Researchers from the Magami Open Sciences Initiative have conducted a rigorous evaluation of five open-source large language models for malaria drug discovery. They created a dataset called Malaria-Instruct, which is used to test the performance of these models in virtual screening tasks. The results show that fine-tuned versions of these models outperform classical machine learning models and even some proprietary models from companies like OpenAI. However, the researchers found that domain-specific fine-tuning is crucial for achieving good results. They also discovered that biomedical pretraining can provide a measurable advantage over other types of pretraining. The study suggests that open-source language models could be a resource-efficient alternative to classical pipelines and proprietary models in antimalarial virtual screening tasks. --- Why it matters: This research matters because it provides insights into the performance of large language models in a specific domain, malaria drug discovery. It highlights the importance of fine-tuning these models for optimal results and suggests that open-source models can be a viable alternative to more expensive proprietary solutions. Source: https://arxiv.org/abs/2608.20418