Graph neural network-based system for metabolomic network analysis and embryo implantation prediction

Abstract

Provided are systems, methods, and computer-readable media for predicting embryo implantation from metabolomic data using graph neural networks (GNNs). Metabolomic data from in vitro fertilization (IVF)-relevant biospecimens are preprocessed and assembled into a heterogeneous or hypergraph comprising nodes representing at least metabolites and metabolic pathways and edges reflecting at least metabolic relationships. A GNN-based neural inference engine performs message passing and subgraph pooling to generate implantation likelihoods and pathway-level explanations. Data-driven and knowledge-based clustering reduces effective dimensionality and mitigates overfitting inherent in thousands of features for few-class outcomes. The system may further integrate with electronic health records and output interpretable network motifs enabling embryo ranking and protocol optimization.

Description

GRAPH NEURAL NETWORK-BASED SYSTEM FOR METABOLOMIC NETWORK ANALYSIS AND EMBRYO IMPLANTATION PREDICTION CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority of U.S. provisional Application No. 63 / 719,660, filed on November 13, 2024, the content of which is incorporated herein in its entirety by reference.FIELD

[0002] The invention relates to reproductive medicine, metabolomics, and artificial intelligence, and more particularly to graph-based deep learning systems that analyze metabolite interaction networks to predict embryo implantation outcomes in in vitro fertilization (IVF) and related settings.BACKGROUND

[0003] In IVF procedures, predicting embryo implantation success is crucial for improving fertility outcomes. Traditional assessment methods focus on morphological evaluation and genetic screening of embryos, which do not capture the full spectrum of factors influencing implantation. These non-invasive embryo assessments omit patient- and cycle-specific biochemical context available in biospecimens, such as follicular fluid, endometrial secretions, serum, urine, and spent embryo culture medium (SECM).

[0004] Incorporating such biochemical information has the potential to improve implantation prediction. However, metabolomic datasets typically quantify thousands of metabolites, while classification targets are binary or few-class. The resulting feature-to-class imbalance and high inter-metabolite correlation often lead to model overfitting. Vector-based models memorize cohort-specific noise and fail to generalize across instruments, labs, and populations.

[0005] Standard approaches (PCA / PLS-DA, SVMs, random forests) treat metabolites as independent features, ignoring chemical topology, stoichiometry, transport relationships, and pathway membership. Ad hoc feature selection reduces variance but discards mechanistic signalsand yields unstable biomarker lists.SUMMARY OF THE INVENTION

[0006] The present invention introduces a system that utilizes graph neural networks (GNNs) to predict embryo implantation success from biochemical graphs derived from metabolomic data. In the constructed graphs, metabolites are represented as nodes and biochemical relations are represented as typed edges, enabling the GNN to perform message passing over reaction, transport, and / or co-abundance structures. By pooling metabolite-level feature signals (e.g., metabolite concentration profiles and propagated messages between biochemically related metabolites) into biologically defined subgraphs derived from data-driven correlations and / or literature-validated metabolic pathways, the GNN imposes a form of topological regularization that constrains learning to physiologically meaningfill connections. This structured aggregation reduces the model’s effective degrees of freedom (DoFs) without arbitrary feature pruning, thereby mitigating overfitting in predicting implantation outcomes, a few-class biomedical classification tasks.

[0007] The system identifies key metabolic pathways related to fertility outcomes, providing fertility specialists with targeted, interpretable insights, enabling personalized treatment strategies and improving the success rates of IVF procedures.

[0008] Specifically, disclosed is a computer-implemented system for predicting embryo implantation. The system acquires metabolomic data from an IVF-relevant biospecimen, constructs a biochemical graph (e.g., a heterogeneous graph or hypeigraph capturing biochemical relations) based at least on the metabolomic data, applies a GNN with subgraph pooling operations to the biochemical graph to generate cluster embeddings representing metabolite clusters, from which an implantation probability is determined. The system may optionally perform model reliability enhancement processes, including but not limited to quality-controlled data preprocessing, uncertainty estimation and calibration, and counterfactual sensitivity analyses configured to identify clinician-modifiable levers.

[0009] In some embodiments, the metabolite clusters are formed via a hybrid clusteringapproach. A data-driven component identifies clusters through community detection at one or more stages, including during graph construction using static adjacency matrices and / or within the GNN-based inference engine using learned adjacencies. A knowledge-driven component defines pathway- or reaction-based groupings based on curated metabolic pathways and reaction ontologies during graph construction. Clustering reduces DoFs by aggregating thousands of raw metabolite features into a smaller set of biologically coherent cluster embeddings used for prediction and interpretation.

[0010] In some embodiments, a series of temporal graphs are constructed for multi-day SECM. In some embodiments, the system further integrates electronic health records (EHR) via HL7 / FHIR standards to enrich the input data.

[0011] In some embodiments, the system further incorporates domain adaptation, federated training, and / or privacy-preserving data management mechanisms to ensure robustness and compliance across multi-center clinical environments. Domain adaptation aligns embeddings across instruments or clinical sites using adversarial representation learning or correlation alignment at the pathway or cluster level. Federated training aggregates model updates from participating clinics via secure aggregation, optionally incorporating differential privacy on gradients.

[0012] The system may be applied to embryo selection, treatment optimization, and personalized interventions based on individual metabolic profiles. IVF clinics, diagnostic laboratories, fertility research centers, and vendors of non-invasive embryo assessment and metabolomic analytics may benefit from the results and analyses generated by the system to reveal insights into other reproductive assessments and metabolomic predictions.BRIEF DESCRIPTION OF THE DRAWING

[0013] For better understanding of the invention regarding the embodiments thereof, reference is made to the accompanying drawings, in which:

[0014] FIG. 1 shows a schematic block diagram illustrating an embodiment of the disclosedsystem for GNN-based implantation prediction. The diagram shows data flow from data acquisition through preprocessing, graph construction, neural inference, and implantation probability output with pathway-level explanations;

[0015] FIG. 2 shows an embodiment of a heterogeneous graph comprising metabolite and pathway nodes, wherein LDH denotes lactate dehydrogenase, and MCT1 denotes monocarboxylate transporter 1 ;

[0016] FIG.3 shows an embodiment of a hypergraph of bipartite representation;

[0017] FIG. 4 shows an embodiment of a reaction layer of a multiplex graph;

[0018] FIG. 5 shows an embodiment of a transport layer of a multiplex graph; and

[0019] FIG. 6 shows an embodiment of a co-abundance layer of a multiplex graph.DEFINITIONS

[0020] As used herein, the term “unit” refers to a functional entity configured to perform one or more specified operations. Each unit may be implemented as a hardware component, a software module executed by a dedicated or general-purpose processor, or a combination thereof.

[0021] The term “metabolite” refers to a chemical compound, typically a small molecule, that is produced, transformed, or consumed during a metabolic process. Metabolites include endogenous and exogenous compounds, as well as isotopologues, adducts, derivatives, or conjugates thereof. Metabolites can be detected or quantified using analytical techniques such as mass spectrometry (MS) or nuclear magnetic resonance (NMR) spectrometry.

[0022] The term “multiplex graph” means a graph with multiple edge types (e.g., reaction, transport, co-abundance, co-elution).

[0023] The term “hypergraph” means a graph where at least one edge connects more than two nodes (stoichiometric reaction modeling).

[0024] “Effective dimensionality” means the dimension of the learned representation used by the classifier after graph pooling and clustering.

[0025] “Calibrated probability” means a score whose empirical frequency matches predictedprobability within tolerance after calibration (e.g., temperature scaling, conformal).DETAILED DESCRIPTION

[0026] The system disclosed herein provides a comprehensive, graph-based framework to analyze metabolomic data of a female subject undergoing IVF and to predict embryo implantation outcomes. The system comprises a data acquisition unit or interface, an optional data preprocessing or harmonization unit, a graph construction unit, a neural inference engine comprising GNN layers configured to determine an implantation probability, an optional pathway identification module configured to produce pathway-level rationales for prediction results, an optional uncertainty / calibration unit for enhancing prediction reliability (e.g., by producing a calibrated probability or a prediction interval), and an output unit or interface configured to provide and / or present a predicted implantation likelihood, such as the implantation probability or prediction interval and preferably the explanatory rationales.

[0027] The system components can be designed flexibly to suit different operational needs. They may be integrated together within a single, unified housing to streamline workflow, improve data transfer speeds, and simplify maintenance.

[0028] Alternatively, the units / modules can be developed as separate, standalone parts that operate in coordination and communicate through secure networks or cloud-based platforms. Models may further be exported to ONNX for low-latency inference. This modular approach allows for greater scalability, easier updates, and customization, enabling parts of the system to be upgraded or replaced without disrupting the system’s workflow.

[0029] The system receives metabolomic data via a data acquisition unit or interface. Metabolomic data may be unprocessed or pre-processed and is derived from a biospecimen related to embryo implantation, including SECM (e.g., day 3 or 5 or 6 post-IVF), follicular fluid, endometrial fluid, serum, and urine. The unit may integrate with analytical techniques and platforms, such as LC-MS / MS (DDA / DIA), GC-MS, CE-MS, and 'H-NMR, which are capable of identifying metabolites as well as other biomolecules in the biospecimen. Various metadata areobtained during or after data acquisition, and may include embryo identifiers, embryo culture conditions, media lot, stimulation protocol, patient physiological parameters (e.g., age, body mass index (BMI), anti-Mullerian hormone (AMH) levels, pre-implantation genetic testing (PGT) status, endometrial parameters), luteal support, and implantation outcomes.

[0030] Prior to graph construction, the collected metabolomic data may be processed through the data preprocessing unit for identification, quantification, and other processing steps as needed. These processing steps may include standardization, log or asinh transformation, and normalization to minimize technical variability introduced during sample preparation or instrument use, ensuring that downstream analyses are both reliable and reproducible. Specifically, the system may perform peak detection, deconvolution, retention-time alignment, isotopologue collapsing, missing-value imputation within batch, and batch-effect correction (e.g., ComBat with run-order covariates).

[0031] Preferably, quality control (QC) is performed during data acquisition using internal standards and pooled QC injections. The data preprocessing unit may further conduct assessment and filtering of data based on quality control results, in addition to standardization, transformation, and normalization steps. For example, data failing pre-specified thresholds (e.g., coefficient of variation >20% for reference ions) may be excluded prior to downstream graph construction and analysis.

[0032] After optional processing, the metabolomic data is passed to the graph construction unit, which models the metabolomic information as a structured biochemical graph comprising nodes and edges.

[0033] The nodes include at least metabolites and metabolic pathways, and optionally include reactions, enzymes, transporters, and compartments (e.g., embryo, medium, serum) associated with the metabolic pathways. For example, the metabolites include glucose, pyruvate, lactate, citrate, glutamine, serine, glycine, arginine, tryptophan, carnitine, phosphatidylcholine, sphingomyelin, glutathione, folate, progesterone, and estradiol; and the pathways includeglycolysis, tricarboxylic acid (TCA) cycle, P-oxidation, amino acid metabolism, one-carbon / folate cycles, sphingolipid metabolism, oxidative phosphorylation, and steroidogenesis.

[0034] The edges represent at least metabolic relationships. Such relationships include: (i) biochemical reaction membership, such as reactant-product pairs in biochemical reactions (e.g., pyruvate + NADH «-> lactate + NAD+, where elevated lactate in SECM correlates with embryos that implant successfully); (ii) enzymatic catalysis, such as reactions catalyzed arginase (arginine ornithine + urea) and indoleamine 2,3-dioxygenase (tryptophan kynurenine), which correlate with implantation success through nitric oxide and immune tolerance pathways; (iii) transport associations, such as camitine-acylcamitine transport, which indicates mitochondrial -oxidation efficiency and where higher acylcamitine ratios are linked to better embryo viability; (iv) literature-curated pathway membership, such as metabolites in folate-methionine cycle and one-carbon metabolism metabolites (e.g., 5-MTHF, SAM, SAH), which are associated with higher implantation probability via methylation support; (v) empirical co-abundance, including partial correlations between metabolites (e.g., positive co-abundance between glutamine and aspartate (TCA anaplerosis) and between glycine and serine (one-carbon flux)) that reflect successful implantations; and (vi) co-elution / spectral similarity, such as metabolite pairs (e.g., sphingomyelins and phosphatidylcholines) co-eluted in LC-MS, where their increased levels and co-variation indicate active membrane remodeling linked to blastocyst expansion and implantation competence. These edges can be weighted to reflect the strength, confidence from databases and empirical statistics, and frequency of interactions, enabling the graphs to capture both qualitative and quantitative aspects of metabolite behavior.

[0035] The graph construction unit may construct biochemical graphs in various embodiments, for example, a heterogeneous graph (FIG. 2), a hypeigraph (FIG. 3), a multiplex graph (FIGS. 4-6), and a series of temporal graphs. In hypergraph embodiments, stoichiometric reactions may be encoded as hyperedges or via incidence matrices. In multiplex embodiments, edge types may be maintained as separate layers.

[0036] In temporal graph embodiments, time-indexed graphs may be constructed for SECM samples across developmental days, where snapshot and continuous-time models may be used to infer, approximate, and represent dynamics and changes.

[0037] In yet another embodiment, multi-omics fusion may be used to augment graphs with transcriptomic / proteomic nodes and cross-modal attention may be used to fuse modalities at the subgraph level. Similarly, lightweight distillation may be used to produce smaller student models for rapid, on-instrument triage.

[0038] Once the biochemical graph is constructed, the neural inference engine proceeds to analyze the data using deep learning techniques for graph structures. Specifically, the engine may integrate heterogenous GNN layers such as relational Graph Convolutional Networks (GCNs) with or without attention and / or DiffPool, Graph Attention Networks (GATs), hypergraph networks, and temporal attention for dynamic graphs to learn from the topology and features of the constructed graphs.

[0039] To further train, recognize, and refine patterns and correlations, the engine may employ self-supervised pretraining (masked-node / edge prediction, contrastive augmentations via edge dropout / feature masking / subgraph sampling) followed by supervised fine-tuning on labeled implantation outcomes. The engine may further employ loss function such as cross-entropy or focal loss and address class imbalance via reweighting to gauge the performance and accuracy of the models.

[0040] In some embodiments, the neural inference engine comprises a clustering module configured to identify metabolite clusters within the biochemical graph, and a pooling module configured to perform subgraph pooling to generate cluster embeddings representing the respective metabolite clusters. A “metabolite cluster” refers to a group of two or more metabolites exhibiting related biochemical functions, co-abundance, or network connectivity, which may be represented as a subgraph within the constructed biochemical graph. In some embodiments, the metabolite clusters are formed by detecting and grouping related metabolites within constructedgraphs based on topology and pathway similarities. For example, community detection algorithms (e.g., Leiden or Louvain clustering) may be applied to attention-weighted adjacency matrices generated by the neural inference engine. In other embodiments, clustering may alternatively or additionally be performed in a knowledge-driven manner by assigning metabolites to curated metabolic pathways or biochemical reactions during graph construction. Examples of metabolite clusters relevant to implantation include: (i) energy metabolism clusters, such as glucose, pyruvate, lactate, and citrate, representing glycolysis and TCA flux; (ii) amino acid clusters, such as glutamine, serine, glycine, alanine, and aspartate, which are linked to one-carbon and nitrogen balance; (iii) lipid clusters, such as phosphatidylcholines, sphingomyelins, and lysophosphatidic acid, reflecting membrane remodeling; (iv) oxidative stress clusters, such as glutathione, taurine, hypotaurine, and ascorbate, indicative of redox control; and (v) methylation clusters, such as methionine, S-adenosylmethionine (SAM), adenosylhomocysteine (SAH), and folate, capturing one-carbon and epigenetic readiness states.

[0041] The cluster embedding (also termed subgraph embeddings) may be obtained by subgraph pooling, such as differentiable pooling (e.g., DiffPool), attention-based pooling (e.g., ASAP), or set-based pooling (e.g., Set2Set). The subgraph pooling operation replaces raw feature vectors of individual metabolites with lower-dimensional representations, thereby reducing the effective dimensionality and mitigating overfitting risk in few-class prediction. In particular, the subgraph embeddings compress thousands of metabolite-level features into compact vectors that capture structural and functional context within the biochemical network. Unlike raw feature vectors, which record independent metabolite intensities, the subgraph embeddings encode topology-aware information, reflecting how metabolites are interconnected through biochemical reactions, transport, or correlations.

[0042] In addition, the neural inference engine may further generate interpretable explanations using interpretability techniques, such as GNNExplainer, PGM Explainer, SHAP, which are applied to nodes, edges, or subgraph embeddings.

[0043] In certain embodiments, the system further comprises an uncertainty / calibration unit configured to quantify and calibrate the prediction confidence of the neural inference engine. In some embodiments, the module performs uncertainty estimation using stochastic inference techniques, such as Monte Carlo dropout and / or deep ensembles, and subsequently applies probability calibration methods such as temperature scaling, isotonic regression, or conformal prediction to produce a calibrated probability and the corresponding prediction interval. The system may further comprise a counterfactual unit configured to perturb cluster- or metabolitelevel features and recompute outputs to estimate changes in the predicted implantation likelihood (A-probability) under clinician-modifiable levers, such as nutrient supplementation candidates, media refresh cadence, or hormone timing windows. The module may further perform provenance logging to trace perturbation sources and output variations. This counterfactual analysis capability aids in clinical decision-making by forecasting implantation outcomes for individual patients based on their unique metabolomic profiles.

[0044] Finally, a user interface presents at least the predicted implantation likelihood. In some embodiments, the user interface is configured to further present prediction-relevant information in an accessible format, including interactive and non-interactive formats, for clinicians. The predicted implantation likelihood comprises at least one of the implantation probability determined by the neural inference engine, a calibrated implantation probability, or a prediction interval. The prediction-relevant information comprises cluster / pathway contribution scores, and recommended what-if simulations, audit logs, data lineage, and model versions. The interface may provide visualizations of metabolite networks and highlight pathways influencing implantation. Predictive information and pathway-level insights may similarly be displayed to enable fertility specialists to make informed, personalized treatment decisions.

[0045] In some embodiments, the system processes metabolomic data from multiple biospecimens respectively corresponding to multiple embryos. The neural inference engine may generate a predicted implantation likelihood for each embryo based on the correspondingbiochemical graph analysis. The system may then rank the embryos according to the predicted implantation likelihoods and provide an embryo-selection suggestion based on the ranking. The embryo-selection suggestion may identify one or more embryos having the highest predicted implantation likelihoods as preferred candidates for embryo transfer.Example 1

[0046] Referring to FIG. 1, the metabolomic data for a biospecimen (101) such as SECM are collected by a data acquisition unit and processed by a data preprocessing unit before graph construction. The preprocessing unit performs processes such as normalization, missing-value imputation, and batch correction. A clustering module detects and groups related metabolites within the constructed biochemical graph based on topology and pathway similarities connecting metabolites by metabolic relationships. The neural inference engine (105) applies a graph neural network to the graph to infer an implantation likelihood. The output unit presents a calibrated implantation probability (110) and pathway-level explanations (111) for clinician interpretation.

Claims

WHAT IS CLAIMED IS:

1. A computer-implemented system for predicting embryo implantation, comprising:a data acquisition unit for receiving metabolomic data from a biospecimen related to embryo implantation;a graph construction unit for constructing a biochemical graph based at least on the metabolomic data, the biochemical graph comprising nodes representing at least metabolites and metabolic pathways and edges representing at least metabolic relationships;a neural inference engine for analyzing the biochemical graph to determine a likelihood of implantation;a clustering module operably coupled to at least one of the graph construction unit and the neural inference engine, and configured to form metabolite clusters within the biochemical graph, wherein a pooling module of the neural inference engine performs subgraph pooling to generate cluster embeddings corresponding to the metabolite clusters, and the neural inference engine aggregates at least the cluster embeddings to determine the likelihood of implantation; and an output unit for presenting at least the likelihood of implantation.

2. The system of claim 1, wherein the biospecimen includes at least one of spent embryo culture medium (SECM), follicular fluid, endometrial fluid, serum, or urine.

3. The system of claim 1, further comprising a data preprocessing unit for processing the metabolomic data, wherein the preprocessing unit performs data processing using at least one of peak detection, alignment, normalization, missing-value imputation, and batch-effect correction.

4. The system of claim 1, wherein the graph construction unit generates temporal graphs for multi-day SECM.

5. The system of claim 1 , wherein the nodes further include enzyme nodes and transporter nodes linked to metabolites to support target prioritization.

6. The system of claim 1, wherein the edges are weighted using curated database confidence scores and empirical statistics based on the metabolomic data.

7. The system of claim 1 , wherein the metabolic relationships include at least one of biochemical reaction membership, enzymatic catalysis, transport association, co-abundance, co-elution, and literature-curated relationships.

8. The system of claim 7, wherein the biochemical reaction membership is stoichiometric reaction membership and the reactions are encoded as hyperedges.

9. The system of claim 1 , wherein the biochemical graph is a multiplex graph with multiple edge types and weights.

10. The system of claim 1, wherein the neural inference engine comprises at least one deep learning algorithm to generate an implantation probability.

11. The system of claim 10, wherein the at least one deep learning algorithms includes at least one of a hypergraph neural network and a temporal graph attention network.

12. The system of claim 1 , wherein the neural inference engine is trained at least on implantation-labeled data and metabolic networks using self-supervised learning techniques.

13. The system of claim 12, wherein the at least one self-supervised learning technique includes masked objective and contrastive objective.

14. The system of claim 1, wherein the clustering module forms metabolite clusters by at least one of knowledge-drive assignment or data-driven community detection.

15. The system of claim 1, wherein the subgraph pooling includes at least one of differentiable pooling, hierarchical pooling, and set-based pooling.

16. The system of claim 1, further comprising an uncertainty / calibration unit for computing a prediction interval, wherein the uncertainty / calibration unit performs uncertainty estimation using at least one of deep ensembles and Monte Carlo dropout and applies calibration using temperature scaling, isotonic regression, or conformal prediction.

17. The system of claim 1, wherein the output unit is configured to further present an interpretation generated using at least one of GNNExplainer, probabilistic graphical-model explainers, and SHAP to attribute pathway and cluster contributions.

18. The system of claim 16, wherein the likelihood of implantation comprises at least one of an implantation probability and the prediction interval.

19. The system of claim 1, wherein the output unit is configured to further present predictionrelevant information, including data integrity, versioning, and lineage related to implantation.

20. The system of claim 1, further comprising a counterfactual unit for perturbing cluster- or metabolite-level features within predefined clinical bounds and estimating changes in the likelihood of implantation.

21. The system of claim 1 , wherein the system is further integrated with electronic health record systems via at least HL7 or FHIR to incorporate clinical covariates into prediction.

22. The system of claim 1 , wherein the system is further configured for federated model training for secure aggregation and differential privacy.

23. The system of claim 1 , wherein the system is further configured to perform domain adaptation using adversarial representation learning or correlation alignment to align pathway or cluster embeddings across instruments or sites.

24. A non-transitory computer-readable medium storing instructions that, when executed by a processing unit, cause the processing unit to predict embryo implantation, comprising:receiving metabolomic data from a biospecimen related to embryo implantation; preprocessing the metabolomic data;constructing a biochemical graph based at least on the metabolomic data, the biochemical graph comprising nodes representing at least metabolites and metabolic pathways and edges representing at least biochemical relationships;performing inference using a trained graph neural network (GNN) applied to the biochemical graph to determine a likelihood of implantation; wherein metabolite clusters are formed within the biochemical graph during the constructing step and / or during inference, and cluster embeddings corresponding to the metabolite clusters are generated via subgraph pooling and aggregated by the GNN to determine the likelihood of implantation; andoutputting at least the likelihood of implantation.

25. The medium of claim 24, wherein the constructing step further comprises constructing temporal graphs from multi-day SECM and applying temporal attention to capture metabolic dynamics.

26. The medium of claim 24, wherein the constructing step further comprises constructing multiplex graphs with multiple edge types and weights.

27. The medium of claim 24, wherein the metabolite clusters are formed by at least one of knowledge-driven assignment and data-driven community detection, wherein knowledge-driven assignment is performed during the constructing step while data-driven community detection is performed during the constructing step and / or during inference.

28. The medium of claim 24, wherein subgraph pooling comprises at least one of differentiable pooling, hierarchical pooling, and set-based pooling.

29. The medium of claim 24, further comprising instructions to perform uncertainty quantification using at least one of deep ensembles and Monte Carlo dropout and to perform calibration using at least one of temperature scaling, isotonic regression, and conformal prediction to produce a prediction interval.

30. The medium of claim 24, wherein the trained graph neural network is trained with at least self-supervised pretraining on metabolic networks, supervised fine-tuning on patient-level splits, and probability calibration.

31. The medium of claim 29, wherein the likelihood of implantation comprises at least one of an implantation probability and the prediction interval.

32. The medium of claim 24, further comprising instructions to perform domain adaption to harmonize embeddings with domain adaptation across instruments or sites.

33. The medium of claim 24, further comprising instructions to implement interpretability to attribute contribution scores to metabolite clusters and pathways.

34. The medium of claim 24, further comprising instructions to implement a counterfactualmodule to estimate changes in the likelihood of implantation under perturbations constrained to clinician-modifiable parameters.

Images