Pfam Domain Pattern Recognition

Pfam Domain Pattern Recognition

Summary

Recognition and extraction of Pfam domain and subPfam patterns from genes within Biosynthetic Gene Clusters to create tokenized sequence representations. This skill enables systematic identification of co-occurring domain motifs that characterize functional gene sub-clusters.

When to use

When you have a set of Biosynthetic Gene Clusters (BGCs) in GenBank format and need to identify statistically significant or topic-modeled co-occurrence patterns of Pfam domains and subPfams across multiple genes. Apply this skill as the initial tokenization step before redundancy filtering and sub-cluster detection, or when you seek to link detected sub-clusters back to conserved domain architectures.

When NOT to use

• Input BGCs have no Pfam domain annotations or lack functional gene annotation — domain pattern recognition requires curated domain assignments as input.
• Goal is to predict biosynthetic product structure directly from gene sequence without intermediate sub-cluster identification — this skill focuses on domain tokenization, not end-to-end product prediction.
• BGC dataset is already redundancy-filtered and you only need to apply statistical or topic-based sub-cluster detection — skip to PRESTO-STAT or PRESTO-TOP instead.

Inputs

• Biosynthetic Gene Clusters in GenBank format
• Gene annotations with Pfam domain assignments
• Pfam domain database with subPfam hierarchy

Outputs

• Tokenized BGC representations (Pfam and subPfam token sequences per gene)
• Domain composition profiles per gene
• Adjacency matrices or co-occurrence frequency tables of domain patterns

How to apply

First, represent each gene in the BGC as a combination of its annotated Pfam domains, using subPfams to increase resolution and granularity of the token representation. This creates a token string per gene that captures the domain architecture. The rationale is that Pfam domains encode functional modules of biosynthetic enzymes; subPfams further discriminate between functional subtypes (e.g., different condensation domain variants). Once tokenized, these representations become amenable to statistical co-occurrence analysis (PRESTO-STAT) or topic modeling (PRESTO-TOP) to detect patterns that recur across many BGCs. The choice of subPfam resolution affects sensitivity: finer tokenization reveals more specific domain combinations but risks over-segmentation; coarser Pfam-only tokenization is more robust but may mask functional distinction.

Related tools

• iPRESTO (Command-line platform that implements Pfam/subPfam tokenization, redundancy filtering, and sub-cluster detection (PRESTO-STAT and PRESTO-TOP) on BGCs)

Evaluation signals

• All genes in input BGCs have been assigned at least one Pfam or subPfam token; no genes are left untokenized.
• Token composition reflects known biosynthetic enzyme architectures (e.g., polyketide synthase genes contain expected condensation, ketoreductase, and ACP domain tokens).
• Adjacency Index or domain co-occurrence frequency matrix is symmetric and non-negative; no NaN or infinite values present.
• When redundancy filtering is subsequently applied, the Adjacency Index correctly identifies highly similar tokenized BGCs as neighbors in the similarity network.
• Detected sub-clusters contain genes whose tokens show statistically significant or topic-weighted co-occurrence patterns that are rare or absent in random BGC samples.

Limitations

• Tokenization accuracy depends on completeness and quality of upstream Pfam domain annotation; unannotated genes or missing domains will degrade pattern detection.
• SubPfam resolution introduces trade-off: finer granularity increases sensitivity to domain variants but risks false fragmentation and reduces statistical power if individual sub-cluster patterns are rare.
• Pfam domains are ortholog-based and may not capture species-specific or novel biosynthetic adaptations that lie outside Pfam HMM definitions.
• The Adjacency Index distance metric assumes co-occurrence patterns are uniform across genomic contexts; gene order and synteny are not explicitly modeled in token representation.

Evidence

• [intro] BGCs are tokenised by representing each gene as a combination of its Pfam domains, where subPfams are used to increase resolution: "BGCs are tokenised by representing each gene as a combination of its Pfam domains, where subPfams are used to increase resolution"
• [intro] For the detection of sub-clusters two methods are used: PRESTO-STAT, which is based on the statistical algorithm from Del Carratore et al. (2019): "For the detection of sub-clusters two methods are used: PRESTO-STAT, which is based on the statistical algorithm from Del Carratore et al. (2019)"
• [intro] Tokenised BGCs are filtered for redundancy using similarity network with an Adjacency Index of domains as a distance metric: "Tokenised BGCs are filtered for redundancy using similarity network with an Adjacency Index of domains as a distance metric"
• [intro] iPRESTO (integrated Prediction and Rigorous Exploration of biosynthetic Sub-clusters Tool) is a command line tool for the detection of gene sub-clusters: "iPRESTO (integrated Prediction and Rigorous Exploration of biosynthetic Sub-clusters Tool) is a command line tool for the detection of gene sub-clusters"