Missing observations are always a challenging problem that we have to deal with in diseases that require follow-up. In hospital records for vesicoureteral reflux (VUR) and recurrent urinary tract infection (rUTI), the number of complete cases is very low on demographic and clinical characteristics, laboratory findings, and imaging data. On the other hand, deep learning (DL) approaches can be used for highly missing observation scenarios with its own missing ratio algorithm. In this study, the effects of multiple imputation techniques MICE and FAMD on the performance of DL in the differential diagnosis were compared. The data of a retrospective cross-sectional study including 611 pediatric patients were evaluated (425 with VUR, 186 with rUTI, 26.65% missing ratio) in this research. CNTK and R 3.6.3 have been used for evaluating different models for 34 features (physical, laboratory, and imaging findings). In the differential diagnosis of VUR and rUTI, the best performance was obtained by deep learning with MICE algorithm with its values, respectively, 64.05% accuracy, 64.59% sensitivity, and 62.62% specificity. FAMD algorithm performed with , , and specificity was found out to be 61.00 with 3 principal components on missing imputation phase. DL-based approaches can evaluate datasets without doing preomit/impute missing values from datasets. Once DL method is used together with appropriate missing imputation techniques, it shows higher predictive performance.

Motivation: Large language models (LLMs) have acquired a remarkable level of knowledge through their initial training. However, they lack expertise in particular domains such as genomics. Variant annotation data, an important component of genomics, is crucial for interpreting and prioritizing disease-related variants among millions of variants identified by genetic sequencing. In our project, we aimed to improve LLM performance in genomics by adding variant annotation data to LLMs by retrieval-augmented generation (RAG) and fine-tuning techniques. Results: Using RAG, we successfully integrated 190 million highly accurate variant annotations, curated from five major annotation datasets and tools, into GPT-4o. This integration empowers users to query specific variants and receive accurate variant annotations and interpretations supported by advanced reasoning and language understanding capabilities of LLMs. Additionally, fine-tuning GPT-4 on variant annotation data also improved model performance in some annotation fields, although the accuracy across more fields remains suboptimal. Our model significantly improved the accessibility and efficiency of the variant interpretation process by leveraging LLM capabilities. Our project also revealed that RAG outperforms fine-tuning in factual knowledge injection in terms of data volume, accuracy, and cost-effectiveness. As a pioneering study for adding genomics knowledge to LLMs, our work paves the way for developing more comprehensive and informative genomics AI systems to support clinical diagnosis and research projects, and it demonstrates the potential of LLMs in specialized domains. Availability and implementation: We used publicly available datasets as detailed in the paper, which can be provided upon request.

Background. Next‐generation sequencing enables massively parallel processing, allowing lower cost than the other sequencing technologies. In the subsequent analysis with the NGS data, one of the major concerns is the reliability of variant calls. Although researchers can utilize raw quality scores of variant calling, they are forced to start the further analysis without any preevaluation of the quality scores. Method. We presented a machine learning approach for estimating quality scores of variant calls derived from BWA+GATK. We analyzed correlations between the quality score and these annotations, specifying informative annotations which were used as features to predict variant quality scores. To test the predictive models, we simulated 24 paired‐end Illumina sequencing reads with 30x coverage base. Also, twenty‐four human genome sequencing reads resulting from Illumina paired‐end sequencing with at least 30x coverage were secured from the Sequence Read Archive. Results. Using BWA+GATK, VCFs were derived from simulated and real sequencing reads. We observed that the prediction models learned by RFR outperformed other algorithms in both simulated and real data. The quality scores of variant calls were highly predictable from informative features of GATK Annotation Modules in the simulated human genome VCF data (R2: 96.7%, 94.4%, and 89.8% for RFR, MLR, and NNR, respectively). The robustness of the proposed data‐driven models was consistently maintained in the real human genome VCF data (R2: 97.8% and 96.5% for RFR and MLR, respectively).

Studies on hybrid data mining approach has been increasing in recent years. Hybrid data mining is defined as an effective combination of various data mining techniques to use the power of each technique and compensate for each other’s weaknesses. The purpose of this study is to present state-of-the-art data mining algorithms and applications and to propose a new hybrid data mining approach for classifying medical data. In addition, in the study, it was aimed to calculate performance metrics of data mining methods and to compare these metrics with the metrics obtained from the hybrid model. The study utilized simulated datasets produced on the basis of various scenarios and hepatitis dataset obtained from the UCI database. Supervised learning algorithms were used. In addition, hybrid models were created by combining these algorithms. In simulated datasets, it was observed that MCC values increased with a higher sample size and higher correlation between the independent variables. In addition, as the correlation between independent variables increased in imbalanced datasets, a noticeable increase was observed in the performance metrics of the group with lower sample size. A similar case was observed with the actual datasets.

Amaç: Makine öğrenimi (ML) ve derin öğrenme (DL) yöntemlerinin genetik ilişki çalışmalarında (GWAS) kullanımı giderek yaygınlaşmaktadır. Bu yöntemler, klasik GWAS metodolojisine ek olarak elde edilen sonuçların doğrulanmasına yardımcı olmaktadır. Bu nedenle; bu çalışmanın amacı, en yaygın kullanılan ML yöntemleri ve DL yönteminin farklı örnek genişliklerinde kıyaslanarak en doğru hasta-kontrol ayırımını yapan sınıflama yöntemini bulmaktır. Gereç ve Yöntemler: ML ve DL yöntemleri için farklı örnek genişliklerinde aynı prevelansa ve minör allel frekansına (MAF) sahip ve eşit sayıda hasta-kontrol içeren veriler üretilmiştir. Tüm veri setlerinde, tüm yöntemler için 10-katlı çapraz geçerlilik yöntemi uygulandı ve tüm yöntemler için, %70 eğitim veri seti %30 test veri seti olarak belirlenerek, modellerin geçerlilikleri ve tahmin güçleri sınandı. Simülasyonlar için PLINK yazılımı, ML için R programlama dili ve DL için de Python programlama dili Microsoft Azure GPU makinalar ile kullanmıştır. Bulgular: Örnek genişliği; N=200 için; en iyi sınıflama performanslarını ML yöntemleri içinde, Random Forest (RF) 0.73 Doğruluk (Acc.) ve CART 0.73 Acc. ile gösterirken, DL 0.60 Acc. değeri elde etmiştir. N=600 için; ML yöntemlerinden DVM 0.78 Acc. ile en iyi sınıflama başarısını elde ederken, DL 0.64 Acc. değeri elde etmiştir. Maksimum örnek genişliğinde, hem ML hem de DL için doğru sınıflama başarısının değişiminin etkilenmediği söylenebilir. Sonuç: Son yıllarda özellikle bağımsız değişkenler arasında ilişkilerin çok yüksek olduğu veri setlerinde yüksek doğru sınıflama oranı DVM ile edilmektedir. Bu çalışmada örneklem genişliğinden bağımsız olarak en yüksek sonuç DVM ile elde edilmiştir. Bu da genetik verilerde sık karşılaşılan doğrusal olmayan ML yöntemlerinin başarısını kanıtlamaktadır.