136879 research outputs found
Sort by
The Impacts of Furrow Diking, Terracing and Contour Cultivation on Water Conservation in Texas Agriculture
No full textProtein Multimodal Learning for Variant Effect Prediction and Protein Engineering
No full textProteins, the building blocks of life, play indispensable roles in diverse functionalities within organisms. Originated as an amino acid sequence, a protein gradually folds into a globular structure. Such 3D conformations enable proteins to carry out unique biological functions alone or with partners. The quest to comprehend proteins through integrating sequence and structure data has emerged as a paramount scientific endeavor. While promising, challenges emerge at various junctures, marked by cross-modal reasoning, heterogeneous data fusion, interpretability and knowledge transfer to downstream tasks.
To identify existing gaps in state-of-the-art protein language models (pLMs) for variant effect prediction (VEP) and motivate learning beyond sequences, we first examined the transferability from the sequence pretraining to VEP by evaluating pLMs��� awareness of sequence semantics and structure properties, in relationship to their sensitivity of variant fitness. We found strong evidence supporting that the amount of protein awareness learned in pLMs is correlated with their abilities for VEP. We also identified the harmful over-finetuning issue in sequence-only modeling and revealed a remedy by injecting structure knowledge. This motivates our first multi-modal learning framework, cross-modal pseudo-likelihood (CMPL) over both sequences and structures. CMPL decomposes the joint probability of two modalities into modeling probabilities of masked sequences and sequenceto-structure, and purposefully attunes the magnitude of structure injection. The CMPL is better and more robust than sequence-only pLMs, and is further enhanced with structure odds integrated in evaluations. We then upgraded the CMPL framework by exact probability factorization across three modalities and specialized attentions over them. A three-phase end-to-end training procedure is designed for the upgraded cross-modality learning framework in an autoregressive manner.
Leveraging pLMs��� ability for zero-shot fitness prediction, we conducted computational antibody engineering without activity labels using our tailored antibody LM. Validated over a COVID-19 antibody benchmark set, we demonstrated the effectiveness of our pipeline to discover antibody variants with improved neutralization in a vast search space at the scale of ��� 20^3
