miRNA Target Prediction

Review on existing algorithms/methods

Outline

• Motivation
• Review of some successful existing algorithms
• miRDB (MirTarget v1 & v2)
• TarBase (DIANA-microT)
• RNA22 (unfinished)
• Some thoughts

​

Motivation

• Identify a reliable/trustable miRNA-mRNA interaction is critical for miRNA functional analysis
• So far, there is no any existing tool can solve the false positive issue.
• We have worked on miRNA more than one year, but we still use the existing tools even we already knew that it certainly causes many problems.

MirTarget

• Developed in 2006, and updated to v2 and v3 in 2007 and 2014, respectively. (v3 is not published yet)
• All results are presented on miRDB
• Several independent reviews/studies rated miRDB as the best computation tool in this field.

MirTarget v1

MirTarget v1, 2006

• miRNA’s seed region: 2-8;
• for each seed, scan all human 3’-UTR sequences
• find perfect base pairing region;
• check other species, must >=3 species;
• calculate MFE of miRNA/target site duplex
• MFE <= -13
• they did not mention why used “-13”
• Local alignment between miRNA and 3’-UTR
• mature miRNA sequence
• 22bp downstream of binding site in 3’-UTR
• “Bases surrounding the seed sequences are important for target recognition. Thus limited seed extension was evaluated for pairing to miRNA positions 1, 9 and 10.”
• “The longest stretch of perfect matches (including positions 2–8) was considered as an extended seed for raw score calculation”.

MirTarget v1, 2006

• terminal matches
• from the TargetScan paper, the terminal position tend to be A/U
• if the potential target and miRNA meet this condition, then increase the score by 1.

MirTarget v1, 2006

• Score:

Wi: seed conservation > limited seed extension > duplex binding stability > terminal base match

Cutoff: score >= 30

​

• False positive control:

shuffled miRNA 100 times and predicted targets on shuffled sequences

e.g. miR-124 target prediction:

85 gene targets were predicted (score >30) for real miR-124;

1.7 predicted targets for one shuffled miR-124 sequence on average.

MirTarget v2, 2007

• Upgrade:
• Integrated miRNA and mRNA expression info
• Used SVM to predict the possible targets
• identified more important features

MirTarget v2, 2007

• Used two sets of microarray data to identify down-regulated genes.
• Normal condition vs HCT116 Dicer^ex5 and DLD-1 Dicer^ex5 cell lines w/ miRNAs
• 6 miRNAs: let-7c, miR-103, miR-106b, miR-141, miR-15a and miR-215.
• 6 unique seed regions to reduce the bias of any specific miRNA sequence
• down-regulated genes: expression level was reduced by at least 40% with P-value < 0.001 in either cell line
• unaffected (normal) genes: expression level was at least 95%, but no more than 120% with P-value > 0.3 in both cell lines.
• 1,401 down-regulated genes and 16,761 normal genes were identified.

​

*However, specific miRNA-mRNA interactions are still unknown. All they achieved here is only a set of mRNAs that are regulated by 6 miRNAs.

​

MirTarget v2, 2007

• After identify dr and normal genes, they retrieved their 3’-UTR from NCBI and conducted further analysis.
• 4 major types of seed pairing:
• seed 8: perfect complementarity on p1-8
• seed 7a: perfect complementarity on p1-7
• seed 7b: perfect complementarity on p2-8
• seed 6: perfect complementarity on p2-7

* In this method, they ignored the sequences that were predicted multiple binding sites.

MirTarget v2, 2007

• Identification of training features
• human 3’-UTR in mouse, rat, dog, chicken
• terminal base: p1, p9, p10
• MFE
• dinucleotide counts
• single nucleotide counts
• position-specific base counts:
• ​

MirTarget v3, 2014

• According to the latest paper of miRDB, the authors have upgraded the algorithm to version 3, which employed CLASH data as additional training data.

DIANA-microT-CDS, 2012

Highlights:

• integrate CLIP data to provide high-confidence miRNA binding site on the target genes.
• PAR-CLIP: true binding sites.
• not only predicted binding site on 3’-UTR region, but also on CDS region.
• However, Dr. Wang (miRDB, 2014) suggested that the binding site on CDS region does not significantly repress gene expression.

DIANA-microT-CDS, 2012

• MREs specified by the PAR-CLIP data are divided in two categories according to the genomic region:
• 5075 MREs in 3’-UTRs
• 6057 MREs in CDSs.
• Feature extraction:
• Alignment for putative MRE identification
• miRNA extended seed sequence [p1–p9] and every 9 nt window on the 3’-UTR or CDS
• minimum: 4nt pairing
• 1 G:U is allowed if >=6nt pairing
• 1 bulge/mismatch if >= 8nt pairing
• Detection of binding categories
• 64 different binding categories
• ‘8mer+3UTR-pairing 1st:mismatch+NotA’
• Conservation measure of the MRE sequence in CDS/3’-UTR
• CDS: 30;
• 3’UTR: 16.
• Detection of significantly accessible locations within MREs
• Sfold: 40 nt upstream and 10 nt downstream of the start of each MRE
• flanking AU content; distance to closest 3 -UTR end; adjacent MRE distance; MFE; etc

DIANA-microT-CDS, 2012

• Feature selection
• used the stepAIC to find a model with minimum AIC
• independently feature selection of CDS and 3’-UTR
• Training and Scoring
• Generalized linear model to calculate the score.
• for each gene, linear combining CDS and 3’-UTR score

​

RNA22, 2006 (unfinished)

Highlights:

• Thinking out of the box
• Pure engineering style
• Signal processing
• Pattern recognition
• IBM Watson Research Center (Now: Thomas Jefferson University)
• Unfinished review.

​

​

RNA22, 2006

• 644 mature microRNA sequences
• remove identical and near-duplicate entries
• no two sequences have more than 90% similarity
• 354 mature microRNA sequences passed the filtering
• “Teiresias” algorithm was used to discover the patterns among the 354 mature miRNA sequences
• unsupervised manner
• eg: [AT][CG].TTTTT[CG]G..[AT]
• 233,554 mature-microRNA patterns remained

​

​

Thank you!!!