This thesis addresses the language barrier in life science NLP by developing TransBERT, a French biomedical language model trained entirely on over 22 million machine-translated abstracts. Leveraging advances in machine translation, the study demonstrates that TransBERT achieves competitive or superior results compared to state-of-the-art French models, even without access to extensive native corpora. Additionally, the research shows that domain-specific tokenization significantly boosts performance, particularly for named entity recognition tasks. The work introduces TransCorpus, the largest French life science corpus to date, and provides a robust evaluation framework for model comparison. These findings highlight scalable strategies for building high-quality language models in low-resource language/domain pairs, paving the way for broader access to advanced NLP tools in virtually any domain or language.