a common set of corpora to perform general purpose tasks (e.g., classifying a sentence into one of several fixed DDI categories). There is a lack of coordination between the drug development and biomedical informatics method development community to develop corpora and methods to perform specific tasks (e.g., extract clinical exposure changes due to PK DDI). We hope that our effort can encourage such a coordination so that more "fit for purpose" NLP methods could be developed and used to facilitate the drug development process.

Not so weak PICO: leveraging weak supervision for participants, interventions, and outcomes recognition for systematic review automation

Objective

The aim of this study was to test the feasibility of PICO (participants, interventions, comparators, outcomes) entity extraction using weak supervision and natural language processing.

Methodology

We re-purpose more than 127 medical and nonmedical ontologies and expert-generated rules to obtain multiple noisy labels for PICO entities in the evidence-based medicine (EBM)-PICO corpus. These noisy labels are aggregated using simple majority voting and generative modeling to get consensus labels. The resulting probabilistic labels are used as weak signals to train a weakly supervised (WS) discriminative model and observe performance changes. We explore mistakes in the EBM-PICO that could have led to inaccurate evaluation of previous automation methods.

Results

In total, 4081 randomized clinical trials were weakly labeled to train the WS models and compared against full supervision. The models were separately trained for PICO entities and evaluated on the EBM-PICO test set. A WS approach combining ontologies and expert-generated rules outperformed full supervision for the participant entity by 1.71% macro-F1. Error analysis on the EBM-PICO subset revealed 18-23% erroneous token classifications.

Discussion

Automatic PICO entity extraction accelerates the writing of clinical systematic reviews that commonly use PICO information to filter health evidence. However, PICO extends to more entities-PICOS (S-study type and design), PICOC (C-context), and PICOT (T-timeframe) for which labelled datasets are unavailable. In such cases, the ability to use weak supervision overcomes the expensive annotation bottleneck.

Conclusions

We show the feasibility of WS PICO entity extraction using freely available ontologies and heuristics without manually annotated data. Weak supervision has encouraging performance compared to full supervision but requires careful design to outperform it.

Chemical-protein relation extraction with ensembles of carefully tuned pretrained language models

The identification of chemical-protein interactions described in the literature is an important task with applications in drug design, precision medicine and biotechnology. Manual extraction of such relationships from the biomedical literature is costly and often prohibitively time-consuming. The BioCreative VII DrugProt shared task provides a benchmark for methods for the automated extraction of chemical-protein relations from scientific text. Here we describe our contribution to the shared task and report on the achieved results. We define the task as a relation classification problem, which we approach with pretrained transformer language models. Upon this basic architecture, we experiment with utilizing textual and embedded side information from knowledge bases as well as additional training data to improve extraction performance. We perform a comprehensive evaluation of the proposed model and the individual extensions including an extensive hyperparameter search leading to 2647 different runs. We find that ensembling and choosing the right pretrained language model are crucial for optimal performance, whereas adding additional data and embedded side information did not improve results. Our best model is based on an ensemble of 10 pretrained transformers and additional textual descriptions of chemicals taken from the Comparative Toxicogenomics Database. The model reaches an F1 score of 79.73% on the hidden DrugProt test set and achieves the first rank out of 107 submitted runs in the official evaluation. Database URL: https://github.com/leonweber/drugprot.