Classification Tasks¶

Leech is a general library for training classifiers on motif-centered nanopore signal: the task, the classes, and what counts as "identity" for a conditional or hierarchical model are all yours to define. This page works through the concrete example that drove leech's development -- tRNA aminoacylation ("charging") state and amino acid identity from direct RNA sequencing -- because it exercises binary classification, hard multi-way discrimination, and identity-conditioned modeling strategies that generalize to other motif-centered classification tasks (e.g., other base modifications identified by a known sequence context).

Within that example, leech supports two tasks of increasing difficulty: binary charged/uncharged classification and amino acid discrimination. This page describes both tasks, the modeling strategies for each, and how tRNA identity can be used to simplify the amino acid problem.

Charged vs. uncharged classification¶

The simpler task is distinguishing charged (aminoacylated) from uncharged tRNAs. Aminoacylation is expected to shift signal amplitude at the CCA tail, which should be large enough for signal-only models to detect with reasonable accuracy.

Dwell features still improve performance on this task because they capture translocation kinetics that amplitude alone misses, but the binary task is largely solvable without them. The more challenging and scientifically interesting task is identifying which amino acid is attached.

Amino acid discrimination¶

The harder, more scientifically interesting task in this example is classifying which of the 20 standard amino acids is attached to a charged tRNA. This is substantially harder than binary classification because the signal differences between amino acids are small.

Consider three amino acids. The values below are illustrative numbers chosen to show the shape of the problem, not measurements:

While Trp is distinguishable by amplitude, Ala and Gly differ by only 1 pA -- within read-to-read noise. In this illustration their dwell times differ substantially, which is why dwell features are hypothesized to help. More broadly, amino acid physical properties are hypothesized to relate to dwell time:

This is the expected physical basis for dwell-based discrimination: larger amino acids may cause more drag during translocation, increasing dwell time at the CCA tail. No correlation coefficients have been measured or recorded in this repository.

Expected confusion patterns¶

Some amino acid pairs are inherently difficult to separate due to chemical similarity:

• High confusion: Ile/Leu, Asp/Glu, Ser/Thr
• Low confusion: Gly/Trp, Lys/Asp

tRNA-conditional strategies¶

At inference time, the tRNA identity is known from sequence alignment. This knowledge can dramatically simplify the classification problem because aminoacyl-tRNA synthetases (aaRSs) are highly specific: mischarging occurs almost exclusively between chemically similar amino acids. Instead of a 20-way classification, the problem reduces to a binary or few-way task.

Common mischarging pairs¶

Each tRNA is charged by a specific aaRS that occasionally mischarges with a chemically similar amino acid:

tRNAs not listed (Ala, Arg, Asn, Cys, Gln, Gly, His, Lys, Met, Pro, Trp) have highly specific synthetases with rare mischarging.

Modeling strategies¶

Four approaches leverage tRNA identity:

tRNA-specific binary classifiers. Train one classifier per tRNA family that distinguishes the cognate amino acid from its known near-cognates. For example, a tRNA-Ile classifier performs 3-way classification (Ile vs. Val vs. Leu). This is the simplest approach and works well when training data is available for each tRNA.

Hierarchical classification. A two-stage pipeline: first identify the tRNA (from sequence), then apply a tRNA-specific second-stage classifier for amino acid identity. This separates the problems cleanly.

Conditional input. A single model receives tRNA identity as an additional input (e.g., a learned embedding). The model uses a loss mask so that only the cognate and near-cognate amino acids contribute to the loss for each tRNA. This allows shared feature learning across tRNAs.

Shared encoder with tRNA-specific heads. A common ConvLSTM encoder processes signal, sequence, and dwell features. The encoder output feeds tRNA-specific classification heads, each performing binary or few-way classification. This balances parameter sharing with task-specific outputs.

Multi-task learning¶

A multi-task model predicts both charging state and amino acid identity simultaneously using a shared encoder with two output heads:

• Charging head -- binary classification (charged vs. uncharged)
• Amino acid head -- 21-way classification (20 amino acids + uncharged)

The combined loss uses a weighted sum:

total_loss = 0.7 * charging_loss + 0.3 * aa_loss

The charging task is easier and provides a strong gradient signal early in training, while the amino acid task drives the model to learn finer-grained representations. The shared encoder benefits from both objectives.

Expected performance¶

Performance depends on data source. Synthetic data (purified tRNAs charged with individual amino acids) provides clean training signal. Biological data introduces variability from cellular context, mixed populations, and post- transcriptional modifications.

The expected gap between 20-way and tRNA-conditional performance motivates the value of incorporating biological knowledge about aaRS specificity.