STEAM — Search with TEA against Many

STEAM is heavily adapted from Foldseek (Van Kempen et al., Nature Biotechnology 2024), replacing Foldseek's 3Di structural alphabet with TEA. This means STEAM can be applied to any protein sequence, no 3D structure required. Like Foldseek, STEAM is built on the MMseqs2 framework.

Requirements

• CMake >= 3.15
• GCC >= 7 or Clang
• For TEA sequence generation: TEA (pip install git+https://github.com/PickyBinders/tea.git)

Installation

# Install build dependencies (if needed)
mamba install -c conda-forge cmake gxx_linux-64

# Build
git clone --recursive https://github.com/PickyBinders/steam.git
cd steam
mkdir build && cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make -j4

The binary will be at build/src/steam.

Quick start

1. Generate TEA sequences

Convert amino acid sequences to the TEA structural alphabet using the tea_convert tool:

tea_convert -f proteins.fasta -o proteins_tea.fasta

This requires a GPU and the TEA package. The output is a FASTA file with TEA sequences in the same order as the input.

2. Search

steam easy-search query_tea.fasta query_aa.fasta \
                   target_tea.fasta target_aa.fasta \
                   result.m8 tmp

Useful flags:

Flag	Default	Notes
-e	100	E-value threshold
--max-seqs	2000	Maximum results per query from prefiltering

3. Cluster

easy-cluster runs cascaded clustering (sensitive) and easy-linclust runs linear-time clustering (faster, less sensitive). Both take paired TEA/AA FASTAs:

# Cascaded clustering
steam easy-cluster proteins_tea.fasta proteins_aa.fasta clusterResult tmp

# Linear-time clustering (large datasets)
steam easy-linclust proteins_tea.fasta proteins_aa.fasta clusterResult tmp

Outputs three files alongside clusterResult:

• clusterResult_cluster.tsv — <representative> <member> adjacency list
• clusterResult_rep_seq.fasta — one AA sequence per cluster representative
• clusterResult_all_seqs.fasta — FASTA grouped by cluster

Useful flags:

Flag	Default	Notes
--min-seq-id	0	Minimum sequence identity for cluster members
-c	0.8	Minimum coverage
--cov-mode	0	0=bidirectional, 1=target, 2=query
--cluster-reassign	off	Cascaded only: corrects criteria-violations from cascaded merging
--single-step-cluster	off	Cascaded only: skip cascading, single pass

Commands

Command	Description
easy-search	Search FASTA pairs against FASTA pairs or a pre-built database
easy-cluster	Cluster paired TEA/AA FASTAs (cascaded, sensitive)
easy-linclust	Cluster paired TEA/AA FASTAs (linear-time, faster)
createdb	Create a STEAM database from paired TEA/AA FASTA files
search	Search pre-built databases (faster for repeated searches)
cluster	Cluster a pre-built database (cascaded)
linclust	Cluster a pre-built database (linear-time)
convertalis	Convert alignment results to various output formats
createsubdb	Subset a STEAM database (keeps _aa companion in sync)

Database workflow

For searching or clustering the same database multiple times, pre-build it:

# Create database (one time)
steam createdb target_tea.fasta target_aa.fasta targetDB

# Search against pre-built database (fast, repeatable)
steam easy-search query_tea.fasta query_aa.fasta targetDB result.m8 tmp

# Cluster the pre-built database
steam cluster targetDB clusterDB tmp

Output format

Default BLAST-tab format (same as MMseqs2/BLAST -outfmt 6):

query  target  fident  alnlen  mismatch  gapopen  qstart  qend  tstart  tend  evalue  bits

Custom output with --format-output adds TEA-specific output columns:

Column	Description
tfident	TEA fractional identity
tpident	TEA percent identity
qteaseq	Query TEA full sequence
tteaseq	Target TEA full sequence
qteaaln	Query TEA aligned sequence
tteaaln	Target TEA aligned sequence

Standard MMseqs2 output columns (fident, alnlen, qcov, tcov, evalue, raw, bits, etc.) are also available.

Scoring

The alignment score at each position is the sum of:

• MATCHA score: substitution score from the TEA alphabet matrix
• AA score: BLOSUM62 substitution score, weighted by --aa-weight (default 1.4)

E-value computation

STEAM uses a log-linear E-value model following Edgar & Sahakyan (2025). E-values are computed as:

E(s) = (H/Q) * 10^(m*s + c)

where s is the raw alignment score, H/Q is the average number of reported hits per query (computed at runtime from prefilter results), and m and c are parameters fitted on SCOP40c.