Bio Small Rna Seq Target Prediction

Version Compatibility

Reference examples tested with: miRanda 3.3a+, BioPython 1.83+, pandas 2.2+, gseapy 1.1+

Before using code patterns, verify installed versions match. If versions differ:

• CLI: <tool> --version then <tool> --help to confirm flags
• Python: pip show <package> then help(module.function) to check signatures

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

miRNA Target Prediction

"Predict target genes for my miRNAs" -> Generate candidate mRNA targets by seed complementarity and thermodynamics, then raise confidence with conservation, validated databases, and matched expression.

• CLI: miranda miRNA.fa UTR.fa -sc 140 -en -20 -strict for de novo prediction
• Python: query TargetScan/miRDB downloads and miRTarBase for validated interactions

The governing principle: a predicted target is a hypothesis, not a finding

Seed-based prediction has a false-positive rate near 50% even for conserved sites (Pinzon 2017), because a 6mer seed match occurs by chance roughly once per 4 kb of sequence, so a multi-kb 3' UTR carries many spurious matches. It also has a RECALL problem in the opposite direction: AGO-CLIP and CLASH show that roughly 60% of real interactions are noncanonical, and the bulged, seedless, and 3'-compensatory sites among them are missed entirely by seed-only tools (3'-supplementary sites keep a canonical seed and are still found), so a clean seed list is incomplete, not just imprecise (Helwak 2013). Worse, a single miRNA represses most of its real targets only modestly - typically less than two-fold at the protein level (Baek 2008; Selbach 2008) - so miRNAs are rheostats, not switches, and large single-target claims should be distrusted. The decisive move is therefore not running more predictors (five seed-based tools agreeing is pseudo-replication, not independent evidence) but climbing an evidence ladder and, above all, intersecting predictions with INVERSELY-correlated differentially expressed mRNA or protein from the SAME samples. Prediction proposes; expression disposes.

Decision: which predictor, and how it scores

De novo prediction with miRanda

Goal: Predict miRNA-mRNA target sites by complementarity and duplex energy.

Approach: Align miRNA sequences against 3' UTRs with a minimum score and a maximum (negative) energy, requiring strict seed pairing.

miranda miRNA.fa UTRs.fa -sc 140 -en -20 -strict -out predictions.txt

# -sc 140: minimum alignment score (keep alignments with score >= 140; default 140)
# -en -20: maximum free energy in kcal/mol (keep energies <= -20; value is negative)
# -strict: require canonical seed pairing at positions 2-8 (no gaps/wobble in seed)
# Tunable: many use -sc 150 / -en -7 (looser) up to -sc 155 / -en -20 (stringent)

Parse miRanda output

Goal: Extract interaction records into a DataFrame.

Approach: Read the lines miRanda prefixes with '>' (per-hit summary) and pull miRNA, target, score, and energy.

import pandas as pd

def parse_miranda(output_file):
    rows = []
    with open(output_file) as f:
        for line in f:
            if line.startswith('>') and not line.startswith('>>'):
                p = line.strip().split('\t')
                if len(p) >= 5:
                    rows.append({'mirna': p[0].lstrip('>'), 'target': p[1],
                                 'score': float(p[2]), 'energy': float(p[3])})
    return pd.DataFrame(rows)

TargetScan context-scores lookup

Goal: Retrieve conserved-site predictions and rank a miRNA's targets.

Approach: Read the downloadable per-site context-scores file and rank by the weighted context++ score (more negative = stronger predicted repression across the whole UTR).

import pandas as pd

def query_targetscan(mirbase_id, ts_file='Predicted_Targets_Context_Scores.default_predictions.txt'):
    # Verified column names: the miRNA column is 'Mirbase ID' (NOT 'miRNA family'),
    # the gene column is 'Gene ID', and 'weighted context++ score' aggregates a UTR's sites.
    df = pd.read_csv(ts_file, sep='\t')
    hits = df[df['Mirbase ID'] == mirbase_id]
    return hits.sort_values('weighted context++ score')   # ascending: most negative first

miRDB (machine-learning) lookup

Goal: Retrieve ML-based target predictions above the conventional confidence cut.

Approach: Read the miRDB prediction download and keep targets with score >= 80.

def query_mirdb(mirna_id, mirdb_file='miRDB_v6.0_prediction_result.txt'):
    df = pd.read_csv(mirdb_file, sep='\t', header=None, names=['mirna', 'refseq', 'score'])
    hits = df[df['mirna'] == mirna_id]
    return hits[hits['score'] >= 80].sort_values('score', ascending=False)

Validated targets and unified lookup

Goal: Anchor target claims in experimental evidence rather than prediction.

Approach: Query miRTarBase for validated interactions and weight by evidence type; use multiMiR (R) to unify predicted and validated sources in one call.

def get_validated_targets(mirna, mirtarbase_file='miRTarBase_MTI.xlsx'):
    df = pd.read_excel(mirtarbase_file)
    hits = df[df['miRNA'] == mirna]
    # 'Support Type' separates strong (reporter/western/qPCR) from less-strong (CLIP/NGS)
    return hits[['Target Gene', 'Experiments', 'Support Type']]

The confidence move: intersect with anti-correlated mRNA DE

Goal: Keep only targets that behave functionally in the actual experiment.

Approach: Intersect predicted (or CLIP-supported) targets of UP miRNAs with DOWN mRNAs from matched samples; note the blind spot that translation-only targets may not move at the mRNA level.

def functional_targets(predicted_targets, mrna_de, mirna_direction):
    # mrna_de: DataFrame with index = gene, column 'log2FC' from matched mRNA-seq.
    # Anti-correlation: an UP miRNA should repress -> targets DOWN (and vice versa).
    # Blind spot: miRNAs also act translationally, so some real targets stay flat at
    # the mRNA level (need ribosome profiling / proteomics to see those).
    want_down = mirna_direction == 'up'
    moved = mrna_de[(mrna_de['log2FC'] < 0) == want_down].index
    return [g for g in predicted_targets if g in set(moved)]