Flexible Kernels for Protein Property Prediction
Researchers have made advancements in predicting protein properties from sparse experimental data, a longstanding challenge in protein design.
A new class of sequence kernels has been introduced, leveraging evolutionary substitution matrices and local linearity to create data-efficient models of protein property landscapes[1]. These Gaussian processes often outperform alternatives relying on foundation model embeddings. The proposed kernels can incorporate structural information from foundation models, making them suitable for multi-task learning across multiple protein property landscapes. Meanwhile, a separate study evaluated MolSight, a system using 10 vision architectures and 7 pre-training strategies across 10 downstream tasks[2]. MolSight achieved top results on 5 of 10 benchmarks and ranked in the top two on all 10, with 80 times lower FLOPs than the nearest multi-modal competitor. The system uses a chemistry-informed curriculum to account for structural complexity across pre-training molecules, and a single rendered bond-line image was found sufficient for competitive molecular property prediction.
research-paper
Background sources we checked (4)
- arxiv.org ↗ Despite its importance to applications in protein design, predicting protein properties like binding affinity and thermostability from sparse experimental data remains a significant challenge. Accordingly, we introduce a class of sequence kernels that exploit evolutionary substit…
- en.wikipedia.org ↗ Protein structure prediction is the inference of the three-dimensional structure of a protein from its amino acid sequence—that is, the prediction of its secondary and tertiary structure from primary structure. Structure prediction is different from the inverse problem of protein…
- en.wikipedia.org ↗ Protein design is the rational design of new protein molecules to design novel activity, behavior, or purpose, and to advance basic understanding of protein function. Proteins can be designed from scratch (de novo design) or by making calculated variants of a known protein struct…
- en.wikipedia.org ↗ Computational methods exploit the sequence signatures of disorder to predict whether a protein is disordered, given its amino acid sequence. The table below, which was originally adapted from and has been recently updated, shows the main features of software for disorder predicti…