ResGene-T: A Tensor-Based Residual Network Approach for Genomic Prediction

Kuldeep Pathak; Kapil Ahuja; Eric de Sturler

cs.LGFeb 28, 2026

ResGene-T: A Tensor-Based Residual Network Approach for Genomic Prediction

Authors: Kuldeep Pathak, Kapil Ahuja, Eric de Sturler

Organizations: Department of Computer Science and Engineering, Indian Institute of Technology Indore, India 453552 · Department of Mathematics, Virginia Tech, Blacksburg, USA 24061

Abstract

In this work, we propose a new deep learning model for Genomic Prediction (GP), which involves correlating genotypic data with phenotypic. The genotypes are typically fed as a sequence of characters to the 1D-Convolution Neural Network layer of the underlying deep learning model. Inspired by earlier work that represented genotype as a 2D-image for genotype-phenotype classification, we extend this idea to GP, which is a regression task. We use a ResNet-18 as the underlying architecture, and term this model as ResGene-2D. Although the 2D-image representation captures biological interactions well, it requires all the layers of the model to do so. This limits training efficiency. Thus, as seen in the earlier work that proposed a 2D-image representation, our ResGene-2D performs almost the same as other models (3% improvement). To overcome this, we propose a novel idea of converting the 2D-image into a 3D/ tensor and feed this to the ResNet-18 architecture, and term this model as ResGene-T. We evaluate our proposed models on three crop species having ten phenotypic traits and compare it with seven most popular models (two statistical, two machine learning, and three deep learning). ResGene-T performs the best among all these seven methods (gains from 14.51% to 41.51%).

Figures & tables

Studies	Dataset	Traits	Methods
Wang et al. [ 13 ] (2018)	Arabidopsis, Maize, Mice, Rice, Simulated data	One hundred fifty seven	SBLUP, CBLUP, GBLUP, Bayesian LASSO
Haile et al. [ 14 ] (2021)	Wheat	Six	rrBLUP, GBLUP, BayesB, BL, RKHS, GS + de novo GWAS
Meher et al. [ 2 ] (2022)	Wheat, Maize, Barley, Simulated data	Nine	GBLUP, CBLUP, SBLUP, BayesA, BayesB, BayesC, BLASSO, BRR

TABLE I: Summary of previous studies based on Statistical Methods

Studies	Dataset	Traits	Methods
Grinberg et al. [ 15 ] (2020)	Yeast, Rice, Wheat	Sixty two	Elastic net, Ridge Regression, LASSO Regression, RF, GBM, SVM, GBLUP, Bloom
Yan et al. [ 7 ] (2021)	Maize	Three	LightGBM, XGBoost, CatBoost, rrBLUP
Wang et al. [ 16 ] (2022)	Pigs	Two	SVR, KRR, RF, AdaBoost.R2, GBLUP, ssGBLUP, BayesHE

TABLE II: Summary of previous studies based on Machine Learning Methods

Studies	Dataset	Traits	Methods
Liu et al. [ 9 ] (2019)	Soybean, Simulated data	Five	DLGWAS, DeepGS, Single CNN, rrBLUP, BRR, BayesA, Bayessian LASSO
Sehrawat et al. [ 17 ] (2023)	Arabidopsis Thaliana, Oryza Sativa	Eight	NovGMDeep, G2Pdeep, DeepGS, GBLUP, rrBLUP
Wang et al. [ 10 ] (2023)	Wheat, Maize, Tomato	Nineteen	DNNGP, DLGWAS, DeepGS, SVR, LightGBM, GBLUP
Lu et al. [ 11 ] (2024)	Soybean	Three	GPFormer, MLP, SVR, RF, XGBoost

TABLE III: Summary of previous studies based on Deep Learning Methods

Varieties	$\boldsymbol{\mathcal{M}_{1}}$	$\boldsymbol{\mathcal{M}_{2}}$	$\cdots$	$\boldsymbol{\mathcal{M}_{d}}$
$\mathcal{V}_{1}$	A	K	$\cdots$	T
$\mathcal{V}_{2}$	A	K	$\cdots$	C
$\vdots$	$\vdots$	$\vdots$	$\vdots$	$\vdots$
$\mathcal{V}_{n}$	R	G	$\cdots$	T

TABLE IV: Structure of the data

Fig. 1: Pipeline for genotype to tensor image transformation. (a) Original genotype represented as a sequence of characters. (b) Conversion of the genotype sequence into a 2D-image. (c) Transformation of the 2D-image into a tensor representation.

Fig. 2: Biological interaction among different SNPs.

	2D-Image Representation	Tensor Representation
	$\mathcal{S}_{2D}\times\mathcal{S}_{2D}=d$	$\mathcal{S_{T}}\times\mathcal{S_{T}}\times\mathcal{C}=d$
	$\mathcal{S}_{2D}=\sqrt{d}$	$\mathcal{S_{T}}=\sqrt{\frac{d}{\mathcal{C}}}$
	$L_{2D}\approx\frac{\mathcal{S}_{2D}}{k}=\frac{\sqrt{d}}{k}$	$L_{T}\approx\frac{\mathcal{S_{T}}}{k}=\frac{\sqrt{d}}{k\sqrt{\mathcal{C}}}$
	Ratio: $\dfrac{L_{2D}}{L_{T}}=\sqrt{\mathcal{C}}$

TABLE V: Comparison of CNN layer requirements for 2D image and tensor representations

Fig. 3: ResNet-18 architecture.

Panel of Figure 9 — Fig. 4: Model Performance on Soybean Dataset.

Dataset	Trait	DLGWAS	DNNGP	GPFormer
Soybean	PH	BS:64, LR:0.001, D:0.3	BS:32, LR:0.01, D:0.1	BS:32, LR:0.001, D:0.3
	NN	BS:32, LR:0.001, D:0.3	BS:32, LR:0.01, D:0.3	BS:64, LR:0.001, D:0.3
	GY	BS:32, LR:0.001, D:0.3	BS:32, LR:0.01, D:0.3	BS:32, LR:0.001, D:0.3
	CT	BS:64, LR:0.001, D:0.3	BS:32, LR:0.01, D:0.1	BS:32, LR:0.001, D:0.3
Rice	PH	BS:64, LR:0.001, D:0.3	BS:64, LR:0.01, D:0.3	BS:64, LR:0.001, D:0.3
	FT	BS:64, LR:0.001, D:0.1	BS:64, LR:0.01, D:0.3	BS:64, LR:0.001, D:0.1

TABLE VI: Optimal Batch Size (BS), Learning Rate (LR), and Dropout (D) configurations for DLGWAS, DNNGP, and GPFormer

Dataset	Trait	ResGene-2D	ResGene-T
Soybean	PH	BS:32, LR:0.001, D:0.3	BS:64, LR:0.001, D:0.1, $\mathcal{C}$ :20
	NN	BS:64, LR:0.001, D:0.3	BS:64, LR:0.001, D:0.1, $\mathcal{C}$ :50
	GY	BS:64, LR:0.001, D:0.1	BS:64, LR:0.001, D:0.3, $\mathcal{C}$ :50
	CT	BS:64, LR:0.001, D:0.3	BS:64, LR:0.001, D:0.3, $\mathcal{C}$ :20
Rice	PH	BS:32, LR:0.001, D:0.3	BS:64, LR:0.001, D:0.1, $\mathcal{C}$ :50
	FT	BS:32, LR:0.001, D:0.3	BS:32, LR:0.001, D:0.3, $\mathcal{C}$ :50

TABLE VII: Optimal Batch Size (BS), Learning Rate (LR), Dropout (D), and Channel (

\mathcal{C}

) configurations for ResGene-2D, and ResGene-T

Panel of Figure 12 — Fig. 5: Model Performance on Rice Dataset.

Panel of Figure 13 — Fig. 6: Model Performance on Sorghum Dataset.

Dataset	Trait	rrBLUP	BayesB	SVR	XGBoost	DLGWAS	DNNGP	GPFormer	ResGene-2D	ResGene-T
Soybean	PH	7	9	4	3	8	6	5	2	1
	NN	5	6	2	9	8	7	4	3	1
	GY	1	2	6	7	7	8	5	4	3
	CT	7	9	5	8	4	6	3	1	2
Rice	PH	8	7	6	3	2	9	4	5	1
	FT	8	6	5	2	3	7	9	4	1

TABLE VIII: Comparative ranking of models across datasets and traits

Explore similar work

AttnGen: Attention-Guided Saliency Learning for Interpretable Genomic Sequence Classification
May 13, 2026Rayhaneh Shabani Nia, Ali KarkehabadiNeural Network Interpretability Attention Mechanisms
University of California, Davis California, USA
JEPA-DNA: Grounding Genomic Foundation Models through Joint-Embedding Predictive Architectures
Feb 19, 2026Ariel Larey, Elay Dahan, Amit Bleiweiss +15Molecular Representation Learning Genomics
Applied AI Architecture, NVIDIA, Israel. · Worldwide Field Ops, NVIDIA, Israel. · Developer Programs, NVIDIA, Israel. +2
A Linear-Transformer Hybrid for SNP-Based Genotype-to-Phenotype Prediction in Grapevine
May 7, 2026Yibin Wang, Murukarthick Jayakodi, Silvas Kirubakaran +2Transformer
Department of Biological and Agricultural Engineering, Texas A&M AgriLife Research, Texas A&M University System, Dallas, TX, 75252, USA · Department of Soil and Crop Sciences, Texas A&M AgriLife Research, Texas A&M University System, Dallas, TX, 75252, USA · USDA-ARS, Grape Genetics Research Unit, 630 West North Street, Geneva, New York 14456, USA

May 13, 2026J/K move · Enter open · S save

AttnGen: Attention-Guided Saliency Learning for Interpretable Genomic Sequence Classification

Rayhaneh Shabani Nia, Ali Karkehabadi

University of California, Davis California, USA

Neural Network Interpretability Attention Mechanisms Feature Selection Genomics Sequence Modeling

Open paper AskPDF arXiv

Deep neural networks have achieved strong performance in genomic sequence classification; however, relating their predictions to biologically meaningful sequence patterns remains challenging. In this work, we present AttnGen, an attention-guided training framework that embeds interpretability directly into the optimization process. AttnGen computes nucleotide-level importance scores using an attention mechanism and progressively suppresses low-contribution positions during training. This encourages the model to focus its predictions on a compact set of informative regions while reducing reliance on noisy sequence elements. We evaluate AttnGen on the standardized demo_human_or_worm benchmark, a binary classification task over 200-nucleotide sequences. With moderate masking, AttnGen achieves a validation accuracy of 96.73%, outperforming a conventional CNN baseline with 95.83% accuracy, while also exhibiting faster convergence and improved training stability. To assess whether the learned importance scores reflect functionally relevant signal, we conduct perturbation-based analysis by removing high-saliency nucleotides. This causes accuracy to drop from 96.9% to near chance level on a 3,000-sequence evaluation set, indicating that the model relies on a relatively small subset of informative positions. Our analysis shows that masking 10--20% of positions provides the most favorable trade-off between predictive performance and interpretability. These results suggest that attention-guided masking not only improves classification performance but also reshapes how models distribute importance across sequence positions. Although this study focuses on short genomic sequences, the proposed approach may extend to more complex interpretable sequence modeling settings.

Abstract

Explore similar work

AttnGen: Attention-Guided Saliency Learning for Interpretable Genomic Sequence Classification

JEPA-DNA: Grounding Genomic Foundation Models through Joint-Embedding Predictive Architectures

A Linear-Transformer Hybrid for SNP-Based Genotype-to-Phenotype Prediction in Grapevine