Exscien (United States)

Abstract Accuratein-silicoprediction of protein-ligand binding affinity is essential for efficient hit identification in large molecular libraries. Commonly used structure-based methods such as docking often fail to rank compounds effectively, and free energy-based approaches, while accurate, are too computationally intensive for large-scale screening. Existing deep learning models struggle to generalize to new targets or drugs, and current evaluation methods often do not accurately reflect real-world performance. We introduceBALM, a deep learning framework that predictsbindingaffinity using pre-trained protein and ligandlanguagemodels. We also propose improved evaluation strategies with diverse data sets and metrics to assess model performance to new targets better. Using the BindingDB dataset, BALM generalises unseen drugs, scaffolds, and targets. In few-shot scenarios for targets such asUSP7andMpro, it outperforms traditional machine learning and docking methods, including AutoDock Vina. Adoption of our target-based evaluation methods will allow a more stringent evaluation of machine learning-based scoring tools. Our protein prediction framework shows good performance, is computationally efficient, and is highly adaptable within this evaluation setting, making it practical for early-stage drug discovery screening.

Abstract Protein representations from deep language models have yielded state-of-the-art performance across many tasks in computational protein engineering. In recent years, progress has primarily focused on parameter count, with recent models’ capacities surpassing the size of the very datasets they were trained on. Here, we propose an alternative direction. We show that large language models trained on codons, instead of amino acid sequences, provide high-quality representations that outperform comparable state-of-the-art models across a variety of tasks. In some tasks, like species recognition, prediction of protein and transcript abundance, or melting point estimation, we show that a language model trained on codons outperforms every other published protein language model, including some that contain over 50 times more parameters. These results suggest that, in addition to commonly studied scale and model complexity, the information content of biological data provides an orthogonal direction to improve the power of machine learning in biology.

Abstract Breast cancer is a global health issue affecting 2.3 million women per year, causing death in over 600,000. Mammography (and biopsy) is the gold standard for screening and diagnosis. Whilst effective, this test exposes individuals to radiation, has limitations to its sensitivity and specificity and may cause moderate to severe discomfort. Some women may also find this test culturally unacceptable. This proof-of-concept study, combining bottom-up proteomics with Matrix Assisted Laser Desorption Ionisation Mass Spectrometry (MALDI MS) detection, explores the potential for a non-invasive technique for the early detection of breast cancer from fingertip smears. A cohort of 15 women with either benign breast disease (n = 5), early breast cancer (n = 5) or metastatic breast cancer (n = 5) were recruited from a single UK breast unit. Fingertips smears were taken from each patient and from each of the ten digits, either at the time of diagnosis or, for metastatic patients, during active treatment. A number of statistical analyses and machine learning approaches were investigated and applied to the resulting mass spectral dataset. The highest performing predictive method, a 3-class Multilayer Perceptron (MLP) neural network, yielded an accuracy score of 97.8% when categorising unseen MALDI MS spectra as either the benign, early or metastatic cancer classes. These findings support the need for further research into the use of sweat deposits (in the form of fingertip smears or fingerprints) for non-invasive screening of breast cancer.