Sequence based virus host prediction: a curated dataset and generalizable framework for training artificial intelligence to identify viruses of humans

Abstract

Understanding how viruses evolve to infect specific hosts is crucial for reliably predicting and preventing emerging virus diseases. While genomic signatures of host adaptation exist within virus sequences, the systematic approaches needed to identify these patterns across the virome are in their infancy and will require well-curated datasets as a foundation for rigorous analysis. Although repositories like GenBank contain extensive genomic data, the critical metadata needed for artificial intelligence (AI) model training often exists in non-standardized, heterogeneous formats, complicating preprocessing, and hindering large-scale analysis across diverse virus families. To overcome these challenges and enable systematic investigation of genomic patterns, we built a dataset of 58 046 virus genomes spanning 15 families that represent diverse genome architectures and host ranges. Each sequence was classified for human host compatibility based on isolation source, creating a foundation for studying sequence-level determinants of host compatibility. We demonstrate the utility of this resource by applying AI through neural networks trained on k-mer frequency patterns and provide multiple analytical frameworks that enable generation of testable hypotheses about the genomic determinants of human-host compatibility. Finally, we apply the model to SARS-CoV-2 genomes to assess its ability to identify host-compatibility signals and characterize lineage- and time-associated shifts in k-mer usage. Using these neural networks to analyze k-mer frequency patterns across viral taxa, we found that sequence-based features can accurately predict human host compatibility and even generalize to some members of virus families not included in the training set. This performance variation suggests different forms of host-specific signal in virus genomes, likely reflecting distinct evolutionary pressures across virus groups. This resource can accelerate exploration of pan-virome lexicographical patterns that define host compatibility and will provide new avenues for identifying the genetic determinants of virus host range evolution.

Introduction

Viruses are masterful opportunists that must overcome a series of host-imposed challenges to infect, replicate, and disseminate. Successful viral reproduction hinges on two factors: exposure, the frequency of contact between viruses and host cells, and susceptibility, the host cell’s permissiveness to viral entry, replication, and dissemination. This requirement can be conceptualized through a lock and key model of host-specificity (Plowright et al. 2017), in which hosts present multiple “locks” (barriers) that viruses must unlock with specific “keys” (biological properties). Although some keys may open similar locks in different hosts, not every key guarantees successful replication. These keys become increasingly specialized as the cumulative set of host-specific obstacles grows, with the number of viable evolutionary solutions becoming progressively narrower with increasing host complexity. The evolutionary tug-of-war between host defenses and viral offenses leaves detectable imprints in virus genomes, which serve as blueprints encoding these host-specific keys—reflecting ongoing co-evolutionary adaptations shaped by the host environment (Pepin et al. 2010, Longdon et al. 2014, Rothenburg and Brennan 2020, Mollentze et al. 2021, Keenan et al. 2024). While these constrained, host-specific solutions are embedded within viral genetic material, their identification has remained elusive due to the extraordinarily high-dimensional spaces they occupy. Characterizing the set of solutions to human-host barriers is crucial not only for understanding the mechanisms driving host specificity, but also for anticipating zoonotic spillover events (Gardy and Loman 2018, Pandit et al. 2022).

Machine learning (ML) and deep learning, fields of study in artificial intelligence (AI), focus on the development of algorithms that learn from data. This provides a powerful framework for modeling these complex host-pathogen interactions and holds promise for advancing virus-host prediction (Pillai et al. 2022). Recent advances in natural language processing (NLP) and the development of large language models (LLMs) have revolutionized our ability to detect patterns in vast, high-dimensional data. Originally designed to understand and generate human language, LLMs excel at recognizing complex, context-dependent structures in such data. As adapted for genomics, these models have unlocked new perspectives on the “language” of genomes. For example, Ji et al. first showcased this potential with DNABERT (Ji et al. 2021), demonstrating how LLMs can identify methylation sites and delineate functional regions in DNA. More recent models have expanded these capabilities: HyenaDNA, a specialized long-range genomic model capable of processing up to 1 million nucleotides at single-base resolution, has proven highly effective at species classification from genomic sequences (Nguyen et al. 2023), while Metagene, a 7-billion-parameter foundation model trained on metagenomic wastewater data, shows promise in detecting pathogenic signals from DNA in environmental samples such as wastewater (Liu et al. 2025). With sufficient data, analogous approaches could one day reveal the genomic “laws” that govern host specificity.

While a wealth of genomic data is available in repositories such as GenBank, preparing the data for use in predictive models remains challenging. Problems such as class imbalance, limited virus diversity, incomplete or inconsistent metadata, and a lack of standardized annotation conventions introduces hurdles that can cause significant errors in downstream analyses (Whalen et al. 2022, Ritsch et al. 2023). While curated resources like GISAID (Shu and McCauley 2017) and the COVID-19 Datahub (Guidotti and Ardia 2020) provide high-quality, well-annotated data, they were not built to capture broad pan-virome sequence patterns.

To address this gap, we developed a virus-genome dataset specifically tailored for AI-based prediction of human host specificity. Here, we detail the dataset’s composition, the curation process, and the quality control measures implemented to ensure its representativeness. We then demonstrate its utility for human-host identification, emerging pathogen prediction, gene-level attribution of viral traits, and generating novel hypotheses related to virus evolution and host-pathogen dynamics.

We present this work with the understanding that our target audience encompasses experts in both statistical analyses and virologic principles. Our intent is to bridge these disciplines by providing sufficient background to enable experts in both domains to understand and make use of our approach and findings.

Materials and methods

Data collection

For our dataset, we selected 15 virus families based on 2 criteria: each family has viruses that infect humans, non-humans, or both, and each family has sufficient diversity (spanning multiple genera) to provide a comprehensive representation. We selected virus families as our highest taxonomic unit because they are specific enough to group viruses with shared evolutionary and biological characteristics, and broad enough to capture substantial diversity. In contrast, orders are overly broad and genera too narrow for our purposes. This initial collection, representing more than half of all known families containing human viruses (Siegel 2018), was assembled with the expectation that it will be expanded in future iterations. As detailed in Table 1 our starting point was 82 513 complete virus genomes obtained from the NCBI Virus database.

Table 1.

Initial sequence distribution across virus families.

Baltimore class	Virus family	Total	Human	Non-human
I (dsDNA)	Adenoviridae	2473	1873	600
	Orthoherpesviridae	2110	1276	834
	Papillomaviridae	4474	3665	809
	Polyomaviridae	2178	653	1525
	Poxviridae	3297	2460	837
II (ssDNA)	Parvoviridae	2405	502	1903
III (dsRNA)	Sedoreoviridae	9743	3398	6345
	Spinareoviridae	2822	83	2739
IV (+)ssRNA	Astroviridae	851	144	707
	Flaviviridae	11 433	6073	5360
	Picornaviridae	10 489	7754	2735
	Togaviridae	2227	969	1258
V ()ssRNA	Orthomyxoviridae	11 535	2999	8536
	Rhabdoviridae	4682	83	4599
VII (dsDNA-RT)	Hepadnaviridae	11 794	9038	2756
	Total	82 513	40 970	41 543

Data preparation and quality control

We developed a genomic data preparation pipeline using BioPython’s SeqIO module (Cock et al. 2009) to exclude incomplete or defective sequences from our dataset. The process employed stringent quality control to filter sequences, excluding entries marked as “partial,” “mutant,” “unverified,” bacterial artificial chromosomes (“bac”), “clone,” and those annotated solely as “gene” sequences. To maintain analytical consistency and minimize modeling artifacts, we only included genomes with standard IUPAC nucleotides (A, C, G, T, where T represents both thymine in DNA and uracil in RNA), processing each virus family independently to facilitate organized processing.

Artificial intelligence for host species labeling

Virus genomes in Genbank provide host information in various formats, ranging from standardized scientific nomenclature to informal descriptions and clinical notes. In many instances, no host information is provided. To address these challenges, we implemented a three-tier host classification system that progressively applies more sophisticated analysis methods to label our dataset (Fig. 1). A detailed summary of the filtering process can be found in Supplemental Table 4. We provide two sets of labels: (i) human or non-human virus and (ii) a standardized host label indicating the specific host source of isolation.

Figure 1.

Multi-tiered host species classification workflow for virus sequence metadata. The system processes input records through three increasingly sophisticated tiers: the simple case was resolved in Tier 1 using direct match and explicit scientific nomenclature (e.g. Homo sapiens from exact matches in host fields), an intermediate case was solved by Tier 2 using database mapping which normalizes non-standard terminology (e.g. mapping “26-day-old piglet” to Sus scrofa through curated species indicators), and complex cases were solved in Tier 3 using AI analysis, which employs Gemini Flash 1.5 to infer host species from complex contextual metadata (e.g. deriving H. sapiens from “asymptomatic carrier” in hepatitis B records). Each tier outputs standardized host information including species name, host category (e.g. mammal), and human/non-human classification.

Tier 1: string matching

Direct string matching was used to identify records where the host species is explicitly stated in standard scientific nomenclature (e.g. “Homo sapiens”) or where common standardized terms are present.

Tier 2: pattern recognition

Sequences not classified by Tier 1 are passed to Tier 2, which applies text cleaning and a regex-based lookup table to normalize non-standard formatting. After removing extraneous characters (e.g. brackets, parentheses), records are matched to a curated dictionary of species indicators. For example, “porcine,” “swine,” and “pig” map to Sus scrofa.

Tier 3: AI-driven analysis

Tier 3 addresses complex cases where host information not provided explicitly, but is implied within contextual metadata. For this, we used Gemini 1.5 Flash, an advanced multimodal LLM, to perform AI-driven analyses of sequence metadata (Team et al. 2024). As shown in Fig. 1, when presented with records lacking explicit host information, such as the complex case where only virus name (Hepatitis B) and isolation source (asymptomatic carrier) are available, the model leverages multiple layers of context. It considers virus type and known host ranges, clinical terminology patterns, and epidemiological context to infer host species. For example, “asymptomatic carrier” in Hepatitis B is recognized as a strong indicator of human origin because chronic asymptomatic infection characterizes human HBV cases. The detailed prompt engineering process and representative model outputs are described in Supplemental Material Section 3. To validate the model’s performance as detailed in Results (section 3.1), we manually evaluated 1000 AI-powered predictions using GenBank records, PubMed literature, and source metadata.

Feature extraction

While all genomes can be characterized through various features, k-mer frequency analysis provides a computationally efficient approach to capture latent sequence patterns. A k-mer is a substring of length k extracted from a nucleic acid or protein sequence, representing all possible subsequences of that fixed length found within the original sequence. K-mer features have been used in phylogenetic inference (Bernard et al. 2016, Bernard et al. 2019), association of virus-host properties (Garcia et al. 2021), and are a powerful approach to alignment-free comparative genomics (Zielezinski et al. 2017, Ren et al. 2018). This enables analysis of complex genomic sequence properties while avoiding the computational burden of traditional sequence alignment methods.

We generated separate feature representations using k-mer lengths from 3 to 8 for the forward strand only. Each k-mer generates a feature vector of dimension , where k is the k-mer length and features correspond to all possible subsequences composed of the four standard nucleotides from the IUPAC alphabet . This multi-scale design captures local nucleotide patterns at smaller k and longer-range motifs at larger k.

The frequency vector for a sequence of length is computed as:

where is the -th k-mer in lexicographic order, denotes the subsequence starting at position with length , and is the indicator function. Subsequences containing ambiguous nucleotides (“N”) are excluded to maintain feature consistency. This normalization enables comparable representations across sequences of varying lengths while preserving relative frequency information.

Data splitting

To prevent information leakage between related sequences while ensuring robust model evaluation, we developed a cluster-based data splitting approach (Fig. 2). Guided by empirical findings from DNABERT-I which demonstrated strong performance with 6-mer contexts (Ji et al. 2021), we adopted 6-mer frequency vectors as our starting approach to capture local sequence composition patterns in virus sequences. These high-dimensional vectors were projected into a 50-dimensional space using Uniform Manifold Approximation and Projection (UMAP), preserving both local and global sequence relationships (McInnes et al. 2020). The reduced representations were then clustered using Scikit-learn’s implementation of DBSCAN (Density-Based Spatial Clustering of Applications with Noise) with an epsilon value of 0.5 and a minimum number of five samples-per-cluster, identifying groups of similar sequences while marking outliers as noise (Ester et al. 1996, Pedregosa et al. 2011).

Figure 2.

Data preprocessing and splitting pipeline for genome analysis. The workflow begins with the extraction of -mer (length 6) frequencies, generating 4096-dimensional feature vectors from input sequences. Dimensionality reduction via UMAP transforms these vectors into a 50-dimensional space while preserving sequence relationships. DBSCAN clustering identifies sequence groups in this reduced space, which are then stratified by host type, virus family, and cluster membership to create balanced dataset splits (80% train, ; 10% validation, ; 10% test, ). The UMAP projection (right) is a 2D representation for visualization, showing the distribution of sequences and highlighting Hepadnaviridae as well as the separation of train (Cluster 1), validation (Cluster 2), and test (Cluster 3) sets. The final data is stored in both parquet format (combined train and validation set, ; separate test set, ) and as H5 files organized by -mer length ( through 8), each containing separate train, validation, and test datasets.

Using these UMAP-derived clusters, we partitioned sequences into training (80%), validation (10%), and test (10%) sets while maintaining three key constraints: (i) cluster integrity, where all sequences within a DBSCAN cluster are assigned to the same split, preventing similar sequences from appearing across different sets; (ii) balanced representation of virus families and host types across splits; and (iii) target split size ratios. The algorithm processes clusters in descending size order, assigning each cluster to the most under-represented split that maintains these constraints. This approach ensures each split contains a representative sample of sequence diversity while preventing potentially confounding sequence similarities from crossing split boundaries. Analyses of the data split and resulting clusters can be found in Supplemental Tables 5 and 6.

Machine learning methods

Neural network architecture

We employ a shallow feed-forward architecture with two hidden layers (64 and 32 units) using Gaussian Error Linear Unit activation functions (Fig. 3). Input vectors range from 64 (k-mer 3 model) to 65 536 (k-mer 8 model) dimensions, with the network learning increasingly abstract representations of sequence features as information flows through the layers. The final sigmoid output layer produces human host probability predictions, with values above 0.5 indicating human-infecting viruses.

Figure 3.

Architecture of neural network for virus-host prediction (k-mer 4 model depicted). The network takes k-mer frequency vectors as input (e.g. AAAA: 0.02, TTTT: 0.03) and processes them through two hidden layers (purple: 64 units, teal: 32 units). An expanded view of a single neuron unit (bottom) demonstrates the computational workflow: input weights (, , ...) are combined to form a weighted sum (), followed by Gaussian Error Linear Unit activation (, where is the normal CDF) to produce the activated output (2.12). The final sigmoid layer outputs a probability between 0 and 1, where values above 0.5 indicate human-infecting viruses. The relative height of hidden layers visually represents their decreasing dimensionality (64 32 units) as information flows through the network.

Based on preliminary experiments in which there was a tendency for the model to overfit by assigning extreme weight to specific k-mers, we employed L2 regularization (), which encourages the model to distribute predictive power across multiple features rather than relying on individual sequences.

We employed label smoothing () to address the inherent uncertainty of host assignment, as many viruses can infect multiple hosts or have imperfect annotation records. By preventing the model from becoming overconfident in its predictions, label smoothing improves generalization to novel virus sequences.

Dropout (rate = 0.3) and batch normalization work together to stabilize training and prevent gradient explosion, promoting robust feature learning. This is particularly important given the sparsity of k-mer frequency vectors when many features may be zero or near-zero for any given sequence.

These parameters were determined through Bayesian optimization using the Optuna framework, which searched the hyperparameter and model configuration space to find optimal values (Akiba et al. 2019). The model was trained using Adam optimization with cosine learning rate decay (initial lr = ) and early stopping (patience = 10) on validation loss. The cyclic learning rate schedule helps the model traverse multiple local optima in the loss landscape, which is critical given the high-dimensional input space and potential for convergence to suboptimal solutions.

Logistic regression classifier

We implemented logistic regression as a linear baseline classifier that models the probability of human host assignment using a sigmoid function applied to a weighted sum of k-mer frequency features:

where is the intercept term, is the vector of feature weights, and represents the standardized k-mer frequency vector. The model parameters are estimated by minimizing the following cross-entropy loss function with L2 regularization:

The L2 regularization term with coefficient prevents overfitting by penalizing large feature weights, which is particularly important given the high dimensionality of the k-mer space. We used the scikit-learn implementation with a maximum of 1000 iterations for convergence. Unlike the neural network, logistic regression cannot learn non-linear relationships between features, making it a useful baseline for quantifying the advantage more complex models have to offer in capturing nuanced sequence patterns that may be indicative of host tropism.

Random forest classifier

As a second baseline, we employed a random forest classifier, an ensemble method that constructs multiple decision trees during training and outputs the mode of their individual predictions (Breiman 2001). Each tree in the forest is trained on a bootstrap sample of the training data and considers only a random subset of features at each split:

where is the number of trees in our forest and is the prediction of an individual tree for input vector .

This approach offers several advantages as a baseline: (i) it naturally models non-linear relationships and feature interactions without explicit specification, (ii) it provides a measure of feature importance through the mean decrease in impurity across all trees, and (iii) it is relatively robust to overfitting compared to individual decision trees.

We allowed trees to grow to their full depth to capture complex patterns in the data, while the ensemble nature of the forest helps control variance. The random selection of features at each split (approximately features where is the feature dimensionality) ensures that the trees are decorrelated, improving the model’s generalization capability.

By comparing the performance of these simpler baselines against our neural network architecture, we can assess whether the additional complexity of the neural network is justified by meaningful performance improvements in virus host prediction.

Evaluation metrics

We selected the Matthews Correlation Coefficient (MCC) as our primary evaluation metric because it provides a conservative measure of performance that avoids overoptimistic biases inherent to metrics like accuracy, even in imbalanced datasets (Chicco and Jurman 2020). MCC captures the correlation between predicted and observed classifications and accounts for all categories of prediction outcomes: true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). Thus, the MCC provides a single measure that considers each predicted outcome simultaneously:

The coefficient ranges from −1 to +1, where +1 indicates perfect prediction, 0 represents random performance, and −1 indicates total disagreement between prediction and observation. Positive MCC values indicate performance better than random, with values closer to +1 representing stronger predictive power. Additional performance metrics including precision, recall, and F1 score were also calculated and are provided in Supplemental Table 7, with the best-performing k-mer configuration for each model summarized in Supplemental Table 8.

Embedding analysis and visualization

t-distributed stochastic neighbor embedding (t-SNE) is a widely used nonlinear dimensionality reduction technique that preserves local relationships between data points when projecting high-dimensional data into a lower-dimensional space (Maaten and Hinton 2008). To visualize and analyze the relationships between virus sequences, we employed t-SNE to generate low-dimensional projections of both the raw k-mer frequency vectors and the learned embeddings from the neural network’s final layer (sklearn.manifold.TSNE; n_components=2, perplexity=20, random_state=12041997, learning_rate=default, n_iter=default, init=default). This was done on the training dataset using scikit-learn (Fig. 4; analysis of the test dataset is presented in Supplemental Figure 2). The resulting visualizations were colored according to three different classification schemes: human/non-human host tropism, model prediction probabilities, and prediction uncertainty. The uncertainty for a prediction probability was calculated as:

Figure 4.

Neural network architecture and embedding visualization for virus sequence analysis. The network processes -mer frequencies as input through an input layer, followed by two hidden layers that learn increasingly abstract representations, culminating in a sigmoid output layer that produces human host probability predictions. Representative accession numbers (NC_006273.2, EU717214.1) demonstrate the processing pipeline, from raw -mer frequencies (e.g. AAAA: 0.02, TTTT: 0.03) through 32-dimensional embeddings ( through ) to final probabilistic predictions ().

where maximum uncertainty (1.0) occurs at and minimum uncertainty (0) at or .

Discriminative host probability analysis in betacoronaviruses

To evaluate the model’s ability to distinguish between human and non-human coronavirus sequences and to assess signals of host adaptation, we performed a distribution analysis of predicted host probabilities using two datasets:

1. Human SARS-CoV-2 sequences: A total of 8683 complete genomes (29–31 kb) with verified collection dates were obtained from NCBI GenBank. Collection dates span December 2019 to January 2025. For each month in this period, ~150 genomes were randomly retrieved. For December 2019 to March 2020, analysis was limited to the Wuhan-Hu-1 sequence, as it was the only complete genome available with a verified collection date.

2. Non-human betacoronavirus sequences: 1188 genomes from non-human isolated sources were collected from the NCBI Virus database.

Predicted host probabilities were generated by our neural network classifier and analyzed using Kernel density estimation with the Seaborn library (Waskom 2021). Density distributions were plotted separately for the human and non-human datasets, and statistical differences between these distributions were evaluated using the Mann–Whitney U test, the Kolmogorov–Smirnov test, and Cohen’s d effect size measure.

The Mann–Whitney U test, a non-parametric test that evaluates whether one distribution is stochastically greater than another (McKnight and Najab 2010), was calculated as:

where and are the sample sizes of the human and non-human distributions, respectively, and is the sum of ranks from the first sample. The P-value derived from this test indicates the probability of observing the difference between distributions by chance.

The Kolmogorov–Smirnov test, which evaluates differences in the entire probability distribution between two samples (Li and Vitányi 1990), was computed as:

where and are the empirical distribution functions of the human and non-human samples, and is the supremum of the set of distances.

To quantify the magnitude of difference between distributions, Cohen’s d effect size was calculated:

where and are the means of the human and non-human distributions, respectively, and are their standard deviations, and and are their sample sizes. Cohen’s d values above 0.8 indicate large effects, with our results showing a substantial difference between distributions.

Covid-19 human signal trend analysis

To investigate temporal shifts in human-host adaptation signals, we analyzed the SARS-CoV-2 dataset described above. For each month (with the exception of December 2019 to March 2020, where only the Wuhan-Hu-1 sequence was available), the ~150 retrieved genomes were divided into three independent batches.

For each genome, 4-mer frequency vectors were computed and processed through our pre-trained neural network classifier to generate human-host probability scores. Daily analyses were performed by calculating the mean probability and standard deviation across the three batches, and a centered 7-day rolling average was applied to smooth short-term fluctuations.

Temporal trends were quantified by aggregating daily means into monthly averages, which were then subjected to ordinary least squares linear regression. The regression slope coefficient, which represents the average monthly change in human-host probability for SARS-CoV-2, was then evaluated using a t-test to determine whether this temporal change was statistically significant.

Gene-level analysis of host prediction

Interpretable ML is essential for understanding how specific genomic regions drive host adaptation by quantifying the impact of individual sequence features. In our work, we adopted a leave-one-out ablation strategy, estimating each k-mer’s contribution by measuring the change in the predicted human probability when that feature is zeroed out. This straightforward approach provides an effective approximation of feature importance, though other methods, such as SHapley Additive exPlanations (SHAP), also offer formal frameworks for breaking down predictions into individual contributions and could be valuable for more detailed analyses in future work.

To map these approximate k-mer contributions to sequence positions, we employ an aggregation strategy where each position’s score is computed as the mean of the contributions of all 4-mers overlapping that position:

where is the set of 4-mers that include position , and is the ablation-derived contribution for k-mer . This produces a position-wise signal profile across the genome.

For gene-level analysis, we compute the average contribution for each gene :

where is the gene length. For visualization, we employ a zero-centered color intensity scale , defined as

which maps to RGB values as follows:

Here, enforcing ensures the visibility of even weak signals, resulting in a spectrum ranging from dark blue () to dark red (), with lighter shades indicating weaker signals.

This approach, while an approximation of true SHAP values, enables the investigation of specific genomic regions that contribute to host adaptation signals. Additionally, pairwise alignments between coronavirus sequences were performed with EMBOSS Needle (Madeira et al. 2024).

Results

Assembly of a virus dataset with high host-labeling accuracy

We constructed our dataset to support the search for pan-virome lexicographical patterns in genomic data that enable prediction of human-host specificity. The resulting dataset comprises 58 046 complete virus genomes representing 15 distinct virus families, with approximately equal distributions between human-associated (52.0%) and non-human-associated (48.0%) viruses (Table 2). Flaviviridae are represented by 9665 sequences, followed closely by Picornaviridae (9256) and Sedoreoviridae (8775). There is substantial variation in host specificity across families, reflecting both natural host range differences and sampling biases in sequence repositories. Sequence collections for some families are strongly biased towards human hosts, such as Hepadnaviridae (94.0% human-associated) and Papillomaviridae (80.6% human-associated), primarily due to their clinical significance and intensive surveillance in human populations. In contrast, Rhabdoviridae (98.7% non-human) and Spinareoviridae (97.2% non-human) are predominantly represented by viruses with non-human hosts, reflecting their broad host range across wildlife and the extensive sampling of agricultural and wildlife disease vectors. Several families show relatively balanced distributions between human and non-human hosts, including Orthoherpesviridae (52.8% human) and Poxviridae (57.8% human), indicating similar sampling efforts across hosts and/or naturally broad host ranges. In addition, Adenoviridae sequences are moderately balanced, though skewed toward human hosts (75.8% human).

Table 2.

Final sequence distribution across virus families.

Virus family	Total	Human	Non-human
Adenoviridae	2305	1747	558
Astroviridae	745	132	613
Flaviviridae	9665	2655	7010
Hepadnaviridae	8281	7780	501
Orthoherpesviridae	1192	629	563
Orthomyxoviridae	929	77	852
Papillomaviridae	3814	3073	741
Parvoviridae	2045	479	1566
Picornaviridae	9256	7253	2003
Polyomaviridae	2012	1447	565
Poxviridae	1445	835	610
Rhabdoviridae	3029	39	2990
Sedoreoviridae	8775	3151	5624
Spinareoviridae	2582	73	2509
Togaviridae	1971	811	1160
Total	58046	30181	27865

To assess the accuracy of the Gemini label assignments (e.g. human or non-human virus), we manually examined 1000 randomly selected sequences it had labeled. As detailed in Table 3, comparison with human labeling (using GenBank and associated publications) showed that Gemini correctly classified 972 (97.2%) sequences and misclassified only 28 (2.8%). The majority of misclassifications occurred within the enterovirus genus, primarily due to insufficient contextual metadata in GenBank entries. For instance, key information about host species and origins for some vaccine-derived polioviruses and enterovirus A are only accessible in associated publications, highlighting the importance of including detailed metadata in GenBank submissions.

Table 3.

Distribution of incorrect entries by genus.

Genus	Count
Enterovirus	11
Orthoflavivirus	4
Alphavirus	3
Alphainfluenzavirus	3
Rotavirus	2
Dependoparvovirus	1
Orthoreovirus	1
Orthopoxvirus	1
Betainfluenzavirus	1
Hepatovirus	1
Total	28

Having compiled this dataset, we examine its applications in human host prediction and outline methods for genomic analysis in the following sections. These approaches provide practical examples and potential starting points for further development of new computational methods for virus-host analysis.

Optimal k-mer length determination: neural networks outperform traditional models with 5-mers

We benchmarked several ML approaches by training three complementary models on k-mer frequency vectors for each sequence: logistic regression (provides a linear baseline), random forest (captures potential non-linear feature interactions), and a neural network (enables deep feature learning). This allowed us to quantify the effects of increasing model complexity with respect to k-mer length. In this comparison across k-mer lengths (k = 3 to 8) and model architectures (Table 4), the neural network consistently outperformed logistic regression and random forest baselines, with k = 5 yielding the highest performance (MCC = 0.820).

Table 4.

MCC scores across k-mer sizes and model architectures.

k-mer	Log regression	Random forest	Neural network
k3	0.723	0.759	0.810
k4	0.775	0.757	0.804
k5	0.767	0.741	0.820
k6	0.654	0.750	0.808
k7	0.669	0.746	0.779
k8	0.594	0.749	0.760

Bold indicates the highest MCC score for each model architecture.

The superior performance of 5-mers suggests this length captures relevant host-specific motifs while avoiding the sparsity issues that emerge at higher k values. Notably, performance degraded for k > 5 across all models, with MCC scores dropping from 0.820 at k = 5 to 0.760 at k = 8. This degradation was particularly pronounced in the logistic regression model, where MCC dropped sharply from 0.767 to 0.594. While the neural network maintained relatively robust performance, the trend suggests that increasing feature sparsity and the exponential growth in feature dimensionality () overwhelm the available training data despite our regularization strategies.

Learned human-host patterns allow for identification of human-viruses in viruses of families not seen during training

To assess the ability of a neural network to generalize human-host predictions to unseen viruses, we performed cross-family validation in which each virus family was held out (excluded from training) while a model was trained on the remaining families (leave-one-out strategy). We then generated human-host predictions on viruses of the held-out family. Based on the k-mer performances of the prior experiment, we performed this test on k-mers 3 to 6 and calculated the MCC for each neural network (Table 5).

Table 5.

MCC comparison across virus families, sorted by dataset size.

Family	k3	k4	k5	k6	Sequences (H/NH)
Flaviviridae	0.442	0.294	−0.027	0.049	9665 (2655/7010)
Picornaviridae	0.390	0.257	0.285	0.292	9256 (7253/2003)
Sedoreoviridae	0.248	0.068	0.276	0.289	8775 (3151/5624)
Hepadnaviridae	−0.320	0.341	0.738	0.828	8281 (7780/501)
Papillomaviridae	0.283	0.539	0.634	0.633	3814 (3073/741)
Rhabdoviridae	−0.020	−0.019	−0.012	−0.005	3029 (39/2990)
Spinareoviridae	−0.049	−0.002	−0.019	0.013	2582 (73/2509)
Adenoviridae	−0.144	−0.205	−0.050	0.000	2305 (1747/558)
Parvoviridae	0.212	0.308	0.099	0.311	2045 (479/1566)
Polyomaviridae	0.300	−0.072	−0.077	0.144	2012 (1447/565)
Togaviridae	−0.454	0.000	0.000	0.000	1971 (811/1160)
Poxviridae	0.436	−0.304	−0.155	−0.110	1445 (835/610)
Orthoherpesviridae	−0.257	−0.224	−0.032	0.155	1192 (629/563)
Orthomyxoviridae	0.059	0.101	0.000	−0.046	929 (77/852)
Astroviridae	0.181	0.013	0.126	0.047	745 (132/613)

For each virus family, the highest MCC value across k-mer sizes is shown in bold.

The highest MCC was achieved with a 6-mer profile for Hepadnaviridae, indicating that 6-mer features provide sufficient information to identify human hepadnaviruses. In contrast, performance markedly declined when relying on 3-mer profiles, indicating that the shorter context fails to capture essential human-host signals for these viruses.

One plausible explanation for this discrepancy is the genome organization of hepatitis B, the most prevalent human hepadnavirus in our dataset. Hepatitis B genomes are circular, compact, and characterized by multiple overlapping genes (Miller et al. 1989). We analyzed genome-wide 3-mer profiles with stride one that are frame mixed, meaning the starting position relative to codons varies across positions (first, second, or third nucleotide). Because two thirds of 3-mers are out of frame, these counts only partially reflect codon usage. In regions where genes overlap, a single nucleotide triplet can encode different amino acids in different frames, which further weakens codon-periodic signals. By contrast, longer k-mers, such as k of six, capture broader sequence context, including di-codons and short motifs, which may be less sensitive to frame mixing and overlap.

Several families—such as Togaviridae and Rhabdoviridae—consistently yielded low MCC values across all k-mer lengths, indicating difficulty in detecting human-host signals in these viruses. With many being vector-borne or zoonotic, it’s possible that these families employ infection mechanisms that differ markedly from the other families in the dataset. As a result, their genomic composition is likely to be more influenced by their natural reservoirs or arthropod vectors than by selective pressures within the human host. Thus, any human-adaptation “signature” might be minimal or altogether absent at the nucleotide level. Additionally, factors like small sample sizes, limited strain diversity, or recent emergence in humans could further hinder the detection of consistent host-related signals. These low MCC scores highlight the complexity and ongoing nature of virus evolution, emphasizing the need for further investigation into whether these families harbor truly human-specific genomic features.

Learned embeddings distinguish human-host viruses and correlate with virulence

To understand how the top-performing neural network, which was trained on 5-mer frequencies, distinguishes human viruses from non-human ones, we used t-SNE to visualize both the k-mer frequency vectors and the model’s learned sequence embeddings. By comparing these two representations, we can see how the model reorganizes the underlying data—transforming k-mer frequency into a more meaningful space. This re-encoding process groups together viruses with shared traits that point to human adaptation, while separating those that lack such features. As a result, we gain a clearer view of the model’s internal structure and the patterns it leverages to identify human viruses.

When t-SNE is applied to the k-mer frequency vectors, minimal clustering of human-host viruses is observed (Fig. 5A), suggesting these raw profiles do not strongly capture signals of human adaptation. In contrast, applying t-SNE to the learned sequence embeddings reveals regions enriched for sequences our model predicts as human versus non-human, with a transitional area between them (Fig. 5B). Within the predicted human region, distinct subgroups are visible rather than a single cloud. We interpret this exploratorily: the embedding may be organizing genomes along axes that correlate with host use, and—because genomes encode determinants of entry, replication, persistence, and immune evasion—such structure could reflect genome-encoded correlates of “infection strategies” (e.g. dinucleotide patterns, envelope status, replication site, motif scale), though we refrain from mechanistic claims from a 2D projection alone. Non-human-labeled sequences are relatively diffuse in the projection.

Figure 5.

t-SNE visualization comparing raw sequence features and learned embeddings of the training data (n = 44 722). (A) t-SNE of raw k-mer frequency vectors shows partial mixing of human (coral) and non-human (turquoise) genomes. (B) t-SNE of neural-network embeddings reveals regions enriched for sequences the model predicts as human versus non-human, with multiple distinct islands rather than a single cloud. (C) Coloring the raw k-mer projection by family shows family-aligned structure with substantial overlap. (D) Coloring the embedding projection by family shows more pronounced family-aligned groupings. Note: t-SNE is a 2D projection used for exploration; local neighborhoods are more reliable than global distances, and hyperparameters are given in Methods. Arrows indicate t-SNE axes. Interactive visualizations: https://hiyata.github.io/host-virus_dataset/.

These patterns become more interpretable when colored by family (Fig. 5C–D). The raw k-mer projection (Fig. 5C) shows some family structure with substantial overlap, whereas the embedding (Fig. 5D) reveals that several subgroups in the predicted human region largely align with specific families (e.g. Hepadnaviridae, Polyomaviridae, Spinareoviridae), while others (e.g. Flaviviridae, Picornaviridae) span both human- and non-human–enriched areas. This family-aligned organization is consistent with the network learning sequence features correlated with taxonomy and host association. We emphasize that t-SNE is used for exploration (local neighborhoods more reliable than global distances), and treat the apparent proximity of certain human-infecting families as a hypothesis-generating observation rather than a definitive map of infection mechanisms.

Further insight is gained by examining the predicted human-host probabilities overlaid on the embedding space. Viruses with extreme probabilities (near 0 or 1) form well-defined clusters (Fig. 6A), while those with scores around 0.5 occupy a transitional zone between human and non-human sequence representations (Fig. 6B). Notably, this transitional region is largely populated by virus species known to cross species barriers—such as rotavirus A, monkeypox, dengue, and West Nile virus. Consistent with the epidemiological dynamics enabled by their broad host ranges, the neural network only weakly classifies them as human viruses, implying that they are not optimized for human-to-human transmission.

Figure 6.

t-SNE visualization of model predictions and classification uncertainty. (A) Distribution of predicted human host probabilities across the embedding space, with colors ranging from 0.00 to 1.00. (B) Classification uncertainty is visualized as the absolute distance from probability 0.50, with lower values indicating greater uncertainty and higher values indicating greater certainty. Regions of high uncertainty identify potential transition zones between host-specific clusters, corresponding to sequences with possible zoonotic potential. Both visualizations use the same t-SNE coordinates, providing complementary views of the model’s classification behavior. Arrows indicate t-SNE axes. Interactive visualizations of this figure are available at https://hiyata.github.io/host-virus_dataset/.

In contrast, viruses that are highly virulent in humans, such as rabies and eastern equine encephalitis viruses, occupy regions of low predicted human adaptation (Leung et al. 2007, Corrin et al. 2021). Although virulence is not the only metric for human disease burden, here it serves as a useful proxy for severity. These observations collectively support the idea that the absence of human-adaptation features leads to greater incompatibility with human biology and, when infection occurs, a higher likelihood of fatality. Human-isolated rabies sequences further illustrate this trend (Fig. 7). Moreover, the observed gradient—viruses in the transitional zone (e.g. monkeypox and dengue) tend to cause epidemics, and those with minimal human adaptation (e.g. rabies) are linked to extreme fatality—is consistent with phylogenomic models of virulence evolution (Geoghegan and Holmes 2018).

Figure 7.

Host-independent clustering of lyssavirus sequences revealed by t-SNE sequence embeddings. Clustering patterns of lyssavirus sequences from both human (coral) and non-human (turquoise) hosts. Despite different host origins, all lyssaviruses cluster together in a distinct region of the embedding space (expanded view, right). Gray points represent sequences from other viruses. Arrows indicate the t-SNE axes.

These findings are further substantiated by comparing viruses by their predicted human-host probability from 5-mer frequency analysis with their reported fatality rates (Table 6). Viruses with high predicted human adaptation (e.g. human papillomavirus 30 and human cytomegalovirus) are associated with very low fatality rates, whereas those with low adaptation (e.g. rabies and eastern equine encephalitis viruses) have substantially higher fatality rates. This concordance between computational predictions and clinical outcomes reinforces the link between genomic adaptation features and virulence.

Table 6.

Viruses and their reported human-fatality rate sorted by human-host probability from 5-mer frequency.

Accession	Virus	Family	Fatality rate	Probability
NC_038889	Human papillomavirus 30	Papillomaviridae	|0.0001%	0.980
NC_006273	Human cytomegalovirus	Orthoherpesviridae	0.0001%	0.920
NC_001538	BK polyomavirus	Polyomaviridae	0.0001%	0.910
NC_003977	Hepatitis B virus	Hepadnaviridae	0.0001%	0.871
MG953831	Human bocavirus 2	Parvoviridae	0.0001%	0.730
NC_063383	Monkeypox virus	Poxviridae	3%–6%	0.680
KX010994	Yellow fever virus	Flaviviridae	20%–60%	0.550
NC_001802	HIV	Retroviridae a	0.0001%b	0.543
NC_045512	SARS-CoV-2	Coronaviridae a	1%	0.445
NC_001477	Dengue virus type 1	Flaviviridae	1%	0.350
NC_001475	Dengue virus type 3	Flaviviridae	1%	0.330
NC_002640	Dengue virus type 4	Flaviviridae	1%	0.320
NC_044855	Norovirus	Caliciviridae a	0.1%–0.001%	0.260
NC_075022	Venezuelan equine encephalitis virus	Flaviviridae	1%	0.214
NC_001474	Dengue virus type 2	Flaviviridae	1%	0.170
NC_001563	West Nile virus	Flaviviridae	3%–15%	0.168
NC_005062	Omsk hemorrhagic fever	Flaviviridae	0.5%–3%	0.168
NC_019843	MERS	Coronaviridae a	35%–36%	0.140
NC_002728	Nipah virus	Paramyxoviridae a	40%–75%	0.124
NC_001608	Marburg virus	Filoviridae a	50%	0.093
NC_006432	Sudan ebola virus	Filoviridae a	40%–75%	0.079
NC_004812	Macacine alphaherpesvirus 1	Orthoherpesviridae	100%	0.075
NC_003899	Eastern equine encephalitis virus	Flaviviridae	30%	0.064
NC_001542	Rabies virus	Rhabdoviridae	100%	0.063
NC_075802	Salmonid herpesvirus 2	Alloherpesviridae a	NAc	0.042
NC_079185	Vibrio phage	Chaseviridae a	NA	0.037
NC_079140	Gordonia phage Azira	Caudoviricetes a	NA	0.036
OR795895	Cucumber mosaic virus	Bromoviridae a	NA	0.035
NC_078671	Grapevine line pattern virus	Bromoviridae a	NA	0.034
NC_077680	Tomato mottle leaf curl virus	Geminiviridae a	NA	0.028
NC_004181	Colorado tick fever virus	Flaviviridae	0.2%	0.022

^aFamilies indicate no viruses from that family were included in the training dataset.

^bFatality rate for HIV refers to patients receiving modern antiretroviral therapy; untreated HIV infection has a significantly higher mortality rate.

^cNA indicates fatality rate data not applicable or not available. Fatality rates were collected from various sources, including the CDC, WHO, and peer-reviewed literature

Overall, while the learned representations might seem paradoxical at first, it seems clear that weaker signals of adaptation imply lower compatibility with human hosts, and increased virulence when infections occur.

Roseolovirus clustering: telomeric integration motifs and genomic signatures

The unusual clustering of human roseoloviruses (Human herpesviruses 6A, 6B, and 7) with non-human and viruses that are highly virulent in humans represents a striking exception to the general virulence pattern observed in our embeddings (Fig. 8). Despite being well-established human viruses that typically cause mild, self-limiting infections in immunocompetent hosts (Ablashi et al. 2014, Krug and Pellett 2022, Mori et al. 2022), our model persistently misclassifies these viruses as having low human adaptation. This counterintuitive result warranted deeper examination of their genomic signatures.

Figure 8.

t-SNE projection of human roseolovirus sequence embeddings localize with highly virulent, non-human viruses, in contrast with other human herpesviruses. Human cytomegalovirus (HCMV), Epstein–Barr virus (EBV), and Kaposi sarcoma herpesvirus (KSHV) form tight clusters distinct from other virus sequences. Human roseoloviruses (HHV-6A, HHV-6B, and HHV-7), represented by diamond symbols, cluster separately in a region dominated by non-human viruses (gray points). The expanded view (right) shows the detailed distribution of roseoloviruses among various non-human virus species, including parvoviruses, influenza viruses, and other animal-specific virus families. Yellow points indicate lyssaviruses from all hosts. Human alphaherpesviruses are absent here as they were assigned to test dataset. Arrows indicate the t-SNE axes.

Examining the viruses that cluster with roseoloviruses reveals that many of them share a common characteristic: the presence of various repetitive elements in their genomes. These repetitive elements often contain low-complexity k-mers, which are short DNA or RNA subsequences (of length k) characterized by a limited diversity of nucleotides, such as homopolymers (e.g. AAAAA), simple repeats (e.g. ATATAT), or other patterns with minimal variation. Low-complexity k-mers typically exhibit low Shannon entropy, a measure of information content, with entropy approaching zero for perfectly repetitive sequences, indicating minimal information and high predictability (Shannon 1948). These low-complexity regions create noise that can confound classification methods, as they occur frequently across diverse genomes without necessarily indicating evolutionary relatedness, and are often masked during sequence alignment and model training (Ren et al. 2018). Murine cytomegalovirus contains extensive microsatellite regions and repetitive gene families (Rawlinson et al. 1996). Fowl aviadenovirus has a “long repeat region” with 13 contiguous copies of a 135-bp sequence spanning ~1.8 kb (Ojkic and Nagy 2001). Parvoviruses in this cluster possess inverted terminal repeats (ITRs), such as the 145-nucleotide palindromic ITRs in adeno-associated virus (Cotmore et al. 2014). RNA viruses in this region also contain repetitive features: Eastern Equine Encephalitis virus terminates its genome with poly-adenosine tails (Lvov et al. 2015, chap. 8), orbiviruses like African horse sickness virus have identical terminal sequences in each genome segment (Mertens and Sangar 1985), and lyssaviruses contain poly-uridine tracts at gene junctions (Albertini et al. 2011). These observations suggest a possible link between repetitive sequence elements and the clustering pattern observed in the embedding space.

To investigate the possible link between repetitive sequence elements and the clustering pattern observed in the embedding space, we analyzed the k-mers with the strongest influence on HHV-6B’s classification according to ablation-derived scores (Table 7). If the clustering was primarily due to low-complexity sequences, we would expect the most influential k-mers to have low Shannon entropy values. However, examination of these entropy values challenges this hypothesis. Most of the influential k-mers for HHV-6B classification have moderate to high entropy (ranging from 1.3710 to 1.9219), with only TTTTT (entropy 0.0000) and TTTTC (entropy 0.7219) being true low-complexity sequences. Thus, the misclassification is not simply due to the presence of general low-complexity regions, but is driven by specific sequence motifs of biological significance.

Table 7.

Comparison of top 10 5-mer ablation-derived scores for HHV-6B, MDV, and HCMV with Shannon entropy.^a

K-mer	HHV-6B	MDV	HCMV	Shannon entropy
TAACC	−0.0034	−0.0017	0.0011	1.5219
TTTTT	−0.0026	−0.0009	−0.0005	0.0000
CGAAA	0.0025	0.0018	0.0028	1.3710
CCCTA	−0.0020	−0.0007	0.0003	1.3710
TTCGA	−0.0020	−0.0028	−0.0001	1.9219
ACCCT	−0.0020	−0.0006	0.0045	1.3710
CGATA	−0.0018	−0.0023	−0.0029	1.9219
TTTTC	−0.0018	−0.0007	−0.0004	0.7219
CCGAA	−0.0017	−0.0021	−0.0046	1.5219
TAACG	−0.0016	−0.0006	−0.0012	1.9219

^aBoldface marks the HHV-6B score and, per row, the comparator (MDV or HCMV) whose value is closest to the HHV-6B score (ties bold both).

Many of the top k-mers influencing HHV-6B classification are fragments or variants of telomeric repeat sequences (TTAGGG/CCCTAA) present at its genomic termini (Fig. 9). These repeats facilitate integration into host chromosome telomeres during latency (Hall et al. 2008, Kaufer et al. 2011). Strikingly, gallid alphaherpesvirus 2 (Marek’s disease virus, MDV) clusters near HHV-6B in the embedding space; it also contains telomeric repeats and integrates into host telomeres (Greco et al. 2014). Like most other herpesviruses, HCMV lacks these telomeric integration motifs and maintains its genome as an episome during latency (Sinclair and Sissons 2006).

Figure 9.

Relationship between top k-mers influencing HHV-6B classification and human telomeric repeat sequences. The table shows how many k-mers with strong negative ablation-derived scores correspond to fragments or variants of telomeric repeat motifs found at the genomic termini of HHV-6B and MDV. In sequence alignments, k-mers are shown in bold underlined text, with vertical bars (|) indicating junctions between adjacent repeat units. These telomeric repeat motifs facilitate integration of virus genomes into host chromosome telomeres and appear to be a key genomic signature driving the unusual clustering of roseoloviruses with non-human viruses in our analysis.

Reexamination of the ablation-derived scores in Table 7 reveals similarities between HHV-6B and MDV for these telomere-related k-mers. For example, the k-mer TAACC, which spans the junction between two complement telomeric repeats, has a strong negative influence on human-host classification for both HHV-6B (−0.0034) and MDV (−0.0017), contrasting a slightly positive value (0.0011) for HCMV. Similar patterns are observed for CCCTA and ACCCT, which represent fragments of the complementary telomeric repeat CCCTAA. These consistent negative scores indicate that the neural network learned to associate telomeric integration motifs with non-human viruses.

The unusual clustering of roseoloviruses with non-human viruses appears to be primarily driven by their telomeric integration strategy rather than by general low-complexity regions. While most well-adapted human viruses in our dataset (such as HCMV, HPV, and EBV) use other persistence strategies, such as episomal maintenance, telomeric integration represents a distinct and relatively uncommon genomic strategy among the human viruses present in our training data. The neural network identified this telomeric integration signature as a feature more commonly associated with non-human viruses, despite its functional importance in enabling persistent human infection by roseoloviruses (Arbuckle et al. 2010).

This reveals an important limitation in using k-mer frequencies alone for predicting host range: genomic features that serve similar functions in different host contexts may confound classification. In this case, telomeric integration motifs generate sequence patterns that our model associates with non-human viruses, despite their role in human infection for roseoloviruses. This suggests that distinct genomic strategies for host persistence can create similar sequence signatures across taxonomic boundaries.

While t-SNE visualization has limitations—including its stochastic nature, focus on local rather than global structure, and sensitivity to hyperparameters (Wattenberg et al. 2016)—the consistent roseolovirus clustering pattern across multiple runs provides compelling evidence that telomeric integration motifs are a dominant feature in the sequence embedding space. This result demonstrates that our neural network can identify shared genomic features across distantly related viruses, revealing unexpected biological connections that transcend traditional taxonomic classifications.

Neural network predictions successfully identify human coronaviruses

The analyses discussed to this point made use of our curated dataset, which did not include coronavirus genomes. Here, we assess real-world utility of genome-based predictions for early detection of emerging pathogens with epidemiological significance by asking the question: Can we predict human compatibility of coronaviruses based on gene patterns learned from unrelated viruses?

We thus extended our evaluations beyond our dataset to human SARS-CoV-2 and non-human betacoronaviruses. We retrieved 1188 betacoronavirus sequences from non-human hosts and 8683 human-derived SARS-CoV-2 sequences from NCBI Virus and generated 4-mer frequency vectors for each sequence. We then used our neural network to generate human-host probabilities for each sequence (Table 8) and plotted them as a density graph (Fig. 10).

Table 8.

Comparison of predicted probabilities: top 10 non-human coronaviruses viruses vs. human SARS CoV-2.

Species	Count	Avg Pred. Prob
SARS-CoV-2 (Human)	8856	0.5453
SARS-CoV-2 (non-human)	315	0.4429
Betacoronavirus Erinaceus	10	0.4099
Horseshoe bat sarbecovirus	31	0.3181
Sarbecovirus sp.	82	0.2881
Betacoronavirus sp.	26	0.2041
Rousettus bat coronavirus HKU9	10	0.1452
Middle East respiratory syndrome-related coronavirus (non-human)	384	0.1383
Rousettus bat coronavirus GCCDC1	10	0.1356
Betacoronavirus 1	220	0.1329
Murine coronavirus	11	0.0924

Bold indicates the highest average predicted probability of human host adaptation among the listed viruses.

Figure 10.

Density distribution of predicted probability of human-host for coronavirus sequences. The plot shows the distribution of human SARS-CoV-2 sequences and non-human betacoronaviruses based on their predicted probability of human host adaptation. The x-axis represents the predicted human host probability while the y-axis shows the density of sequences. Dashed vertical lines indicate the mean probability values for each distribution.

The predicted probabilities for human SARS-CoV-2 sequences form a sharp, concentrated peak around 0.56, indicating a weak but consistent human-host signature. In contrast, betacoronaviruses isolated from non-human hosts exhibit a broader distribution centered near 0.2. The size and shape of these distributions are significantly different (Mann–Whitney U test: ; Kolmogorov–Smirnov test: ), with a large effect size (Cohen’s ) indicating a statistically significant difference in the distributions of predicted probabilities, despite partial overlap due to bimodality in animal viruses. Notably, the non-human distribution also features a secondary peak in the 0.4–0.6 range, which largely represents SARS-CoV-2 sequences isolated from animals. This human/non-human overlap in probability distribution aligns with epidemiological observations that SARS-CoV-2 can transmit from humans to animals (Tan et al. 2022), reinforcing the concept that virus genomes exist along a continuum between strict host-adapted genotypes and phenotypes.

Remarkably, the neural network assigns SARS-CoV-2 sequences human-adaptation probabilities comparable to those of other zoonotic viruses. This suggests that although these viruses are not fully optimized for human-to-human transmission, they harbor genomic features that facilitate infection in humans—a nuance of considerable epidemiological importance. The clear separation between non-human betacoronavirinae and human-derived coronaviruses further demonstrates the model’s capacity to discern subtle host-specific signals in novel virus sequences, highlighting its potential utility in the early detection of emerging pathogens.

Quantifying human adaptation of SARS-CoV-2 from 2019 to 2024

Inspired by our success in identifying human betacoronaviruses, we examined how predicted human-host probabilities have changed over time in SARS-CoV-2 sequences. To do this, we analyzed SARS-CoV-2 sequences collected over a 5-year period, correlating each sequence’s collection date with its predicted human-host probability. Importantly, this analysis does not imply that the prediction for any single sequence changes over time; rather, it assesses whether more recent isolates—after 5 years of human circulation—exhibit stronger human-host signals. If the model has truly learned generalizable host features from the genomes of the virus families in our dataset, then SARS-CoV-2 sequences from later time points should have higher human predicted probabilities in comparison to those from the early outbreak period.

For this analysis, we organized our dataset into three randomized batches, each containing ~50 COVID-19 sequences per month from March 2020 to January 2025 (the 8683 sequences used in our previous assessment). The dataset includes only a single sequence from December 2019 (Wuhan-Hu-1); no other sequence data was available for the period between December 2019 and March 2020.

The predicted human-host probability gradually increased from December 2019 to December 2024, with an average monthly increase of 6.83 (Fig. 11). Starting with the initial Wuhan-Hu-1 strain from December 2019, which had a human host probability of 0.538, there were consistent evolutionary adaptation patterns across the three randomized cohorts.

Figure 11.

Temporal analysis of SARS-CoV-2 human host adaptation from December 2019 to December 2024. The plot shows the predicted probability of human host adaptation over time, with the y-axis representing predicted probability and the x-axis showing collection dates. Three independent batch analyses are shown. The Wuhan-Hu-1 reference strain (December 2019) is marked with a star. Shaded regions represent 95% confidence intervals for each batch. A linear regression analysis is indicated by a dashed line show an average probability increase of 6.83 × 10⁻³ per month. Variant bands (Alpha, Delta, Omicron) mark periods of global dominance (largest share of sequences) from OWID (GISAID EpiCoV data); annotations are contextual and non-causal.

Two key features characterize this progression. (i) Synchronized oscillatory patterns are readily visible in all cohorts, especially between March 2021 and September 2022, suggesting that global selective pressures may drive systematic evolutionary responses. (ii) Moreover, the predicted human-host probabilities markedly increased during the latter half of 2022, with average values peaking around 0.553 before transitioning into a more moderate upward trend. For context, we annotate approximate global dominance windows for Alpha, Delta, and Omicron—defined as periods when each lineage held the largest global share of sequenced genomes—based on aggregates from Our World in Data (GISAID, via CoVariants.org, processed by Our World in Data 2025). These labels are contextual and were not used to train the model or fit the trend.

The confidence intervals (shaded regions) indicate that prediction reliability was largely consistent over the study period, widening slightly during phases of rapid change. This increased variance coincided with the emergence of major variants of concern, underscoring the dynamic nature of virus evolution. Although the absolute change in human host probability might seem modest (ranging from ~0.515 to 0.595), the statistically significant positive trend () reinforces the notion that the virus was gradually adapting to human hosts.

Isolating coronavirus gene signatures allows for identification of changes that drive human predictions

To identify the genomic changes driving human adaptation in coronaviruses, we computed gene-level ablation-derived scores that quantify the impact of individual 4-mer features on our model’s human-host probability predictions. These scores quantify the impact of individual 4-mer features on our neural network’s predictions by measuring the change in predicted human-host probability when each feature is individually omitted.

We applied this strategy to three representative coronavirus genomes: bat coronavirus RaTG13 (2013), SARS-CoV-2 Wuhan-Hu-1 (2020), and a recent SARS-CoV-2 variant isolated from a human in 2024. Notably, RaTG13—the non-human virus sequence with the highest identity with SARS-CoV-2 known to date—serves as our non-human host baseline (Zhou et al. 2020). Our intention is not to imply a direct evolutionary descent from RaTG13 to Wuhan-Hu-1, but to use it as a comparative anchor to highlight the evolution of human adaptation signals.

As illustrated in Fig. 12, the RaTG13 genome has predominantly neutral to slightly negative scores, suggesting a scarcity of human-adaptive features (detailed scores for each gene are provided in Supplemental Table 9). In contrast, Wuhan-Hu-1 (December 2019) has increased positivity in multiple genes (notably ORF3a, E, M, and N), while ORF1ab presents a mixed profile indicative of an ongoing evolutionary transition. The 2024 variant, however, reveals a marked shift, with most genomic regions—especially ORF1ab—displaying positive scores. This progression demonstrates the model’s sensitivity to subtle shifts in k-mer frequencies that collectively drive a significant enhancement in human host compatibility.

Figure 12.

Comparative genomic analysis of coronavirus evolution and human host adaptation signals. Genome maps of three coronavirus strains are shown: bat coronavirus RaTG13 (July 2013), SARS-CoV-2 Wuhan-Hu-1 (December 2019), and SARS-CoV-2/human/USA/2545087/2024 (December 2024). The genomic organization is displayed with annotated open reading frames and structural genes. Importance scores, which indicate the contribution of sequence features to human host prediction, are represented by a color gradient from strong negative to medium positive. The x-axis shows genome position, with major genes labeled: ORF1ab (replicase polyprotein), S (spike), E (envelope), M (membrane), N (nucleocapsid), and accessory proteins [ORF3a, ORF6, ORF7a, ORF7b, ORF8, ORF10 (ORF10 is present in all three genomes; however, it was not annotated in the RaTG13 entry in GenBank and thus, not included in the visualization. The region maintains a human predicted probability)].

To further quantify the relationship between sequence divergence and the emergence of human adaptation signals, we performed pairwise alignments of the genomes and compared them with their predicted probabilities (Table 9).

Table 9.

Human-host probability and identity comparisons for coronavirus strains.^a

	RaTG13 (2013)	Wuhan-Hu-1 (2020)	SARS-CoV-2 (2024)
Human-host probability (%)	52.5	53.8	55.1
Pairwise comparisons
Comparison	Probability (%)	Sequence identity (%)	Identity difference (%)
2013 vs. 2020	+1.3	96.3	3.6
2020 vs. 2024	+1.3	99.0	1.0
2013 vs. 2024	+2.6	95.3	4.6

^aBold values highlight the highest predicted human-host probability and the largest differences.

Although the changes in human-host probability predictions are modest—rising from 52.5% in RaTG13 (2013) to 53.8% in Wuhan-Hu-1 (2020) and 55.1% in the 2024 variant—the impact of such changes on human-host predictions increases markedly with time in human hosts. For example, between RaTG13 and Wuhan-Hu-1, a 3.6% difference in sequence identity resulted in a 1.3% increase in human-host probability; between Wuhan-Hu-1 and the 2024 variant, a sequence identity difference of only 1% produced a similar increase (+1.3%). This suggests that once human-to-human transmission is established, even minimal mutations can amplify the human-adaptive signal captured by our model. The neural network’s ability to detect these subtle genomic changes underscores its potential as a tool for identifying the genetic markers of virus-host adaptation.

Discussion

Identification of lexicographical patterns in virus genomes

There exists a vast yet finite number of genome sequences capable of encoding viruses that can infect and replicate within human hosts. Collectively, these sequences represent the pan-virome genomic set of human-infective viruses, encompassing a wide spectrum of natural hosts and virulence potential. The key scientific challenge lies in identifying lexicographical patterns embedded within virus genomes that confer membership to this human-compatible set amidst the vast landscape of genomic data. To empower research aimed at addressing this challenge, we present a curated viral genome dataset spanning 15 virus families, designed explicitly to support research into sequence-level determinants of host specificity. By providing standardized genome sequences, host annotations, and carefully constructed dataset splits, we aim to enable the discovery of pan-viral patterns of host specificity. The dataset’s structure is particularly suited for both traditional bioinformatics approaches and emerging AI methodologies, especially those inspired by NLP, to investigate virus-host relationships.

Insights from initial computational analyses

Our initial analyses demonstrate how this dataset can be used to (i) explore genetic properties that define groups of viruses and (ii) generate hypotheses about virus evolution. Notably, the models detected host-adaptation signals even in families excluded from training, suggesting convergent genomic features underlying cross-species transmission. Intermediate prediction scores for zoonotic viruses highlight possible genomic correlates of host-jumping potential, while temporal trends in SARS-CoV-2 and correlations with virulence point to additional avenues for hypothesis-driven research. Although biological significance remains to be validated, these findings illustrate the potential value of the dataset for probing virus-host interactions.

Parallels between linguistic and virus evolution

NLP approaches applied to virus genomics offer analytical frameworks for characterizing host-specific patterns in virus genomes; the dataset we built and described here is structured and available to enable such computational investigations. Previous work has shown that language model approaches can reveal patterns in DNA sequences, from regulatory elements to functional domains. The emergence of AI algorithms capable of detecting sequence relationships within or between virus species creates new opportunities to search for pan-viral patterns of host adaptation. While our k-mer frequency analysis represents a relatively simple approach to sequence encoding, its utility suggests that even basic sequence features may contain meaningful host signals. More sophisticated language models that recognize sequence patterns as “biological grammar” could reveal deeper insights into host specificity patterns, decoding genetic information in a manner parallel to natural language syntax.

The parallel between language evolution and virus evolution offers an interesting perspective for understanding host adaptation patterns. Just as words evolve over time, accumulating changes while maintaining core meaning within specific linguistic environments (geographical regions, culture, mixing of dialects), virus genomes evolve within the constraints of their host environments. Etymology traces how words adapt to new contexts while retaining functional elements, analogous to how viruses preserve, modify, or eliminate sequence motifs while adapting to new hosts.

Limitations and future directions

Data and annotation limitations

The dataset provided here has several limitations that will inform future development directions. The binary human/non-human classification scheme enables systematic ML approaches but simplifies the complex reality of virus host ranges. Sequence repositories have substantial sampling biases favoring clinically relevant viruses, leading to disproportionate representation of human pathogens across numerous virus families. These biases create significant challenges for achieving balanced representation across virus and host species diversity, disproportionately affecting genomes with limited samples. Technical challenges also arise from the structure of genomic data and annotation practices, especially for segmented viruses within Orthomyxoviridae, creating potential data leakage.

Limitations of k-mer frequency analysis

Though k-mer frequency proved successful as an input feature for many viruses, important limitations remain. Genomes containing low complexity regions present particular classification challenges when using k-mer frequencies, exemplified by the misclassification of roseoloviruses. Nevertheless, the observed proximity of these misclassified viruses to MDV in the embedding space indicates genuine biological signal capture beyond random noise, highlighting the necessity for nuanced interpretation and virological domain knowledge when evaluating classification outcomes and genomic embeddings.

Directions for future development

To address the current limitations and enhance dataset utility, we identify four primary future development directions:

1. Taxonomic expansion: incorporate additional virus families and develop comprehensive multi-host annotations to capture the complexity of virus host ranges.

2. Methodological advancement: explore alternative sequence vectorization approaches and genome-specific data augmentation methods to address class imbalance and enhance model generalization.

3. Infrastructure development: develop a collaborative database system with community curation capabilities for regular updates and improved annotation quality control.

4. Standardization initiative: collaborate with national genomic databases to establish standardized formats and benchmarks for virus host prediction, including consensus protocols for sequence annotation and validation.

Genomes represent nature’s most fundamental language—a code orchestrating life across all biological domains on our planet.

Just as AI has demonstrated remarkable generalization capabilities across human languages, we stand at the threshold of a similar revolution in genomic interpretation. The dataset and methods presented here can inspire and facilitate new discoveries about lexicographical pan-virus human host signals. By systematically translating virus genome sequences into predictive frameworks, we envision a future where virus emergence is anticipated prospectively. This convergence of computational biology and ML promises transformative advancements in pandemic preparedness, creating early warning systems attuned to subtle genetic changes long before global health emergencies manifest.

Conflicts of interest

The authors declare no conflict of interest.

Funding

None declared.

Data availability

The data splits and vectorized dataset are publicly available in .h5 format at: https://github.com/hiyata/host-virus-dataset. The repository includes all data files, detailed documentation, and usage examples to facilitate integration into ML workflows. The trained model weights are available at: https://huggingface.co/hiyata, and the parquet format with metadata is available at https://huggingface.co/datasets/hiyata/Virus-Host-Genomes. A demonstration space to recreate and analyze FASTA sequences is accessible at: https://huggingface.co/spaces/hiyata/HostClassifier. A summary of available resources and dataset features can be found in Supplemental Tables 1 and 2. The distribution of genome types and host classifications is illustrated in Supplemental Figure 1, and the k-mer lengths selected for each experimental analysis are outlined in Supplemental Table 3. We encourage contributions from the community to enhance and expand the dataset.