Claim Levels#

PhaseLab uses a hierarchical system to report uncertainty honestly. This document explains each claim level and when they apply.

Why Claim Levels?#

Most prediction tools report point estimates without uncertainty. PhaseLab acknowledges that reliability depends on:

  1. Data availability: Have we seen similar cases?

  2. Validation depth: Single dataset or cross-validated?

  3. Experimental support: Structural priors or tiling data?

Rather than hiding this, we report it explicitly.

The Four Levels#

UNKNOWN#

Meaning: Cannot assess reliability

Requirements: Insufficient data or context

When it applies:

  • No tiling data for the region

  • Novel target with no structural homologs

  • Missing critical information

Interpretation: Don’t make predictions. Gather more data.

result.claim_level == ClaimLevel.UNKNOWN
# "We can't tell you anything reliable"

EXPLORATORY#

Meaning: Preliminary assessment only

Requirements: Structural priors available

When it applies:

  • Before tiling screen (Stage I)

  • Using ENCODE/Roadmap chromatin data

  • Structural predictions (AlphaFold, etc.)

Interpretation: Directional guidance only. Not for decisions.

result.claim_level == ClaimLevel.EXPLORATORY
# "This is our best guess, but unvalidated"

CONTEXT_DEPENDENT#

Meaning: Valid within a specific context

Requirements: Single tiling dataset validation

When it applies:

  • One tiling screen analyzed

  • Spatial coherence validated (p < 0.05)

  • Same cell type/conditions as tiling

Interpretation: Reliable IF conditions match the validation context.

result.claim_level == ClaimLevel.CONTEXT_DEPENDENT
# "Validated in K562, may differ in other cells"

STRONG_COMPUTATIONAL#

Meaning: High computational confidence

Requirements: Cross-validated across datasets

When it applies:

  • Multiple independent tiling screens

  • Validated across cell types

  • Consistent coherence-outcome correlation

Interpretation: Strong computational evidence. Still needs wet lab.

result.claim_level == ClaimLevel.STRONG_COMPUTATIONAL
# "Validated across 6 genes, 115k guides"

How Claims Propagate#

When combining analyses, the overall claim level is the minimum of component levels:

# Example: Combining CRISPRa and circadian simulations
crispra_claim = ClaimLevel.CONTEXT_DEPENDENT
circadian_claim = ClaimLevel.EXPLORATORY

overall_claim = min(crispra_claim, circadian_claim)
# Result: EXPLORATORY (weakest link)

This ensures you never overstate confidence.

The Claim Level Hierarchy#

UNKNOWN < EXPLORATORY < CONTEXT_DEPENDENT < STRONG_COMPUTATIONAL

More data / validation required ←──────────→ More confidence

Upgrading Claims#

Claims only upgrade with additional evidence:

Claim Upgrades#

From

To

Required Evidence

UNKNOWN

EXPLORATORY

Structural priors or homolog data

EXPLORATORY

CONTEXT_DEPENDENT

Tiling screen validation (p < 0.05)

CONTEXT_DEPENDENT

STRONG_COMPUTATIONAL

Cross-validation across datasets

Important

Claims never upgrade automatically. PhaseLab requires you to provide additional evidence to increase confidence.

Using Claim Levels in Code#

from phaselab.fusion import ClaimLevel

# Check claim level
if result.claim_level >= ClaimLevel.CONTEXT_DEPENDENT:
    # Safe to make recommendations
    print("Validated prediction")
else:
    # Need more data
    print("Preliminary only - validation required")

# Filter by claim level
reliable_regions = [
    r for r in result.regions
    if r.claim_level >= ClaimLevel.CONTEXT_DEPENDENT
]

Decision Guide#

What to Do at Each Level#

Claim Level

Action

Use For

UNKNOWN

Gather data first

Nothing

EXPLORATORY

Run tiling screen

Experimental planning

CONTEXT_DEPENDENT

Proceed with caution

Guide selection, pilot studies

STRONG_COMPUTATIONAL

Proceed confidently

Large-scale experiments

Therapeutic Claims#

For therapeutic applications (e.g., SMS pipeline), claim levels are particularly important:

Warning

Even STRONG_COMPUTATIONAL claims require wet lab validation before any clinical application. PhaseLab provides computational guidance, not clinical recommendations.

The SMS pipeline generates falsification tests specifically designed to validate or invalidate computational predictions.

See Also#