Computational Molecular Biology 2025, Vol.15, No.6, 265-272 http://bioscipublisher.com/index.php/cmb 269 This not only ensures stability but also minimizes the risk of allergies as much as possible (Figure 2). Simulation data show that the overall immune-inducing potential is good. Although there is still a way to go before clinical application, the direction is clear (Kardani et al., 2020). Figure 2 Discontinuous B-cell epitopes predicted by ElliPro. (A-E): 3D representation of conformational or discontinuous epitopes of the most antigenic chimeric protein fromT. cruzi CL Brenner. Epitopes are shown as yellow surfaces, and the bulk of the protein is represented in grey sticks (Adopted from Rawal et al., 2021) 5.3 Cancer neoantigen vaccines: from prediction to clinical trials In the field of cancer vaccines, predicting individual-specific neoantigens is becoming increasingly realistic. By using AI to analyze tumor mutation sites, identifying which fragments might become "targets", and then combining structural modeling to confirm whether they can trigger immune attacks, this set of processes is no longer just theoretical. At present, some neoantigen vaccines screened out based on these algorithms have entered the clinical trial stage. Although the mutations of each patient are different, this personalized strategy does provide a breakthrough for the design of tumor vaccines (Guarra and Colombo, 2023). 6 Challenges, Limitations, and Validation Bottlenecks 6.1 Accuracy and generalizability issues in predictive models Calculating vaccine design is not as "automatic" as it seems. Often, the model works well on a certain type of pathogen at the beginning, but once the target is changed, its performance immediately drops. Especially when the data volume is small and the sources are diverse, many AI algorithms will fall into the trap of overfitting. This situation is not uncommon - for some models, the longer they are trained, the worse their generalization ability becomes. Not to mention that the existing data report formats are diverse. Many times, even the connection between different data becomes a problem, let alone expecting to extract from them which factors are truly meaningful for immune protection (Dalsass et al., 2019; Bravi, 2024). In other words, a model is not a universal key. The more complex the scene is, the more tailor-made algorithms and features selection skills are needed. 6.2 The necessity of experimental validation: bridging the computational-experimental gap No matter how accurate the calculation is, it still has to pass the test. Many epitope predictions that seemed "promising" ultimately failed to pass the in vivo and in vitro experimental test. Between calculation and reality, it's not something that can be easily bridged with just a few sets of data. This also makes experimental verification the most time-consuming yet indispensable part of the entire process. Although some high-throughput technologies and community standard testing platforms have been established, to be honest, the investment cost is still high and they cannot completely replace traditional verification methods. Unless computational prediction, immune experiments and clinical research are integrated, the entire process can truly succeed. Otherwise, no matter how intelligent AI is, it can only remain at the "hypothesis" stage (Hashim and Dimier-Poisson, 2025).
RkJQdWJsaXNoZXIy MjQ4ODYzNA==