Overview #
Computational predictive tools are essential for variant interpretation, providing quantitative evidence on how genetic changes may affect gene or protein function.
Within the ACMG/AMP framework:
- PP3 supports pathogenicity,
- BP4 supports benignity, and
- BP7 applies to synonymous variants with no predicted impact on splicing.
While early ACMG guidelines did not specify which computational tools or thresholds to use, subsequent refinements by the ClinGen SVI Working Group introduced standardized approaches for meta-predictors and score-based evidence strength.
SeqSMART builds upon these principles with a proprietary AI-driven missense prediction model — the SeqSMART Missense Score — trained on over 300,000 curated missense variants. This model integrates gene-specific biological, biochemical, and computational data, enabling highly accurate PP3/BP4 assignments.
1. SeqSMART Missense Score #
1.1 Model Overview #
The SeqSMART Missense Score provides a unified, data-rich probability between 0 and 20, representing the likelihood that a missense variant disrupts normal gene function.
The model was trained using:
- 300,000 curated ClinVar and expert-reviewed missense variants,
- Gene-level and domain-specific features (e.g., UniProt functional regions, active sites, transmembrane domains),
- Biochemical substitution metrics (charge, polarity, hydrophobicity shifts),
- Aggregated results from multiple computational predictors (CADD, PolyPhen, SIFT, PROVEAN, conservation metrics, etc.), and
- Contextualized gene behavior, weighting evidence based on constraint scores, inheritance mode, and known variant mechanisms.
This integrative approach allows SeqSMART to predict variant impact in gene-specific biological context, rather than through generalized thresholds alone.
1.2 Score-to-Evidence Mapping #
The SeqSMART Missense Score outputs a continuous value between 0 and 1.
The resulting score is automatically translated into ACMG evidence strength for PP3 or BP4 using the following standardized thresholds:
| SeqSMART Missense Score | Interpretation |
| ≥ 18.64 | PP3 – Strong |
| 18.64 – 15.46 | PP3 – Moderate |
| 15.46 – 12.88 | PP3 – Supporting |
| 5.8 – 3.66 | BP4 – Supporting |
| 3.66 – 0.32 | BP4 – Moderate |
| 0.32 – 0.06 | BP4 – Strong |
| < 0.06 | BP4 – Very Strong |
This mapping aligns with the ClinGen evidence-weighting structure but leverages SeqSMART’s enhanced model precision and expanded feature space.
2. SpliceAI Integration #
After evaluating the Missense Score, SeqSMART automatically checks SpliceAI outputs for potential splice alterations.
- If any SpliceAI score ≥ 0.5 → PP3 (Supporting) is applied.
- If all SpliceAI scores < 0.5 → The variant supports BP4 (if not already assigned).
This ensures splicing-related evidence is incorporated seamlessly alongside missense impact prediction.
3. Variants Without Missense Scores #
For non-missense variants, SeqSMART first evaluates potential splicing effects using SpliceAI.
- If SpliceAI ≥ 0.5, the variant meets PP3 – Supporting (suggesting a likely splice impact).
- If SpliceAI < 0.5, the system proceeds to evaluate CADD scores according to the following calibrated thresholds:
| CADD Score Range | Interpretation |
| ≥ 28.1 | PP3 – Moderate |
| 28.1 – 25.3 | PP3 – Supporting |
| 22.7 – 17.3 | BP4 – Supporting |
| 17.3 – 0.15 | BP4 – Moderate |
| < 0.15 | BP4 – Strong |
This multi-layered decision framework maintains analytical consistency even when one or more prediction tools lack data coverage, ensuring that every variant receives a balanced and evidence-based computational evaluation.
4. BP7 – Synonymous Variants and Splicing Predictions #
BP7 applies to synonymous variants predicted to have no impact on splicing.
SeqSMART uses SpliceAI to make this determination:
- If all SpliceAI scores < 0.5, BP7 is considered met.
- If any score ≥ 0.5, BP7 is unmet.
To avoid redundant evidence counting, if BP4 was already assigned based on SpliceAI data, BP7 is still computed but excluded from the weighted ACMG scoring.
This maintains balanced computational evidence while preserving transparency.
5. Transparency and Expert Review #
All computational evidence is fully accessible within the variant detail view:
- SeqSMART Missense Score (with numerical output and evidence tier)
- SpliceAI probabilities (acceptor/donor loss/gain)
- CADD score and percentile rank
Expert reviewers can:
- Inspect tool outputs and their thresholds.
- Adjust or override PP3, BP4, or BP7 if new evidence arises.
- Record notes or institutional interpretations directly in the variant record.
All modifications are logged with reviewer identity, timestamp, and rationale to ensure traceable reproducibility.
6. Why the SeqSMART Missense Score Matters #
Traditional tools (e.g., REVEL, MutPred2) apply global weights without accounting for gene context or variant type diversity.
SeqSMART’s AI-driven score improves upon this by:
- Incorporating gene-specific functional architecture,
- Recognizing biochemical substitution effects,
- Integrating constraint, conservation, and domain-level data, and
- Continuously retraining on newly validated variants.
This enables the SeqSMART platform to deliver more precise and biologically grounded predictions than static meta-predictors.
Summary #
SeqSMART replaces external meta-predictors with an advanced AI-based Missense Score that encapsulates structural, biochemical, and gene-level context into a single interpretable metric.
Together with SpliceAI and CADD integration, this system ensures that computational evidence for PP3, BP4, and BP7 is accurate, consistent, and transparent.
SeqSMART Predictive Logic Principle:
Artificial Intelligence meets biological intuition — one score, all evidence.