Comparison between parameter-efficient techniques and full fine-tuning: A case study on multilingual news article classification.

Adapters and Low-Rank Adaptation (LoRA) are parameter-efficient fine-tuning techniques designed to make the training of language models more efficient. Previous results demonstrated that these methods can even improve performance on some classification tasks. This paper complements existing research by investigating how these techniques influence classification performance and computation costs compared to full fine-tuning. We focus specifically on multilingual text classification tasks (genre, framing, and persuasion techniques detection; with different input lengths, number of predicted classes and classification difficulty), some of which have limited training data. In addition, we conduct in-depth analyses of their efficacy across different training scenarios (training on the original multilingual data; on the translations into English; and on a subset of English-only data) and different languages. Our findings provide valuable insights into the applicability of parameter-efficient fine-tuning techniques, particularly for multilabel classification and non-parallel multilingual tasks which are aimed at analysing input texts of varying length.

The next-generation Open Targets Platform: reimagined, redesigned, rebuilt.

The Open Targets Platform (https://platform.opentargets.org/) is an open source resource to systematically assist drug target identification and prioritisation using publicly available data. Since our last update, we have reimagined, redesigned, and rebuilt the Platform in order to streamline data integration and harmonisation, expand the ways in which users can explore the data, and improve the user experience. The gene-disease causal evidence has been enhanced and expanded to better capture disease causality across rare, common, and somatic diseases. For target and drug annotations, we have incorporated new features that help assess target safety and tractability, including genetic constraint, PROTACtability assessments, and AlphaFold structure predictions. We have also introduced new machine learning applications for knowledge extraction from the published literature, clinical trial information, and drug labels. The new technologies and frameworks introduced since the last update will ease the introduction of new features and the creation of separate instances of the Platform adapted to user requirements. Our new Community forum, expanded training materials, and outreach programme support our users in a range of use cases.

Open Targets Platform: supporting systematic drug-target identification and prioritisation.

The Open Targets Platform (https://www.targetvalidation.org/) provides users with a queryable knowledgebase and user interface to aid systematic target identification and prioritisation for drug discovery based upon underlying evidence. It is publicly available and the underlying code is open source. Since our last update two years ago, we have had 10 releases to maintain and continuously improve evidence for target-disease relationships from 20 different data sources. In addition, we have integrated new evidence from key datasets, including prioritised targets identified from genome-wide CRISPR knockout screens in 300 cancer models (Project Score), and GWAS/UK BioBank statistical genetic analysis evidence from the Open Targets Genetics Portal. We have evolved our evidence scoring framework to improve target identification. To aid the prioritisation of targets and inform on the potential impact of modulating a given target, we have added evaluation of post-marketing adverse drug reactions and new curated information on target tractability and safety. We have also developed the user interface and backend technologies to improve performance and usability. In this article, we describe the latest enhancements to the Platform, to address the fundamental challenge that developing effective and safe drugs is difficult and expensive.