Machine learning for functional protein design

Nature Biotechnology, Journal Year: 2024, Volume and Issue: 42(2), P. 216 - 228

Published: Feb. 1, 2024

Language: Английский

---

Simulating 500 million years of evolution with a language model

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: July 2, 2024

Abstract More than three billion years of evolution have produced an image biology encoded into the space natural proteins. Here we show that language models trained on tokens generated by can act as evolutionary simulators to generate functional proteins are far away from known We present ESM3, a frontier multimodal generative model reasons over sequence, structure, and function ESM3 follow complex prompts combining its modalities is highly responsive biological alignment. prompted fluorescent with chain thought. Among generations synthesized, found bright protein at distance (58% identity) Similarly distant separated five hundred million evolution.

Language: Английский

---

Simulating 500 million years of evolution with a language model

Science, Journal Year: 2025, Volume and Issue: unknown

Published: Jan. 16, 2025

More than three billion years of evolution have produced an image biology encoded into the space natural proteins. Here we show that language models trained at scale on evolutionary data can generate functional proteins are far away from known We present ESM3, a frontier multimodal generative model reasons over sequence, structure, and function ESM3 follow complex prompts combining its modalities is highly responsive to alignment improve fidelity. prompted fluorescent Among generations synthesized, found bright protein distance (58% sequence identity) proteins, which estimate equivalent simulating five hundred million evolution.

Language: Английский

---

Sparks of function by de novo protein design

Nature Biotechnology, Journal Year: 2024, Volume and Issue: 42(2), P. 203 - 215

Published: Feb. 1, 2024

Language: Английский

---

Feature Reuse and Scaling: Understanding Transfer Learning with Protein Language Models

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: Feb. 8, 2024

Abstract Large pretrained protein language models (PLMs) have improved property and structure prediction from sequences via transfer learning, in which weights representations PLMs are repurposed for downstream tasks. Although shown great promise, currently there is little understanding of how the features learned by pretraining relate to useful We perform a systematic analysis learning using PLMs, conducting 370 experiments across comprehensive suite factors including different tasks, architectures, model sizes, depths, time. observe that while almost all down-stream tasks do benefit compared naive sequence representations, majority performance does not scale with pretraining, instead relies on low-level early pretraining. Our results point mismatch between current PLM paradigms most applications these models, indicating need better methods.

Language: Английский

---

Enhancing efficiency of protein language models with minimal wet-lab data through few-shot learning

Nature Communications, Journal Year: 2024, Volume and Issue: 15(1)

Published: July 2, 2024

Accurately modeling the protein fitness landscapes holds great importance for engineering. Pre-trained language models have achieved state-of-the-art performance in predicting without wet-lab experimental data, but their accuracy and interpretability remain limited. On other hand, traditional supervised deep learning require abundant labeled training examples improvements, posing a practical barrier. In this work, we introduce FSFP, strategy that can effectively optimize under extreme data scarcity prediction. By combining meta-transfer learning, to rank, parameter-efficient fine-tuning, FSFP significantly boost of various using merely tens single-site mutants from target protein. silico benchmarks across 87 mutational scanning datasets demonstrate FSFP's superiority over both unsupervised baselines. Furthermore, successfully apply engineer Phi29 DNA polymerase through experiments, achieving 25% increase positive rate. These results underscore potential our approach aiding AI-guided

Language: Английский

---

PTM-Mamba: a PTM-aware protein language model with bidirectional gated Mamba blocks

Nature Methods, Journal Year: 2025, Volume and Issue: unknown

Published: April 10, 2025

Language: Английский

---

Generative artificial intelligence for de novo protein design

Current Opinion in Structural Biology, Journal Year: 2024, Volume and Issue: 86, P. 102794 - 102794

Published: April 24, 2024

Engineering new molecules with desirable functions and properties has the potential to extend our ability engineer proteins beyond what nature so far evolved. Advances in so-called 'de novo' design problem have recently been brought forward by developments artificial intelligence. Generative architectures, such as language models diffusion processes, seem adept at generating novel, yet realistic that display perform specified functions. State-of-the-art protocols now achieve experimental success rates nearing 20%, thus widening access de novo designed proteins. Despite extensive progress, there are clear field-wide challenges, for example, determining best silico metrics prioritise designs testing, designing can undergo large conformational changes or be regulated post-translational modifications. With an increase number of being developed, this review provides a framework understand how these tools fit into overall process protein design. Throughout, we highlight power incorporating biochemical knowledge improve performance interpretability.

Language: Английский

---

Zero-shot prediction of mutation effects with multimodal deep representation learning guides protein engineering

Cell Research, Journal Year: 2024, Volume and Issue: 34(9), P. 630 - 647

Published: July 5, 2024

Abstract Mutations in amino acid sequences can provoke changes protein function. Accurate and unsupervised prediction of mutation effects is critical biotechnology biomedicine, but remains a fundamental challenge. To resolve this challenge, here we present Pro tein M utational E ffect P redictor (ProMEP), general multiple sequence alignment-free method that enables zero-shot effects. A multimodal deep representation learning model embedded ProMEP was developed to comprehensively learn both structure contexts from ~160 million proteins. achieves state-of-the-art performance mutational effect accomplishes tremendous improvement speed, enabling efficient intelligent engineering. Specifically, accurately forecasts consequences on the gene-editing enzymes TnpB TadA, successfully guides development high-performance tools with their engineered variants. The efficiency 5-site mutant reaches up 74.04% (vs 24.66% for wild type); base editing tool basis TadA 15-site (in addition A106V/D108N double renders deoxyadenosine deaminase activity TadA) exhibits an A-to-G conversion frequency 77.27% 69.80% ABE8e, previous TadA-based adenine editor) significantly reduced bystander off-target compared ABE8e. not only showcases superior predicting proteins also demonstrates great capability guide Therefore, exploration gigantic space facilitates practical design proteins, thereby advancing studies biomedicine synthetic biology.

Language: Английский

---

Multi-modal deep learning enables efficient and accurate annotation of enzymatic active sites

Nature Communications, Journal Year: 2024, Volume and Issue: 15(1)

Published: Aug. 26, 2024

Annotating active sites in enzymes is crucial for advancing multiple fields including drug discovery, disease research, enzyme engineering, and synthetic biology. Despite the development of numerous automated annotation algorithms, a significant trade-off between speed accuracy limits their large-scale practical applications. We introduce EasIFA, an site algorithm that fuses latent representations from Protein Language Model 3D structural encoder, then aligns protein-level information with knowledge enzymatic reactions using multi-modal cross-attention framework. EasIFA outperforms BLASTp 10-fold increase improved recall, precision, f1 score, MCC by 7.57%, 13.08%, 9.68%, 0.1012, respectively. It also surpasses empirical-rule-based other state-of-the-art deep learning method based on PSSM features, achieving ranging 650 to 1400 times while enhancing quality. This makes suitable replacement conventional tools both industrial academic settings. can effectively transfer gained coarsely annotated databases smaller, high-precision datasets, highlighting its ability model sparse high-quality databases. Additionally, shows potential as catalytic monitoring tool designing desired functions beyond natural distribution. Wang et al. propose efficient algorithm, advance various

Language: Английский

---