Scoring Overview

siRNAforge uses research-backed thermodynamic metrics to rank siRNA candidates. Higher composite scores indicate better predicted efficacy.

Quick Reference

Metric	Optimal Range	What It Means
composite_score	0-100 scale, higher is better	Overall quality
asymmetry_score	≥0.65	Guide strand selection preference
gc_content	40-55%	Stability vs. accessibility balance
melting_temp_c	60-78°C	Duplex stability (nearest-neighbour Tm)
mfe	-4 to -7 kcal/mol	Secondary structure stability
duplex_stability_dg	-32 to -43 kcal/mol for a 21mer	Guide:passenger duplex ΔG

Composite Score

composite_score is a weighted sum of seven sub-scores, each normalised to [0, 1], computed once per candidate by sirnaforge.core.scoring.compute_composite:

• Thermodynamic asymmetry (weight 0.12) - Guide strand preferentially enters RISC

• GC content (weight 0.10) - Balance between stability and accessibility

• Target accessibility (weight 0.13) - mRNA region accessibility

• Empirical rules (weight 0.15) - Position-specific sequence features

• Off-target specificity (weight 0.25) - Post-screen: decays with the genuine off-target count (on-target, ortholog and repeat-mediated hits excluded), exp(-count / 10)

• Isoform coverage (weight 0.15) - Post-screen: fraction of the query gene’s protein-coding isoforms the guide hits

• Conservation (weight 0.10) - Post-screen: fraction of the non-query species handed to the aligner with an ortholog hit. That set is wider than --genome-species: species reaching the pipeline only through --genome-indices, --genome-fastas or --transcriptome-indices are screened, so they count. A species whose alignment produced nothing stays in the denominator — it cannot contribute an ortholog hit, so conservation becomes a lower bound. Dropping it instead would let a degraded run outscore the complete run it degraded from, because a conservation term that goes inactive has its weight redistributed to the surviving terms

These weights (ScoringWeights, weight_set_version = "2.0.0") sum to 1.00 and are the post-screen set. Off-target, isoform coverage and conservation cannot be evaluated until transcriptome/miRNA screening has run, so the score is computed once, after screening, not at design time. A candidate produced by the standalone sirnaforge design path (no screening) still gets a composite_score from the same scorer, with those three terms simply inactive (see below) — it is not a second, cheaper score.

Weights are renormalised over the active term set

A term is active for a candidate only when its sub-score could actually be computed:

• off_target, isoform_coverage and conservation are inactive before screening has run.

• isoform_coverage stays inactive if the query gene has no protein-coding transcript (an annotation gap, not a candidate defect).

• conservation stays inactive when no species beyond the query species was handed to the aligner — a single-species run has no evidence to compute it from.

compute_composite renormalises the remaining weights to sum to 1 before combining them, so a missing term is dropped rather than scored zero. This is why a single-species run is neither rewarded nor penalised for lacking a conservation term: the weight that would have gone to conservation is redistributed proportionally across the terms that did run, not left on the table as an implicit zero.

scored_after_screening (bool) tells you which regime produced a given row’s score, and weight_set_version records which weight set. Scores from weight_set_version 1.x (the pre-issue-#80 five-term set) are not comparable to 2.x scores — always compare candidates within one run, one weight-set version.

What top_candidates excludes

top_candidates is rebuilt after screening from the candidates that clear three gates, so it can be shorter than min(top_n, number passing). (candidates_all.csv and candidates_pass.csv are filtered on passes_filters alone, but both are written in the re-ranked order.) The gates:

1. passes_filters is PASS — a failed off-target filter is a rejection, not a low score;

2. repeat_flagged is False — a guide that saturates the query transcriptome is excluded even when screening never ran;

3. scored_after_screening is True, whenever some but not all candidates were scored after screening. A design-time score lacks the three post-screen terms and is systematically the more optimistic number, so letting it compete against post-screen neighbours would put exactly the candidates whose evidence is missing at the top. Those rows stay in candidates_all.csv with their design-time score, and the run logs an ERROR naming the count. If no candidate was scored after screening (a wholly failed or wholly pre-screen run) the list is internally consistent and this gate does not apply.

Per-term contribution columns

Each candidate carries score_asymmetry, score_gc_content, score_accessibility, score_empirical, score_off_target, score_isoform_coverage and score_conservation: the renormalised-weight × sub-score × 100 contribution of each active term. They are None/empty for any term that was inactive for that candidate, so you can see exactly which terms carried a given score rather than inferring it from the total alone.

In siRNA mode they sum to composite_score (to floating-point tolerance). In --design-mode mirna they do not, and not merely by the bonus: that mode divides by the maximum attainable bonus to keep the range at 0-100, so

composite_score = (Σ contributions + mirna_bonus × 100) / (1 + max_mirna_bonus)

with max_mirna_bonus = 0.25 under the default miRNA weights (ago-start 0.10 + position-1 pairing 0.05 + 3’ supplementary 0.10). A candidate that earns no bonus therefore reports a composite_score 20% below the sum of its own contribution columns; the columns still show the relative weight each term carried, and the divisor is the same for every row of a miRNA run (including injected dirty controls), so within-run comparisons hold.

The design-time off-target proxy is now a diagnostic only

Before this weight set, the off_target term at design time was a proxy for guide self-repetitiveness (repeated 7-mers within the guide, unrelated to alignment against a reference). That computation still runs, but it no longer feeds composite_score under any name — it survives only as the unweighted diagnostic component_scores["design_off_target_proxy"].

Asymmetry Score

The most important single predictor of siRNA efficacy.

Score	Interpretation
0.8-1.0	Excellent - strong guide strand bias
0.65-0.8	Good - likely correct strand selection
0.5-0.65	Moderate - mixed strand loading possible
<0.5	Poor - passenger strand may dominate

Research basis: Khvorova et al. (2003), Schwarz et al. (2003)

GC Content

Affects duplex stability and target accessibility.

Range	Effect
<35%	Unstable duplex, poor RISC loading
35-40%	Acceptable, monitor stability
40-55%	Optimal range
55-60%	Acceptable, may reduce accessibility
>60%	Overly stable, poor target release

Melting Temperature

Temperature at which 50% of duplexes dissociate.

• <55°C: Unstable, may dissociate prematurely

• 55-65°C: Optimal for mammalian cells

• 65-75°C: Acceptable, verify experimentally

• >75°C: May resist RISC processing

Minimum Free Energy (MFE)

Predicts secondary structure stability of the guide strand.

• >0 kcal/mol: Unstable, poor structure

• -2 to -4 kcal/mol: Minimal structure (good)

• -4 to -8 kcal/mol: Moderate structure (optimal)

• <-10 kcal/mol: Strong self-structure (may reduce activity)

Filtering Recommendations

Standard (most applications)

sirnaforge workflow GENE --gc-min 35 --gc-max 60

Stringent (publication-quality)

sirnaforge workflow GENE --gc-min 40 --gc-max 55 --top-n 30

Relaxed (difficult targets)

sirnaforge workflow GENE --gc-min 30 --gc-max 65

Output Columns

The candidates_pass.csv and candidates_all.csv files include:

Column	Description
sirna_id	Unique identifier
guide_sequence	21nt guide strand (5’→3’)
passenger_sequence	Passenger/sense strand
position	Start position in transcript
composite_score	Overall quality score
asymmetry_score	Thermodynamic asymmetry
gc_content	GC percentage
melting_temp_c	Melting temperature (°C)
mfe	Minimum free energy (kcal/mol)
duplex_stability_dg	Guide:passenger duplex ΔG (kcal/mol)
dg_5p / dg_3p	Terminal 7 bp ΔG at each duplex end
delta_dg_end	dg_5p - dg_3p; positive favours guide loading
off_target_screened	False means the screen was incomplete, so the hit counts are a lower bound, not a total
off_target_count	Genuine off-target hits only (on-target, ortholog and repeat-mediated hits excluded)
on_target_hits / ortholog_hits / repeat_hits	The other three classes from the same four-way split
ortholog_species	Comma-joined canonical species names with at least one ortholog hit
repeat_flagged / repeat_transcript_fraction	Design-time k-mer repeat verdict and the frequency it was based on
isoform_coverage / conservation_score	Post-screen sub-scores (empty when inactive for that candidate)
score_*	Per-term contribution to composite_score (see Composite Score above)
scored_after_screening / weight_set_version	Which scoring regime produced this row’s composite_score
passes_filters	PASS or the first failed filter

References

1. Khvorova A et al. (2003) - Thermodynamic asymmetry and RISC loading

2. Schwarz DS et al. (2003) - Asymmetry rule for siRNA strand selection

3. Reynolds A et al. (2004) - Rational siRNA design guidelines

4. Ui-Tei K et al. (2004) - Guidelines for effective siRNAs