Biology subjects

Ke, G.

Publications and source records attributed to Ke, G..

5 recordsLinked to original sources

bioRxiv2023

Amalga: Designable Protein Backbone Generation with Folding and Inverse Folding Guidance

Recent advances in deep learning enable new approaches to protein design through inverse folding and backbone generation. However, backbone generators may produce structures that inverse folding struggles to identify sequences for, indicating designability issues. We propose Amalga, an inference-time technique that enhances designability of backbone generators. Amalga leverages folding and inverse folding models to guide backbone generation towards more designable conformations by incorporating "folded-from-inverse-folded" (FIF) structures. To generate FIF structures, possible sequences are predicted from step-wise predictions in the reverse diffusion and further folded into new backbones. Being intrinsically designable, the FIF structures guide the generated backbones to a more designable distribution. Experiments on both de novo design and motif-scaffolding demonstrate improved designability and diversity with Amalga on RFdiffusion.

bioinformatics↗

bioRxiv2023

UNI-RNA: UNIVERSAL PRE-TRAINED MODELS REVOLUTIONIZE RNA RESEARCH

AO_SCPLOWBSTRACTC_SCPLOWRNA molecules play a crucial role as intermediaries in diverse biological processes. Attaining a profound understanding of their function can substantially enhance our comprehension of lifes activities and facilitate drug development for numerous diseases. The advent of high-throughput sequencing technologies makes vast amounts of RNA sequence data accessible, which contains invaluable information and knowledge. However, deriving insights for further application from such an immense volume of data poses a significant challenge. Fortunately, recent advancements in pre-trained models have surfaced as a revolutionary solution for addressing such challenges owing to their exceptional ability to automatically mine and extract hidden knowledge from massive datasets. Inspired by the past successes, we developed a novel context-aware deep learning model named Uni-RNA that performs pre-training on the largest dataset of RNA sequences at the unprecedented scale to date. During this process, our model autonomously unraveled the obscured evolutionary and structural information embedded within the RNA sequences. As a result, through fine-tuning, our model achieved the state-of-the-art (SOTA) performances in a spectrum of downstream tasks, including both structural and functional predictions. Overall, Uni-RNA established a new research paradigm empowered by the large pre-trained model in the field of RNA, enabling the community to unlock the power of AI at a whole new level to significantly expedite the pace of research and foster groundbreaking discoveries.

bioinformatics↗

bioRxiv2023

Uni-Fold MuSSe: De Novo Protein Complex Prediction with Protein Language Models

AO_SCPLOWBSTRACTC_SCPLOWAccurately solving the structures of protein complexes is crucial for understanding and further modifying biological activities. Recent success of AlphaFold and its variants shows that deep learning models are capable of accurately predicting protein complex structures, yet with the painstaking effort of homology search and pairing. To bypass this need, we present Uni-Fold MuSSe (Multimer with Single Sequence inputs), which predicts protein complex structures from their primary sequences with the aid of pre-trained protein language models. Specifically, we built protein complex prediction models based on the protein sequence representations of ESM-2, a large protein language model with 3 billion parameters. In order to adapt the language model to inter-protein evolutionary patterns, we slightly modified and further pre-trained the language model on groups of protein sequences with known interactions. Our results highlight the potential of protein language models for complex prediction and suggest room for improvements.

bioinformatics↗

bioRxiv2022

Uni-Fold Symmetry: Harnessing Symmetry in Folding Large Protein Complexes

Deep folding models have revolutionized the conventional methods of protein complex prediction. However, applying them to large protein oligomers is not easy. These models generally require copying the sequences of identical subunits to capture the in-between relationships. Accordingly, the scales of target protein complexes are strictly limited due to the cubic complexity of these models. To address this issue, we propose UF-Symmetry (Uni-Fold Symmetry), which is extricated from the need of sequence copying via harnessing the intrinsic symmetry of large protein oligomers. Taking the sequences of the asymmetric unit (AU) and a pre-specified symmetry group, UF-Symmetry learns to fold the AU and to assemble the complex structure in an end-to-end manner. By reducing the input scales from entire assemblies to AUs, UF-Symmetry allows to predict much larger assemblies with significant acceleration: for a complex of 4-fold cyclic symmetry (C4) and AU size of 512, UF-Symmetry achieves approximately 20 times acceleration to current methods. On a benchmark of recently released PDB multimers, UF-Symmetry approximately halves the failure rate of current methods and achieves approaching accuracy on commonly successful cases.

bioinformatics↗

bioRxiv2022

Uni-Fold: An Open-Source Platform for Developing Protein Folding Models beyond AlphaFold

Recent breakthroughs on protein structure prediction, namely AlphaFold, have led to unprecedented new possibilities in related areas. However, the lack of training utilities in its current open-source code hinders the community from further developing or adapting the model. Here we present Uni-Fold as a thoroughly open-source platform for developing protein folding models beyond AlphaFold. We reimplemented AlphaFold and AlphaFold-Multimer in the PyTorch framework, and reproduced their from-scratch training processes with equivalent or better accuracy. Based on various optimizations, Uni-Fold achieves about 2.2 times training acceleration compared with AlphaFold under similar hardware configuration. On a benchmark of recently released multimeric protein structures, Uni-Fold outperforms AlphaFold-Multimer by approximately 2% on the TM-Score. Uni-Fold is currently the only open-source repository that supports both training and inference of multimeric protein models. The source code, model parameters, test data, and web server of Uni-Fold are publicly available3.

bioinformatics↗