Abstract Study question How powerful is the combination of artificial neural networks that combine embryo's protein profile with its automatic score provided by time-lapse videos to predict ploidy? Summary answer An artificial neural network (ANN) that considers proteomics and automatic score assigned by deep learning achieves 71% accuracy in distinguishing between euploid and aneuploid embryos. What is known already Currently the most widely used technique for detecting chromosomal abnormalities involves biopsy of the developing embryo. However, it has several disadvantages related to invasiveness, technical difficulty, high economic costs, etc. Therefore, different non-invasive techniques are being studied for the detection of aneuploid embryos. The discovery of cell-free DNA (cfDNA) released by the embryo to the culture media during its development marked the beginning of a new era of noninvasive PGT (niPGT) but some factors require adaptation for the analysis. Also, artificial Intelligence (AI) represents a valuable alternative to developing new models for predicting PGT outcomes without disturbing the embryo. Study design, size, duration This study included 294 samples of culture medium from 81 treatments of the PGT-A program. Out of the total, 23 were control samples (medium in which no embryos had been cultured) and 271 were samples where there was a developing embryo. Embryos were cultured until the blastocyst stage in EmbryoScope systems (Vitrolife, Sweden) with single-step medium (Gems, Genea) and automatically scored by iDAScore v2 algorithm from 1 to 9.9. Participants/materials, setting, methods The spent culture medium was collected on day 5/6 of embryo development and chromosome analysis was performed using next-generation sequence technology (Juno Genetics, Valencia). The relative concentrations of 92 proteins were analysed using Proseek Multiplex Assays (Olink Bioscience) using 1 μl of each sample. The final assay readout was presented as Normalized Protein eXpression (NPX) values. Finally, we developed our own ANN algorithms considering protein profile and automatic embryo score. Main results and the role of chance For euploid embryos (n = 101), 35 protein samples analysed had different NPX values between conditioned and control media*. The relative concentration was reduced for 11 proteins (consumption by the embryo) and 24 proteins increased their levels (secretion). For aneuploid embryos (n = 170), 33 protein samples analysed had different NPX values between conditioned and control media*. The relative concentration was reduced for 4 proteins and 29 proteins increased their levels. Out of the total, only 6 proteins had on average different concentrations between normal and abnormal embryos*: MCP_1, IL_17A, CXCL1, IL18, IL_22RA1 and CSF_1. Additionally, the automatic embryo score provided by iDAScore v2 algorithm was higher for euploid embryos than for aneuploid embryos (5.9 ± 2.7 vs. 5.1 ± 2.6)*. For ploidy prediction, three architectures of ANN were developed with different input data ANN1 (six discriminatory proteins), ANN2 (automatic embryo score) and ANN3 (six discriminatory proteins and automatic embryo score). Our dataset was divided into 68% training, 16% validation and 16% test. The accuracy, sensibility and specificity for the test phase were as follows: 63.6%, 76.9% and 44% for ANN1; 61%, 11.8% and 95.8% for ANN2; and 71.1%, 62.5% and 77.3% for ANN3. Limitations, reasons for caution Only one laboratory by using single step culture medium from one brand was involved in this study. Wider implications of the findings Our study showed a new approach to avoid transferring aneuploid embryos in cases where embryo biopsy is not performed. In addition, further studies on this field may result in a new non-invasive methodology for detecting aneuploidy. Trial registration number PI21/00283
Read full abstract