Ancestral Genome Reconstruction Analysis Based on Artificial Intelligence and Evolutionary Algorithms
Data publikacji: 17 mar 2025
Otrzymano: 01 lis 2024
Przyjęty: 19 lut 2025
DOI: https://doi.org/10.2478/amns-2025-0833
Słowa kluczowe
© 2025 Minglu Zhao, published by Sciendo
This work is licensed under the Creative Commons Attribution 4.0 International License.
Ancestral genome reconstruction is a cornerstone of computational biology, providing critical insights into evolutionary dynamics and genomic adaptations across species. The reconstruction process requires highly sophisticated computational techniques due to the complexity of genomic data, including massive data volumes, incomplete sequences, and evolutionary variations. With the exponential growth of sequencing technologies, the demand for accurate, efficient, and scalable computational models has increased significantly. Existing efforts in ancestral genome reconstruction rely heavily on computational methods to predict genomic structures of extinct ancestors based on the genomic sequences of contemporary species. These predictions are vital for understanding mutation patterns, genome rearrangements, and evolutionary pathways. However, traditional methods often struggle with issues such as data sparsity, noise, and computational inefficiencies, especially when handling high-dimensional genomic datasets. As artificial intelligence (AI) and evolutionary algorithms have demonstrated immense potential in solving complex optimization problems, their application to ancestral genome reconstruction offers a promising avenue for addressing these challenges [1-2]. Leveraging AI’s deep learning capabilities for feature extraction and the adaptive search mechanisms of evolutionary algorithms creates opportunities to refine and enhance the reconstruction process [3-4]. This study aims to explore and evaluate the performance of such integrative approaches in the context of ancestral genome reconstruction [5-6].
Over the years, various methods have been developed for ancestral genome reconstruction, ranging from parsimony-based approaches to probabilistic models and phylogenetic algorithms. Parsimony-based methods aim to minimize evolutionary changes between genomes but often oversimplify the reconstruction process, leading to inaccurate results when applied to large, complex datasets. Probabilistic models, such as Bayesian inference, provide a more robust framework by incorporating evolutionary uncertainties, but their computational demands grow exponentially with dataset size. Similarly, phylogenetic algorithms, which reconstruct evolutionary relationships to infer ancestral genomes, face challenges in balancing accuracy and computational efficiency. Recent advances in machine learning have introduced deep learning models such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) into the domain. While these models excel in handling large datasets and extracting intricate patterns, they often require extensive training data and are susceptible to overfitting in scenarios with limited ancestral genome references [7-8]. Evolutionary algorithms, including genetic algorithms and particle swarm optimization, have also been utilized to optimize genome reconstruction. Despite their success in addressing specific aspects of the problem, such as parameter tuning, these algorithms often converge prematurely or lack the precision required for reconstructing high-dimensional genomic features [9-10]. These limitations underscore the need for a unified framework that combines the strengths of deep learning and evolutionary algorithms to overcome the shortcomings of existing methods [11-12].
In this study, we propose an innovative approach that integrates the Improved Whale Optimization Algorithm (IWOA) with Deep Belief Networks (DBN) to enhance the accuracy and efficiency of ancestral genome reconstruction. The IWOA algorithm introduces advanced mechanisms, such as nonlinear convergence, chaotic disturbance, and improved population initialization, to optimize the DBN’s initial parameters. By leveraging these improvements, the model addresses common issues in deep learning, such as poor initialization and local minima, which hinder performance in high-dimensional genomic data. The DBN component excels in extracting deep, hierarchical features from complex datasets, enabling the model to identify subtle evolutionary patterns and genomic structures. This integrative framework not only enhances the precision of reconstruction but also ensures computational efficiency, making it suitable for large-scale genomic analyses. Unlike existing methods, the IWOA-DBN model balances global and local search capabilities, ensuring robust optimization while maintaining high accuracy. Experimental results demonstrate that the proposed approach outperforms traditional methods in reconstruction accuracy, scalability, and reliability, establishing its potential as a state-of-the-art solution for ancestral genome reconstruction.
Literature[13] explores parsimony-based methods for ancestral genome reconstruction, which aim to minimize evolutionary changes when inferring ancestral states. However, these methods often oversimplify complex genomic variations such as insertions, deletions, and gene duplications, leading to reduced accuracy in large-scale datasets. To address these issues, Literature[14] introduces probabilistic models, particularly Bayesian inference, which incorporate evolutionary uncertainties and provide statistically robust reconstructions. Despite their advantages, these models suffer from high computational complexity, making them impractical for large genomic datasets. Literature[15] investigates phylogenetic tree-based reconstruction, which utilizes comparative genomics to infer ancestral genome structures. While this method offers biological relevance, it heavily depends on the accuracy of phylogenetic trees, which are often affected by incomplete or noisy genomic data.Literature[16] applies deep learning techniques, particularly convolutional neural networks (CNNs) and recurrent neural networks (RNNs), to genomic reconstruction. These models excel in identifying patterns from genomic sequences but require large annotated datasets and lack interpretability, limiting their application in ancestral genome research. Literature[17] attempts to mitigate these limitations by introducing hybrid deep learning models that integrate domain-specific constraints, yet the challenge of overfitting remains unresolved. To improve optimization in genome reconstruction, Literature[18] explores evolutionary algorithms such as genetic algorithms and particle swarm optimization. These methods enhance parameter tuning but often face issues of premature convergence, failing to balance global and local search strategies. Literature[19] proposes hybrid approaches that combine evolutionary algorithms with deep learning, showing improvements in reconstruction performance but still struggling with parameter initialization and search space exploration.Literature[20] investigates the Whale Optimization Algorithm (WOA) as an effective evolutionary optimization technique for genomic analysis. While WOA demonstrates strong global search capabilities, its standard version lacks adaptability to high-dimensional genomic structures. Literature[21] enhances WOA by incorporating nonlinear convergence mechanisms and chaotic disturbance strategies, significantly improving optimization performance in genomic applications. Inspired by these findings, this study proposes an improved hybrid model, IWOA-DBN, which integrates an optimized WOA with Deep Belief Networks to enhance ancestral genome reconstruction. By leveraging evolutionary optimization for parameter tuning and deep learning for hierarchical feature extraction, the proposed approach aims to improve reconstruction accuracy, computational efficiency, and model interpretability.
By revealing itself to be a probabilistic generation model upon dissection, a DBN exemplifies the characteristics of a conventional deep learning model. The input layer uses RBM to translate the data before sending the feature vectors from the nonlinear process to the hidden layers of the DBN. The network extracts characteristics from the data by continually mapping from the lower levels of the hidden layer to its higher layers. The RBM is the foundational component of the DBN model. The Boltzmann-like distribution of its samples characterizes this probabilistic model that is based on energy. Any possible statistical probability distribution may be modelled using energy. The strength of RBM lies in its ability to construct a learning model for the relevant data regardless of our knowledge of the statistical variables' internal distribution. Only bidirectional weights are used in the RBM model's connections between the visible and hidden layers. In an RBM model, which consists of two layers (the revealed and concealed layers), there are no link weights between units within the same layer. You can see the RBM structure in Figure 1.
RBM structure.
It is possible to determine the joint state energy function between every pair of neurons in the visible and hidden layers of the RBM by providing the layer vectors, weight, and bias of the RBM:
The RBM network topology achieves its best state for a given probability distribution when the energy function reaches its minimal value. The RBM model has to be trained endlessly to reach its ideal state, which is done by reducing the energy function. The combined probability distribution of the revealed and buried layers may be found. In RBM, there are restricted and autonomous linkages between any two units on distinct tiers. Given an input vector and an exposed layer vector, we can calculate the hidden layer's probability using the joint probability distribution formula. You may get the probability of the visible layer vector in the same manner when the input vector is the unit from the hidden layer.
As a superposition of many RBMs, DBN is limited to learning shallow features due to its one-layer network design. As a result, it is possible to stack many RBMs in order to extract deep features from complex data. These recovered deep features could be useful for classification and regression. However, data classification and regression cannot be accomplished only with RBM stacking. One must apply a classification or regression layer on top of the stacked RBM in order to get the full DBN model. Several RBMs and DBNs with several classification layers are used in the classification process. A vector of raw sample data is sent to the first RBM's visible layer. Learning then proceeds to establish the first RBM's visible and buried layer parameters to a fixed set. As it sets its settings, each RBM learns a new layer of knowledge; the input to the visible layer of the second RBM comes from the first RBM's hidden layer. The softmax function converts the features produced by the final RBM into category labels that correspond to the sample data. After the continuous RBM, the item layer takes over as the categorization layer. The more RBMs there are, the better the abstract features can be extracted from the sample data, and stacked RBMs stand for DBN. Another benefit of layer training is that it allows us to achieve near-perfect parameter matching between every pair of layers. The whole DBN requires more training due to the fact that the present model parameters are not ideal.
To begin, the structure of the network is first separated into several stacked two-layer networks. When running in an unsupervised environment, the greedy pre-training algorithm can only access data from the next two-layer network that has been partitioned. Its principal role is to facilitate communication between the exposed and hidden layers by serving as a map between the two. Next, the output from the hidden layer is utilized as input to the visible layer using the contrastive divergence approach. The error that results is then used to repair and fine-tune the internal parameters of the two-layer network, thereby concluding the pre-training phase. Lastly, the pre-training phase of the two-layer network is carried out layer by layer until the maximum number of layers is attained. The best answer for the parameters of a DBN model for a given set of sample data may be found by using supervised inverse fine-tuning. After determining the discrepancy between the actual and predicted results of the regression or classification layers, the DBN model is adjusted using the gradient descent technique. By using back-propagation and a top-down method, the parameters are adjusted via trial and error until they approach the best solution. The training phase of the DBN model is finished when the number of rounds of reverse fine-tuning approaches the maximum number of iterations.
Given that DBN employs initial weights and thresholds that are chosen at random, the network can fail to converge to the global optimum. Therefore, the WOA approach is used to identify the beginning parameters of this investigation. Humpback whales utilize a unique spiral bubble net to find food while they forage. After plunging to depths of about 15 meters, humpback whales corkscrew their way to the top of the sea. The last bubbles rise to the top at a pace almost identical to the first ones, forming a network of bubbles strong enough to surround the prey. Humpback whales may catch more food if they use this strategy, which involves diving under the school and quickly swimming to the center of the bubble net, where they hope to ingest it. Humpback whales will alternate between releasing bubbles and observing their prey in order to teach their fellow whales a hunting technique. The WOA idea was derived from the humpback whale's underwater hunting and predation behaviors. Surround, spiral bubble net, and global search are the three main stages of WOA. Similar to how whale swarms forage, the algorithm's exploration phase involves the swarm members following the leader and randomly searching for prey in a certain area. The whale swarm is given a dual strategy during the algorithm's mining phase. One is for the school of whales to get nearer to its prey as its members draw closer to the leader. Another tactic is to swim in a spiral pattern while releasing bubbles to lure in prey. This technique is called spiral bubble net hunting. By using global search, traditional whale optimization techniques may potentially attain global optimization performance. During the global search phase, the whale pod goes into a state of random movement as it can't determine the precise location of its prey. At now, whale pods find food by combining the whereabouts of a random selection of whales, who in turn update the next position.
Where
During the encirclement and predation phases of the whale pod's early feeding search, there is a lot of ambiguity since each whale moves in reaction to the leader's movements. As the number of prey rises, the whale pod's position is adjusted to coincide with the locations of the prey. The humpback whale has basically mastered the use of bubble nets as a means of locating food. Spiral bubble net hunting is a method wherein whales swim in a spiral pattern while spitting bubbles. This increases the likelihood of capturing food by drawing it closer to the top of the water. At its core, hunting is a two-part process that involves surrounding predation and spiral bubble net predation. To simulate the hunting behavior of whales that use both encircling and spiral bubble net techniques concurrently, a random variable is set up to randomly choose between the two position update strategies. In order to change its hunting stance when the probability surpasses a certain level, the whale pod employs the spiral bubble net predation strategy. As soon as the probability drops below a certain level, the whale aggregation changes its hunting stance to accommodate the local prey strategy.
When applied to difficult global optimization problems, WOA suffers from a slow convergence time and inaccurate solution accuracy. The design of IWOA has been greatly enhanced by this endeavor. Swarm intelligence uses a population-iteration based method, which is directly correlated to the search effectiveness of a high-quality beginning population. Applying the algorithm to a more diverse population from the outset may improve its global convergence speed and solution quality. Before using the WOA approach, no knowledge about the best solution to the optimization issue is available. Because of the reasons mentioned above, a random generation method is usually used to create the initial population. Nevertheless, stochastic approaches produce an initial population that lacks variety and has an unequal distribution in the solution space. Search efficiency suffers as a result of ineffective information extraction from the solution space. A nonlinear phenomena, chaos exhibits both randomness and regularity, as well as ergodicity. As a result, the chaotic sequence is used to create the starting population in this article. A great number of chaotic models may be used to produce chaotic sequences. For the purpose of population initialization, this work employs the Skew Tent mapping model to produce chaotic sequences:
Where
A combination of global and local searches is an essential component of any intelligent optimization algorithm that employs population evolution. Without proper coordination between these two types of actions, the evolutionary phase convergence speed of the algorithm might be slowed down or made to converge too rapidly. Expanding their search to include more of the world is the only way for companies to find anything. This prevents the algorithm from being trapped in a local optimum and keeps the population diverse. The swarm may accelerate the algorithm's convergence via local search by focusing on a smaller part of the solution space. Finding a way to integrate global and local search would greatly improve the performance of this new swarm intelligence optimization method. There is a strong correlation between the convergence factor and the WOA algorithm's ability to search both locally and globally. Adjusting the amount of the convergence factor is necessary to achieve an ideal equilibrium between the WOA algorithm's global search capacity and its local search ability. The convergence factor of WOA, on the other hand, falls linearly with time. In the end, the algorithm is satisfied enough to do a global search since it begins with a larger number, which enables it to select a wider search zone. With further development, the algorithm will eventually reach an improved state where a small convergence factor is sufficient for accurate local search. A strategy that relies on a linearly decreasing convergence factor is unable to capture the optimal search process because of the nonlinear changes that happen in the WOA approach during evolutionary search. This research outlines a nonlinear convergence factor method that takes its cues from the PSO algorithm's inertia weight setting:
Where
A flaw exists in the basic WOA algorithm, as it is in other swarm intelligence optimization methods. Particularly in the subsequent cycles, a local optimum might easily be trapped. In order to address this problem, this study applies chaotic disruption to the current optimal individual by drawing on the ergodicity, regularity, and unpredictable nature of chaotic sequences. This could trigger the WOA algorithm to jump out of its local optimum in an effort to enhance search capabilities and solution accuracy:
Where
Figure 2 shows the pipeline for this study, which aims to address DBN's shortcomings by optimizing IWOA's starting parameters to construct IWOA-DBN.
IWOA-DBN structure.
The procedure begins by setting the appropriate parameters of IWOA and the DBN to beginning values. The next step is to determine the IWOA fitness and then execute the appropriate IWOA operation. The initial weights and thresholds of the DBN are set using the parameter values after the termination criteria are satisfied. The last step is to train the DBN till the criteria is satisfied.
To train and test the model, relevant data is first collected. This study includes a total of 30,199 training samples and 18,902 test samples, with each sample consisting of eight ancestral genome reconstruction evaluation metrics, as detailed in Table 1. The corresponding labels represent the respective reconstruction evaluation levels.
Genomic reconstruction evaluation indicators.
| Index | Meaning |
|---|---|
| A | Structural Accuracy |
| B | Comparison Accuracy |
| C | Mutation Recovery Rate |
| D | Fragment Integrity |
| E | Functional Completeness |
| F | Mutability |
| G | Retention Rate |
| H | Sequence Consistency |
Training the IWOA-DBN model follows the crucial step of determining the necessary data samples used in this study. The effectiveness of the model's performance is highly dependent on the training process, which is closely linked to the assessment of its accuracy and generalization ability. A comprehensive analysis of the training process provides valuable insights into how well the model adapts to the data and improves over time. One of the key evaluation metrics in this study is the training loss, which is depicted in Figure 3. The training loss serves as a crucial indicator of how well the model minimizes errors over successive iterations. To provide a clearer view of the training process, an interval of 10 epochs is used to visualize the trend of training loss in this work. Initially, the training loss is relatively high, but as the model undergoes more iterations, it exhibits a steady decline, demonstrating the effectiveness of weight updates in optimizing the network. As the training progresses, the loss continues to decrease, indicating improved performance in reconstructing patterns and learning feature representations. The network gradually converges when training reaches 50 epochs, at which point the loss stabilizes, signifying that the model has effectively learned from the training data without further significant improvements. Notably, after the model reaches its convergence point, the final training loss stabilizes at approximately 0.5, demonstrating that the model has successfully minimized errors while maintaining its generalization capability. This pattern of loss reduction followed by stabilization reflects the robustness of the IWOA-DBN model in handling the dataset efficiently. Additionally, the convergence behavior indicates that the model avoids issues such as overfitting, ensuring a balanced performance across training and unseen data. The results further validate the efficacy of the IWOA-DBN in achieving reliable training outcomes, making it well-suited for practical applications requiring ancestral genome reconstruction assessment.
Training analysis.
To demonstrate the superiority of IWOA-DBN, this study conducts a comparative analysis with several other state-of-the-art approaches. A series of controlled experiments were designed to ensure a fair and objective comparison. Figure 4 presents the experimental results while maintaining consistency in parameter settings, thereby guaranteeing the credibility and reproducibility of the evaluation process. By keeping the hyperparameters, data input, and evaluation criteria uniform across all models, the study ensures that the observed differences in performance are genuinely attributable to the effectiveness of the respective methods rather than external variables. In terms of overall performance, IWOA-DBN consistently outperforms traditional approaches, exhibiting superior accuracy, robustness, and stability in assessing ancestral genome reconstruction. The enhanced performance can be attributed to the intelligent parameter optimization of IWOA, which refines the deep belief network (DBN) to achieve better convergence, reduced error rates, and improved generalization capability. The experimental findings reinforce that IWOA-DBN is a reliable and efficient tool for evaluating ancestral genome reconstruction. Its ability to extract meaningful patterns, adapt to complex genomic data, and minimize loss makes it a powerful choice in bioinformatics applications. This study provides compelling evidence that IWOA-DBN offers significant advantages over conventional methods, making it a valuable framework for future research and practical implementations in genomic sequence analysis and evolutionary studies.
Comparison with different methods.
In this study, the parameter control mechanism is implemented through the use of a nonlinear convergence factor, which plays a crucial role in optimizing the learning process of IWOA-DBN. By adjusting the convergence behavior dynamically, this factor helps the model achieve a more effective balance between exploration and exploitation during training. Figure 5 illustrates the impact of this approach by comparing the performance of IWOA-DBN with and without the nonlinear convergence factor. To rigorously assess its effectiveness, this study evaluates key performance metrics, including accuracy and F1 score, under both scenarios. The results demonstrate that incorporating the nonlinear convergence factor significantly enhances the model's ability to learn meaningful patterns, leading to a higher accuracy and improved F1 score compared to when a linear convergence factor is applied. This improvement indicates that the nonlinear approach effectively refines parameter adjustments, allowing the model to adapt dynamically throughout the training process. Moreover, the findings highlight the robustness and resilience of the nonlinear convergence factor. Unlike linear convergence, which may limit the model’s capacity to fine-tune parameters efficiently, the nonlinear factor ensures a more adaptive and flexible learning trajectory, preventing premature convergence and enabling better generalization. This validates the effectiveness of the nonlinear approach in improving the performance and stability of IWOA-DBN, making it a more reliable choice for complex tasks such as ancestral genome reconstruction assessment.
Comparison of convergence factor.
This study presents an advanced framework for evaluating ancestral genome reconstruction performance by integrating the Improved Whale Optimization Algorithm (IWOA) with Deep Belief Networks (DBN). By leveraging the global search capabilities of IWOA and the hierarchical feature extraction power of DBN, the proposed model effectively optimizes parameter initialization and improves reconstruction accuracy. The experimental results demonstrate that IWOA-DBN outperforms traditional methods in both computational efficiency and reliability, making it a promising approach for ancestral genome analysis. The model's ability to balance exploration and exploitation, combined with its adaptability to high-dimensional genomic data, highlights its potential for further applications in evolutionary biology and comparative genomics. However, challenges remain, including the need for improved interpretability of deep learning models and enhanced robustness in handling highly fragmented or noisy genomic data. Future research will focus on extending the framework by incorporating self-adaptive evolutionary mechanisms, hybridizing with transformer-based architectures for sequence modeling, and developing domain-specific loss functions to further refine reconstruction accuracy. Additionally, integrating multi-omics data, such as transcriptomics and epigenomics, could provide a more comprehensive perspective on ancestral genome evolution. These advancements will contribute to more precise and scalable ancestral genome reconstruction methods, further bridging the gap between computational biology and evolutionary genetics.