Polee

Polee constructs a computationally efficient approximation of the RNA-seq likelihood to quantify transcript expression and propagate uncertainty into downstream statistical inference, mitigating biases from ambiguous read assignment such as multi-mapping reads, similar isoforms, paralogs, or duplicated loci.

Key Features:

• Uncertainty-aware inference: Propagates quantification uncertainty into downstream statistical analyses rather than treating point estimates as fixed inputs.
• Compact likelihood approximation: Replaces the full read-level likelihood with a low-dimensional approximation using the Pólya tree transformation to improve tractability.
• Scalable computation: Remains feasible on large RNA-seq datasets without requiring explicit probabilistic modeling of every read.
• Improved transcript-level analyses: Suited for differential transcript expression where isoform ambiguity is a primary source of variance.

Scientific Applications:

• Differential transcript expression: Provides more reliable detection and ranking of transcript-level changes across conditions by incorporating quantification uncertainty.
• Isoform-resolved regulation studies: Supports analysis of alternative splicing and isoform switching under experimental perturbations.
• Complex transcriptomes: Applicable to organisms or tissues with extensive paralogy, repeats, or closely related gene families where multi-mapping is prevalent.
• Budget-constrained pipelines: Enables probabilistic-style inference that captures uncertainty with moderate computational resources compared to full read-level models.

Methodology:

Conventional workflows follow a two-stage procedure: (i) estimate gene/transcript abundances and (ii) treat those estimates as fixed inputs for downstream models.

Polee constructs a computationally efficient approximation of the likelihood function for expression parameters using the Pólya tree transformation and substitutes this approximate likelihood for the full read-level likelihood.

The approximate likelihood is then used in downstream inference (e.g., differential transcript expression) to incorporate uncertainty without modeling billions of reads directly.