Meta-Layer Documentation

Last Updated: December 2025
Status: Active Development

---

Overview

This directory contains documentation for the meta-learning layer that recalibrates base model splice site predictions to improve variant effect detection.

---

Document Index

Core Architecture

Document	Description	Status
ARCHITECTURE.md	System architecture and design	✅
LABELING_STRATEGY.md	Label derivation from SpliceVarDB	✅
DATA_FORMAT_AND_LEAKAGE.md	Data format and avoiding leakage	✅
TRAINING_GUIDE.md	Step-by-step training instructions	✅

Methodology

Document	Description	Status
methods/README.md	Method taxonomy overview	✅
methods/ROADMAP.md	Development roadmap	✅
methods/PAIRED_DELTA_PREDICTION.md	Siamese/paired prediction	✅
methods/VALIDATED_DELTA_PREDICTION.md	Single-pass with ground truth (BEST)	✅
MULTI_STEP_FRAMEWORK.md	Decomposed classification	✅

Experiments

ID	Experiment	Outcome	Details
001	Canonical Classification	Partial Success	docs
002	Paired Delta Prediction	r=0.38	docs
003	Binary Classification	AUC=0.61	docs
004	Validated Delta (BEST)	r=0.41	docs

---

Quick Links

Getting Started

1. Read ARCHITECTURE.md for system overview
2. Review methods/ROADMAP.md for methodology context
3. Check experiments/004_validated_delta/ for best approach

For Training

1. Read TRAINING_GUIDE.md
2. Use methods/VALIDATED_DELTA_PREDICTION.md approach

For Understanding Results

1. Check experiments/README.md for experiment index
2. Review individual experiment docs for details

---

Package Documentation vs Project Documentation

This documentation is package-level (inside src/agentic_spliceai/docs/), focusing on: - Implementation details and R&D insights - Experiment logs and methodology development - Technical deep-dives

For project-level documentation (user-facing guides), see: - /docs/ - High-level project documentation - /README.md - Project overview

---

Related Code

Package	Path	Description
meta_layer.core	splice_engine/meta_layer/core/	Configuration, artifact loading
meta_layer.models	splice_engine/meta_layer/models/	Model implementations
meta_layer.training	splice_engine/meta_layer/training/	Training pipelines
meta_layer.workflows	splice_engine/meta_layer/workflows/	High-level workflows

---

Naming Conventions

We avoid cryptic names like "Phase 1", "Approach A", "Approach B" in favor of descriptive names:

Old Name	New Name	Rationale
Phase 1	Canonical Classification	Describes training data
Approach A	Paired Delta Prediction	Describes input format
Approach B	Validated Delta Prediction	Describes target source
Phase 1 Workflow	CanonicalTrainingWorkflow	Self-explanatory

---

Ported from meta_spliceai with improved naming conventions.