Recommending Bacterial Cultivation Conditions from 16S rRNA

Juntang Wang; Hao Wu; Yihan Wang; Runkun Guo; Shixin Xu; Dongmian Zou

cs.LGOct 20, 2026

Recommending Bacterial Cultivation Conditions from 16S rRNA

Authors: Juntang Wang, Hao Wu, Yihan Wang, Runkun Guo, Shixin Xu, Dongmian Zou

Organizations: Harvard University USA · Duke Kunshan University China · Duke University USA · Boston University USA

Abstract

Recommending bacterial cultivation conditions from sequence data is a challenging structured problem characterized by missing standardized benchmarks and limited use of sequence structure. To address this, we introduce DSNI, a benchmark of 65,023 bacterial strains with phylogeny-aware splits, pairing 16S rRNA sequences with four cultivation conditions (media, temperature, pH, oxygen). Exploiting the mosaic structure of 16S rRNA, where hypervariable regions support discrimination and conserved regions support primer design, we propose a Variability-Gated Encoder that learns position-dependent importance weights conditioned on biological features (region type, position, and local complexity). The gating module adds minimal overhead (1,280 parameters, <1%) while providing an explicit inductive bias valuable in low-data regimes. Experiments demonstrate that the learnable gate improves flat Transformers by 4.7% in joint accuracy and complements RNABERT, yielding 44.2% exact-match joint accuracy. The model also recovers 90.2% of ground-truth cultivation protocols within its top-25 recommendations, narrowing the exhaustive search space by approximately 60 times.

Figures & tables

Panel of Figure 1 — Figure 1: DSNI benchmark statistics and label distributions. Left: summary statistics for 65,023 strains. Right: phylogenetic, temperature, oxygen, and media distributions. DSNI is dominated by Proteobacteria and Firmicutes, shows strong mesophilic bias, and has long-tailed media categories.

Panel of Figure 2 — Figure 2: Architecture of the variability-gated encoder. The input 16S rRNA sequence is embedded with positional encodings, then modulated by a lightweight gating layer (dashed box) that computes position-wise importance weights. The gate conditions on biological features: region tags indicating hypervariable (V1–V9) vs. conserved positions, positional encoding, local GC content, and Shannon entropy. Gated embeddings are processed by a Transformer encoder, and a [CLS] token representation feeds four task-specific heads for media, temperature, pH, and oxygen requirement predictions.

Traditional ML
Method	Params	Media Acc.	Temp. Acc.	pH Acc.	O 2 Acc.	Joint Acc.
$k$ -mer + RF	—	38.2 $\pm$ 1.1	86.4 $\pm$ 0.8	89.1 $\pm$ 0.7	72.3 $\pm$ 0.9	24.7 $\pm$ 1.2
$k$ -mer + XGBoost	—	41.5 $\pm$ 0.9	87.8 $\pm$ 0.6	90.3 $\pm$ 0.5	74.6 $\pm$ 0.8	27.3 $\pm$ 1.0
Phylogenetic / Similarity-Based
GROWREC	—	46.3 $\pm$ 0.0	85.2 $\pm$ 0.0	88.4 $\pm$ 0.0	71.8 $\pm$ 0.0	28.9 $\pm$ 0.0
$k$ -NN (Sequence)	—	43.8 $\pm$ 0.0	87.1 $\pm$ 0.0	89.7 $\pm$ 0.0	73.5 $\pm$ 0.0	27.6 $\pm$ 0.0

Table 1: Main results on the DSNI test set. Best in bold , second-best underlined ; mean

\pm

std over 5 seeds. Params

=

trainable parameters in millions.

Panel of Figure 3 — Figure 3: Few-shot learning curves for joint accuracy across training set sizes. Shaded regions indicate $\pm$ 1 standard deviation over 3 random seeds. Gating improves performance most strongly in the low-data regime, and RNABERT+Gating performs best across data sizes.

Table 5

Panel of Figure 4 — Figure 4: Average gate activation across 16S rRNA sequence positions. Shaded regions indicate hypervariable regions (V1–V9); the model learns to assign higher gate values to these regions, particularly V3 and V4 which provide the finest species-level discrimination. Conserved regions receive lower weights (near the 0.5 baseline), consistent with their limited discriminative power for cultivation prediction.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 5 — Figure 5: Entity-relationship diagram of the DSNI database schema. The central STRAINS table contains 16S rRNA sequences and cultivation conditions, linked to MEDIA through a many-to-many relationship. Media compositions are decomposed into SOLUTIONS and INGREDIENTS, while preparation protocols are captured in STEPS. Atmospheric requirements are stored in the GAS table.

Table	Field	Description
STRAINS	strain_id	Unique identifier (DSM/BEI number)
	genus	Taxonomic genus assignment
	species	Species epithet
	sequence_16s	Full or partial 16S rRNA sequence
	opt_temp	Optimal growth temperature ( ∘ C)
	opt_ph	Optimal growth pH

Appendix

Table 4: Key fields in the DSNI database schema.

Resource	Organisms	Labels	Split strategy
KOMODO [ 22 ]	${\sim}18$ K pairs	Media only	Random
MediaDB	${\sim}800$ orgs.	Media composition	None (catalog)
DSNI (ours)	65,023 strains	4 conditions	Genus-aware

Appendix

Table 5: Cultivation-prediction benchmarks. DSNI is the largest resource and the only one to provide genus-aware splits and standardized labels for all four primary conditions.

Method	Input	Throughput	Coverage	Limitation
Trial-and-error	Isolate	50–200 cond./org.	Single organism	Slow, expensive
Culturomics [ 16 ]	Fecal/env. sample	${\sim}200$ cond./run	Broad, untargeted	5–30% success
Dilution-to-extinction	Liquid sample	Low	Specialized	Low-nutrient only
DSNI (ML)	16S only	Instant	Any 16S-tagged org.	Needs validation

Appendix

Table 6: Wet-lab cultivation methods vs. DSNI-based ML recommendation. ML provides instant, organism-agnostic ranked shortlists that complement rather than replace wet-lab methods.

Gating Layer
Hyperparameter	Transformer	RNABERT
Hidden dimension $d$	256	768 (fixed)
Number of layers $N$	6	12 (fixed)
Attention heads	8	12 (fixed)
Feed-forward dim	1024	3072 (fixed)
Dropout	0.1	0.1

Appendix

Table 7: Hyperparameters for all models. Values were selected via grid search on the validation set.

$A\to B$	$H(A)$	$H(A\mid B)$	$\mathrm{MI}(A;B)$
Media $\to$ Temperature	4.61	4.38	0.23
Media $\to$ Oxygen	4.61	4.27	0.34
Temperature $\to$ pH	0.61	0.57	0.04

Appendix

Table 8: Empirical mutual information

\mathrm{MI}(A;B)

between cultivation-condition label pairs on the training split.

\mathrm{MI}/H(A)

measures how much of

A

’s entropy is explained by

B

Temperature
	Flat Trans.	+Gating	RNABERT+Gating
Psychrophile (6%)	52.3	58.7	61.2
Mesophile (89%)	94.2	95.1	95.8
Thermophile (5%)	48.6	55.4	58.9
pH
Acidophile (4%)	45.2	52.8	56.3

Appendix

Table 9: Per-class F1 scores for temperature and pH prediction. Transformer+Gating model shows improved performance on minority classes (psychrophile, thermophile, acidophile, alkaliphile).

Media category	# samples	F1
TSA (most common)	8,247	68.4
BHI	5,182	61.7
Nutrient Agar	4,891	59.3
Marine Agar	3,256	52.8
Acidic mineral (tail)	47	12.3
Methanogen medium (tail)	31	8.7

Appendix

Table 10: Per-class F1 scores for the media task (RNABERT+Gating), illustrating the long-tail collapse. Sample counts are training-set frequencies.

Method	Media	Temp.	pH	O 2
$k$ -mer + RF	28.4	62.3	58.7	54.2
$k$ -mer + XGBoost	31.2	65.1	61.4	57.8
CNN Encoder	34.6	67.8	63.2	60.4
Transformer (Flat)	38.2	71.5	66.8	64.1
Transformer + Gating	43.7	76.2	72.4	69.8
RNABERT	41.8	74.6	70.9	68.2

Appendix

Table 11: Macro-F1 scores on the DSNI test set. Macro-averaging weights all classes equally, emphasizing performance on rare classes.

Distance to nearest training genus	$N$	RNABERT+Gating	GROWREC	$\Delta$
Close ( $>$ 90% 16S identity)	3,180	52.1%	38.4%	$+13.7$ pp
Medium (80–90%)	4,130	43.8%	27.6%	$+16.2$ pp
Distant ( $<$ 80%)	2,443	33.1%	16.3%	$+16.8$ pp

Appendix

Table 12: Joint accuracy stratified by 16S identity to the nearest training-set genus. RNABERT+Gating outperforms the phylogenetic-matching baseline GROWREC by an increasing margin at greater phylogenetic distances.

# conditions correct	% of test strains	Cumulative
4/4 (exact match)	44.2%	44.2%
3/4	32.1%	76.3%
2/4	16.8%	93.1%
$\leq$ 1/4	6.9%	100%

Appendix

Table 13: Distribution of correctly predicted conditions for RNABERT+Gating on the DSNI test set.

Task	Classes	Top-1	Top-2	Top-3	Top-5
Media	42	54.8%	66.3%	74.8%	82.1%
Temperature	3	92.1%	98.4%	—	—
pH	3	93.5%	98.8%	—	—
Oxygen	4	82.7%	93.2%	97.1%	—

Appendix

Table 14: Per-task top-

k

accuracy for RNABERT+Gating. “—” indicates that

k

exceeds the number of classes for that task.

Split strategy	Flat Trans.	Trans.+Gating	RNABERT+Gating
Random split	58.2%	61.7%	64.3%
Genus-aware (ours)	36.8%	41.5%	44.2%
Inflation	$+21.4$ pp	$+20.2$ pp	$+20.1$ pp

Appendix

Table 15: Effect of split strategy on joint accuracy. Random splits inflate apparent predictability by

\sim

20 pp across encoders.

Comparison	$\Delta$ Joint Acc.	95% Bootstrap CI	McNemar $p$
Transformer+Gating vs. Flat Transformer	$+4.7$ %	$[+3.1\%,+6.2\%]$	$<0.001$
RNABERT+Gating vs. RNABERT	$+4.5$ %	$[+2.9\%,+6.0\%]$	$<0.001$
RNABERT+Gating vs. Flat Transformer	$+7.4$ %	$[+5.6\%,+9.1\%]$	$<0.001$

Appendix

Table 16: Statistical significance of the gating-induced improvements in joint accuracy.

Method	Full length	440-token crop	$\Delta$ (crop)
Flat Transformer	36.8%	33.4%	$-3.4$ pp
Transformer+Gating	41.5%	38.2%	$-3.3$ pp
RNABERT (fixed 440)	—	39.7%	—
RNABERT+Gating (fixed 440)	—	44.2%	—

Appendix

Table 17: RNABERT truncation control: joint accuracy under matched 440-token center cropping. Gating gains are preserved under matched inputs.

Bioinformatics

See all 40 papers

A Boltzmann-machine-enhanced Transformer For DNA Sequence Classification
Mar 27, 2026Zhixuan Cao, Yishu Xu, Xuang WUTransformer Transformer Attention
Tsinghua University · University of California, Berkeley
HyperTaxa: Multiscale Genome Representations for Generalized Microbial Taxonomic Inference
Oct 19, 2026Akash Rao, Sushim Gupta, Vishalini Ramnath +2Bioinformatics Representation Learning
CSSE Department, Auburn University, Auburn, AL, USA · Oklahoma Animal Disease Diagnostic Laboratory, Oklahoma State University, Stillwater, OK, USA
Continuous Diffusion Transformers for Designing Synthetic Regulatory Elements
Mar 11, 2026Jonathan Liu, Kia GhodsMolecular Generation Diffusion Transformer
Department of Computer Science Princeton University

Mar 27, 2026J/K move · Enter open · S save

A Boltzmann-machine-enhanced Transformer For DNA Sequence Classification

Zhixuan Cao, Yishu Xu, Xuang WU

Tsinghua University · University of California, Berkeley

Transformer Transformer Attention Genomics

Open paper AskPDF arXiv

DNA sequence classification requires not only high predictive accuracy but also the ability to uncover latent site interactions, combinatorial regulation, and epistasis-like higher-order dependencies. Although the standard Transformer provides strong global modeling capacity, its softmax attention is continuous, dense, and weakly constrained, making it better suited for information routing than explicit structure discovery. In this paper, we propose a Boltzmann-machine-enhanced Transformer for DNA sequence classification. Built on multi-head attention, the model introduces structured binary gating variables to represent latent query-key connections and constrains them with a Boltzmann-style energy function. Query-key similarity defines local bias terms, learnable pairwise interactions capture synergy and competition between edges, and latent hidden units model higher-order combinatorial dependencies. Since exact posterior inference over discrete gating graphs is intractable, we use mean-field variational inference to estimate edge activation probabilities and combine it with Gumbel-Softmax to progressively compress continuous probabilities into near-discrete gates while preserving end-to-end differentiability. During training, we jointly optimize classification and energy losses, encouraging the model to achieve accurate prediction while favoring low-energy, stable, and interpretable structures. We further derive the framework from the energy function and variational free energy to the mean-field fixed-point equations, Gumbel-Softmax relaxation, and the final joint objective. The proposed framework provides a unified view of integrating Boltzmann machines, differentiable discrete optimization, and Transformers for structured learning on biological sequences.

Abstract

Bioinformatics

Explore similar work

A Boltzmann-machine-enhanced Transformer For DNA Sequence Classification

HyperTaxa: Multiscale Genome Representations for Generalized Microbial Taxonomic Inference

Continuous Diffusion Transformers for Designing Synthetic Regulatory Elements