Models - Samtool Supported

This guide provides a comprehensive overview of the models supported by SAMtools, the ubiquitous suite of utilities for interacting with high-throughput sequencing data.

Because "model" can refer to two different things in bioinformatics—sequencing technologies (how data is generated) or genomic feature formats (how data is described)—this guide covers both contexts to ensure you have the complete picture.

SAMtool Supported Models

SAMtool (Segment Anything Model tool) provides a unified interface for various promptable segmentation models. Below is the complete list of officially supported model architectures and specific weight variants as of the latest release (v1.2.0).

Future Roadmap: Upcoming Model Support

The Samtool development team has announced support for the following models in the v2.5 release (Q3 2025):

Additionally, Samtool will introduce automatic operator substitution for unsupported ops using MLIR dialects. samtool supported models

Beyond the Mask: A Deep Dive into SAMtool Supported Models

In the rapidly evolving landscape of computer vision, the release of Meta AI’s Segment Anything Model (SAM) was a watershed moment. However, deploying SAM effectively often requires more than just the base model; it requires robust tooling. Enter SAMtool—a collection of utilities, wrappers, and extensions designed to streamline segmentation workflows.

But one question dominates the technical forums: Which models actually work with SAMtool? The answer is more nuanced than a simple list. Here is your comprehensive guide to the architectures, checkpoints, and custom variants supported by the modern SAMtool ecosystem.

4.2 The Somatic (Tumor/Normal) Model

For somatic mutation calling, SAMtools provides samtools mpileup with the -B (disable BAQ) flag for tumor samples to avoid over-filtering, followed by BCFtools with the -c (consensus caller) or somatic-specific priors.

Critical addition: samtools calmd (Calculate MD tags) is used post-alignment to realign reads based on the reference, reducing false positives in homopolymer regions. This guide provides a comprehensive overview of the

2.1 Core Utilities as Model Components

A SAMtools-supported model treats each command as a node in a directed acyclic graph (DAG):

| Node | Function | Input | Output | I/O Model | | :--- | :--- | :--- | :--- | :--- | | samtools view | Conversion/filtering | SAM/BAM/CRAM | BAM | Stream | | samtools sort | Coordinate sorting | Unsorted BAM | Sorted BAM | Disk-heavy | | samtools index | BAI indexing | Sorted BAM | .bai file | Random access | | samtools mpileup | Genotype likelihood | Multiple BAMs | BCF/VCF | Streaming + math | | samtools depth | Coverage histogram | BAM | Text table | Streaming |

5. Case Study: Reducing False Positives in Low-AF Variants

Problem: In circulating tumor DNA (ctDNA) with 1% allele frequency (AF), standard SAMtools models produce high false positives due to sequencing errors.

Solution: A SAMtools-supported model with hard filtering and strand bias correction. Stable Diffusion 3 – Multi-modal text-to-image with DiT

Workflow:

  1. samtools view -q 30 -F 0x04 in.bam (Keep only reads with MAPQ >= 30, not unmapped).
  2. samtools mpileup -d 1000 -Q 20 -f ref.fa in.bam > pileup.txt
  3. Custom Python script to compute Fisher’s exact test on strand bias.
  4. bcftools filter -e 'INFO/DP<20 || INFO/AF<0.01 || INFO/STB>0.05'

Result: Sensitivity dropped from 95% to 89%, but precision improved from 0.45 to 0.82 (validated by ddPCR).

4.3 The Coverage-Based CNV Model

A lightweight model for copy number variation:

samtools depth -a -b regions.bed in.bam | \
    awk 'sum+=$3 END print sum/NR'   # Mean depth per region

This model bypasses complex segmentation algorithms and uses SAMtools' native -d (max depth) flag to normalize GC bias.