Generation of functional oligopeptides that promote osteogenesis based on unsupervised deep learning of protein IDRs

Mingxiang Cai,Baichuan Xiao,Yao Sun,Jiaqi Wu,Yongbiao Zhang,Fujun Jin,Xiaopeng Xu,Xiaogang Wang,Pingping Niu,Yong Zhang,Rui Yue,Yuwei Hua,Meijing Liu,Zuolin Wang,Junhui Li

doi:10.1038/s41413-022-00193-1

Abstract

Deep learning (DL) is currently revolutionizing peptide drug development due to both computational advances and the substantial recent expansion of digitized biological data. However, progress in oligopeptide drug development has been limited, likely due to the lack of suitable datasets and difficulty in identifying informative features to use as inputs for DL models. Here, we utilized an unsupervised deep learning model to learn a semantic pattern based on the intrinsically disordered regions of ~171 known osteogenic proteins. Subsequently, oligopeptides were generated from this semantic pattern based on Monte Carlo simulation, followed by in vivo functional characterization. A five amino acid oligopeptide (AIB5P) had strong bone-formation-promoting effects, as determined in multiple mouse models (e.g., osteoporosis, fracture, and osseointegration of implants). Mechanistically, we showed that AIB5P promotes osteogenesis by binding to the integrin α5 subunit and thereby activating FAK signaling. In summary, we successfully established an oligopeptide discovery strategy based on a DL model and demonstrated its utility from cytological screening to animal experimental verification.

Highlights

Peptide drugs are known to be highly selective, efficacious, and well tolerated by patients,[1] and very short peptide drugs have been attracting increasing attention because of their high bioavailability and low cost of synthesis.[2–4] A variant of the artificial intelligence method has been harnessed to substantially increase the efficiency of peptide drug development efforts[5]; these gains have been enabled by the abundant databases of available protein sequence and spatial structural information.[6]
These methods have been less impactful in the field of oligopeptide drug development due to issues including the relatively small amount of available data for oligopeptide drugs and the fact that the very short lengths of oligopeptides result in relatively few of the distinguishable features that are exploited by common machine learning approaches.[7]
A report by Stavros et al explored the interesting concept of the “no free lunch” theorem, which may yield insights that can help overcome the present shortage of available datasets for functional oligopeptides.[8]

Summary

PGESRDAQTHK NMV Y L F I WC Amino acid

The new oligopeptides obtained in each cycle serve as oligopeptides awaiting extension in the cycle until the end condition of the cycle is reached (here, until the oligopeptides have been extended to a total of 10 residues). c Protein candidates were retrieved from UniProt based on their reported involvement in bone formation by using four osteogenic GO terms: “ossification”, “osteogenesis”, “osteoblast development”, and “osteoblast differentiation”. Intravenous injection of AIB5P had no significant effect on the body weight of mice. Micro-CT scanning showed that the trabecular bone mineral density, cortical bone mineral density, and bone volume of the AIB5P treatment group were all significantly increased compared to those of the vehicle controltreated OVX model animals (Fig. 5a, b). Both bone masses of the fifth lumbar vertebrae (Fig. S3B) and the bone formation rate were significantly enhanced in the AIB5P group (Fig. 5c, d).

GGGAGGGG PPGPPGPP

Double labeling

FAK P

Functional peptide learning using a deep neural network

Findings

AUTHOR CONTRIBUTIONS