Remote Homology Usage Guide

Overview

Detect distant homologs using profile and structure-aware methods. Encodes the modern landscape: PSI-BLAST and jackhmmer for iterative profile search; HHblits/HHsearch for profile-profile (PDB70, Pfam); MMseqs2 and DIAMOND as the modern blastp replacements (100-10,000x faster); Foldseek (van Kempen 2024) for structure-aware homology via the 3Di alphabet, with ProstT5 for sequence-only access to structural search.

Prerequisites

conda install -c bioconda hmmer mmseqs2 diamond hhsuite foldseek blast
hmmsearch -h | head -3       # HMMER 3.4+
mmseqs version               # MMseqs2 15+
diamond --version            # DIAMOND 2.1+
hhblits -h | head -3         # HH-suite3 3.3+
foldseek --version           # Foldseek 9+

For Foldseek databases:

mkdir -p foldseek_dbs tmp
foldseek databases --help    # list available DBs

Quick Start

"Find structural homologs of this protein via Foldseek against AlphaFoldDB Swiss-Prot subset"
"Run jackhmmer 3 iterations against UniRef90; save the final HMM for downstream hmmsearch"
"Search a protein query with MMseqs2 at -s 7.5 (HMMER-equivalent sensitivity) instead of BLAST"
"Annotate domains with hmmscan against Pfam-A using --cut_ga calibrated thresholds"
"DIAMOND --ultra-sensitive against UniRef90 for a metagenomic protein set"

Example Prompts

Foldseek for divergent structural homologs

"I have a protein where blastp finds nothing significant. Predict the structure (or use ProstT5 for sequence-only) and search AlphaFoldDB with Foldseek easy-search. Report TM-scores and probability cutoff -- hits with prob > 0.9 are structurally confident."

PSI-BLAST with drift protection

"Run psiblast for 3 iterations against UniRef90 with -inclusion_ethresh 0.002 (stricter than the 0.005 default). Save the PSSM with -out_pssm and the included sequence set so I can audit for paralog contamination. Don't iterate to convergence -- 4+ iterations drift."

MMseqs2 as PSI-BLAST replacement

"Same iterative profile search as PSI-BLAST but use MMseqs2 with --num-iterations 3 and -s 7.5 (HMMER-equivalent sensitivity). It's 100x faster."

Pfam domain annotation

"Annotate domains of these 5,000 protein sequences using hmmscan against Pfam-A with --cut_ga. Use the calibrated gathering thresholds, not arbitrary E-value cutoffs."

HHsearch against PDB70

"Build an HHblits profile of this protein against UniRef30 (3 iterations), then hhsearch against PDB70 for the deepest possible structural homology to known PDB entries."

DIAMOND for metagenomic scale

"I have 1 million predicted ORFs from a metagenome. Use DIAMOND blastp --ultra-sensitive against UniRef90 with -p 32. blastp would take days; DIAMOND will finish in an hour."

What the Agent Will Do

Identify whether the problem is single-query distant homology (PSI-BLAST, jackhmmer, Foldseek) or large-batch (MMseqs2, DIAMOND).
Default to MMseqs2 (-s 7.5) or DIAMOND --ultra-sensitive over BLAST for any batch >50 sequences.
For very deep homology to known structures, run Foldseek (with ProstT5 if no structure available) against AlphaFoldDB or PDB100.
For domain annotation, use hmmscan vs Pfam-A with --cut_ga.
For PSI-BLAST, cap iterations at 3 and use stricter inclusion threshold; save PSSM for reproducibility.
For HHsearch profile-profile, build query MSA via HHblits first.
Recommend mixing sequence + structure evidence for the toughest cases.
Note when fold-level similarity alone (Foldseek to a TIM barrel) is not evidence of homology.

Tips

The twilight zone of sequence homology is 20-35% pairwise identity (Rost 1999). Below that, structure beats sequence -- reach for Foldseek.
ProstT5 (Heinzinger 2024) is the bridge: sequence -> 3Di alphabet directly, no AF2 step needed for Foldseek.
For metagenomic-scale work, MMseqs2 easy-cluster is the modern de-facto for clustering hundreds of millions of sequences.
PSI-BLAST is non-deterministic in detail (input order matters). For reproducibility, save the PSSM and re-use with -in_pssm.
HHsearch vs PDB70 is still the gold standard for deepest possible homology to PDB; for AlphaFoldDB coverage, prefer Foldseek.
DIAMOND v2 adds frameshift-aware mode (--frameshift 15) for nanopore / PacBio long-read protein search.
Common folds (TIM barrel, Rossmann, ABC) appear across many superfamilies. A high Foldseek score on a common fold is necessary but not sufficient for homology.
The Foldseek-multimer mode (Foldseek 9+) searches protein complexes against multimer databases -- relevant for interactome work.

blast-searches - Remote BLAST baseline
local-blast - Local BLAST+ for moderate-scale work
ortholog-inference - Distinct topic: orthology vs homology (RBH, OrthoFinder, OMA)
alignment/multiple-alignment - Build MSAs for HMM profiles
structural-biology/alphafold-predictions - Predict structures for Foldseek queries