Method and system for predicting curative effect of polygonatum sibiricum extract on treating neurodegenerative diseases based on computer deep learning
By constructing a multimodal dataset and utilizing a computer deep learning model for cross-modal alignment and dose consistency constraints, the interpretability problem of predicting the efficacy of traditional Chinese medicine extracts was solved, enabling interpretable prediction of efficacy and personalized dosage recommendations for neurodegenerative diseases.
Patent Information
- Application Number
- CN202511604734.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-05
- Publication Date
- 2026-02-03
AI Technical Summary
In the existing technology, the efficacy prediction method of traditional Chinese medicine extracts for the treatment of neurodegenerative diseases fails to effectively combine pharmacokinetic parameters and efficacy characteristics, lacks interpretability, makes it difficult to identify key driving factors, and fails to meet the medical field's requirements for model transparency and verifiability.
By collecting multimodal datasets, filtering, imputing missing values, and temporal alignment are performed to generate trainable temporal multimodal feature sequences. Cross-modal contrast alignment is performed using computer deep learning models, dose consistency constraints are introduced, efficacy prediction is performed, and individualized dose recommendations are provided through interpretability explanation and counterfactual simulation.
It enables interpretable prediction of treatment efficacy for neurodegenerative diseases, identifies key driving factors, meets the transparency requirements of the medical field, and provides verifiable evidence for clinical decision-making.
Smart Images
Figure CN121460221A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of computer deep learning, and particularly relates to a method and system for predicting the curative effect of a Huangjing extract on treating neurodegenerative diseases based on computer deep learning. BACKGROUND
[0002] Neurodegenerative diseases (such as Alzheimer's disease and Parkinson's disease) are a class of chronic diseases characterized by progressive damage to neurons, and their incidence continues to rise as the population ages. However, there is still a lack of effective and radical treatment to slow down the progression of the disease. Huangjing, as a traditional Chinese medicine, has shown potential activity in anti-inflammatory, antioxidant and neuroprotective aspects, and has gradually become a research hotspot for the auxiliary treatment of neurodegenerative diseases. However, due to the complexity of the disease pathogenesis, significant differences in individual pathological characteristics and drug metabolism, the evaluation of the curative effect of Huangjing extract and its clinical transformation face many bottlenecks.
[0003] Currently, there are still obvious deficiencies in the curative effect prediction method of traditional Chinese medicine extract for treating neurodegenerative diseases: the existing prediction model often ignores the coupling relationship between pharmacokinetic parameters and curative effect characteristics, does not effectively constrain the dose consistency, and the deep learning model often lacks explainability due to the "black box" characteristics, which cannot clearly determine the internal relationship between Huangjing dose, drug metabolism and curative effect, and is difficult to identify key curative effect driving factors, and cannot meet the requirements of model transparency and verifiability in the medical field.
[0004] To this end, the present application proposes a curative effect prediction method and system for Huangjing extract in treating neurodegenerative diseases based on computer deep learning, to solve the above problems. SUMMARY
[0005] The purpose of the present application is to provide a curative effect prediction method and system for Huangjing extract in treating neurodegenerative diseases based on computer deep learning, to solve the problems raised in the background.
[0006] To achieve the above purpose, the present application provides the following technical solutions:
[0007] The curative effect prediction method for Huangjing extract in treating neurodegenerative diseases based on computer deep learning comprises:
[0008] Collecting time series and end point observation data related to Huangjing extract administration and the subject, to obtain a unified time-labeled multi-modal data set;
[0009] Filtering, missing value imputation, time series alignment and generating synthetic samples based on Huangjing pharmacokinetics and pathological model under biological constraints are performed on the multi-modal data set, to obtain a trainable time series multi-modal feature sequence;
[0010] extracting intra-modal representation from the time-series multi-modal feature sequence and aligning through cross-modal comparison to obtain a sample-level cross-modal semantic vector;
[0011] coupling the administration information and the pharmacokinetic parameters with the sample-level cross-modal semantic vector, and introducing a dose consistency constraint in the training to obtain model learning of dose response interpretable explanation;
[0012] regressing or classifying the model learning of dose response interpretable explanation with the comprehensive efficacy endpoint and outputting an efficacy prediction distribution to obtain an efficacy prediction result with a confidence interval;
[0013] performing interpretable explanation on the efficacy prediction result with a confidence interval and simulating a counterfactual-based administration scenario to obtain a corresponding individualized dose duration recommendation.
[0014] Preferably, the multi-modal data set comprises: pre- and post-administration of Huangjieling behavioral time-series data, Huangjieling administration information, blood concentration or pharmacokinetic curve, endpoint pathological quantification, and individual metadata;
[0015] The pre- and post-administration of Huangjieling behavioral time-series data includes escape latency, crossing platform times, target quadrant residence time, swimming trajectory and speed;
[0016] The Huangjieling administration information includes dose, administration frequency, administration duration, and preparation form;
[0017] The endpoint pathological quantification includes Aβ, tau, and inflammatory markers;
[0018] The individual metadata includes gender, age, and batch.
[0019] Preferably, the multi-modal data set is filtered, missing value is supplemented, time-series is aligned, and synthetic samples are generated based on Huangjieling pharmacokinetics and pathology models under biological constraints to obtain a trainable time-series multi-modal feature sequence, including:
[0020] Filtering and scale normalization are used for behavioral time series, and interpolation or statistical completion is used for missing time points;
[0021] Color normalization and patch segmentation are performed on image-based pathological data;
[0022] Batch effect correction is performed on omics data;
[0023] And based on the one-compartment, two-compartment pharmacokinetic model or the pathology progression model of Huangjieling, limited simulation samples are generated to expand the training set.
[0024] Preferably, the extracting intra-modal representation from the time-series multi-modal feature sequence and aligning through cross-modal comparison to obtain a sample-level cross-modal semantic vector comprises:
[0025] adopting time convolution network or Transformer for future window prediction task to extract time sequence representation for behavior time sequence;
[0026] adopting contrastive learning or autoencoder to extract visual representation for image;
[0027] adopting variational autoencoder to learn latent representation for omics or biochemical scalar;
[0028] aligning cross-modal vectors by InfoNCE or contrastive loss to obtain sample-level cross-modal semantic vector.
[0029] Preferably, the administration information and pharmacokinetic parameters are coupled with the sample-level cross-modal semantic vector, and a dose consistency constraint is introduced in the training to obtain model learning of dose response interpretable explanation, including:
[0030] The blood concentration-time curve or fitted pharmacokinetic parameters of Huangjing are spliced or attentionally fused with the sample-level cross-modal semantic vector; and a dose consistency regularization term is added in the training target, so that samples with similar Huangjing doses or similar PK characteristics have smaller distance in the representation space, thereby obtaining the model learning of dose response interpretable explanation.
[0031] Preferably, the model learning of dose response interpretable explanation is regressed or classified trained with a comprehensive efficacy endpoint and outputs an efficacy prediction distribution to obtain an efficacy prediction result with a confidence interval, including:
[0032] A joint loss function (including regression or classification loss and uncertainty related term) is used to train the model with a comprehensive efficacy score or a single endpoint as a supervised target; and a Bayesian regression head, MC-Dropout or deep Gaussian process is used to estimate the uncertainty of the output, and finally an efficacy prediction distribution and corresponding confidence interval are output.
[0033] Preferably, the efficacy prediction result with a confidence interval is explained for interpretability and a dosing scenario simulation based on counterfactuals to obtain a corresponding individualized dose duration recommendation, including:
[0034] Attention-based or SHAP feature importance analysis is applied to a single sample to identify key driving factors related to Huangjing efficacy;
[0035] And by changing the Huangjing dose or administration duration in the input (generating counterfactual PK curves) to input the training model to estimate the efficacy increment and uncertainty under different dosing scenarios, an individualized dose duration recommendation with a confidence interval is output.
[0036] Preferably, the method further comprises performance evaluation of the efficacy prediction model, the evaluation indicators including: RMSE, MAE, R 2 in the regression case, AUROC and accuracy in the classification case, and prediction calibration indicators (such as prediction interval coverage or CRPS), and multiple batches or external verification are used to prove the robustness and transferability of the Huangjing-based efficacy prediction model.
[0037] The computer deep learning-based efficacy prediction system for treating neurodegenerative diseases by using Huangjing extract includes:
[0038] A multi-source data collection module is configured to collect time series and endpoint observation data related to administration of Huangjing extract and a subject, and obtain a unified time-labeled multi-modal data set.
[0039] A time series preprocessing and data enhancement module is configured to filter, impute missing values, align time series, and generate synthetic samples based on Huangjing pharmacokinetics and pathology models under biological constraints, and obtain trainable time series multi-modal feature sequences.
[0040] A self-supervised representation learning module is configured to extract intra-modal representations from the time series multi-modal feature sequences and align them through cross-modal comparison, and obtain sample-level cross-modal semantic vectors.
[0041] A pharmacodynamic and pharmacokinetic coupling representation learning module is configured to couple the administration information and pharmacokinetic parameters with the sample-level cross-modal semantic vectors, and introduce a dose consistency constraint in the training, and obtain a model learned dose response interpretable explanation.
[0042] An uncertainty quantification module is configured to perform regression or classification training on the model learned dose response interpretable explanation with respect to a comprehensive efficacy endpoint and output an efficacy prediction distribution, and obtain an efficacy prediction result with a confidence interval.
[0043] A counterfactual dose optimization module is configured to perform interpretable explanation on the efficacy prediction result with a confidence interval and simulate a counterfactual-based administration scenario, and obtain a corresponding individualized dose duration recommendation.
[0044] Compared with the prior art, the present application has the following advantages:
[0045] (1) The present application integrates behavioral time series data before and after administration of Huangjing, pharmacokinetic parameters, pathological quantitative indicators, and individual metadata, and constructs a unified time-labeled multi-modal data set. Through filtering, time series alignment, and synthetic sample generation under biological constraints, the problems of noise interference, missing values, and sample scarcity in real data are effectively solved. The collaborative modeling of multi-modal features significantly improves the model's ability to capture complex disease mechanisms, and the prediction effect is better than that of a single modal model.
[0046] (2) The present application couples the pharmacokinetic parameters with the cross-modal semantic vector, and introduces a dose consistency regularization term, forcing the model to learn an interpretable explanation of the dose-response relationship. Samples with similar doses or similar pharmacokinetic characteristics are constrained to be adjacent in the representation space, so that the model prediction not only focuses on the therapeutic effect, but also explicitly reflects the influence mechanism of drug dosage. At the same time, through SHAP value or attention mechanism, the feature importance analysis of single sample can identify the key driving factors related to the therapeutic effect. It not only meets the strict requirements of the medical field for model transparency, but also provides verifiable decision basis for clinicians. BRIEF DESCRIPTION OF DRAWINGS
[0047] Figure 1 The present application is a computer deep learning-based efficacy prediction method for treating neurodegenerative diseases with Huangjing extract;
[0048] Figure 2 The present application is a water maze positioning navigation swimming trajectory diagram;
[0049] Figure 3 The present application is a mouse 5-day escape latency change diagram;
[0050] Figure 4 The present application is a diagram of the number of times each group of mice crosses the platform and the percentage of time spent in the target quadrant in the spatial search experiment (##P<0.01 compared with the blank group; **P<0.01 compared with the model group);
[0051] Figure 5 The present application is a block diagram of the computer deep learning-based efficacy prediction system for treating neurodegenerative diseases with Huangjing extract. DETAILED DESCRIPTION
[0052] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.
[0053] Example 1:
[0054] Please refer to Figure 1 The computer deep learning-based efficacy prediction method for treating neurodegenerative diseases with Huangjing extract comprises:
[0055] S1, collect time series and end point observation data related to Huangjing extract administration and test subjects to obtain a unified time-labeled multi-modal data set;
[0056] In step S1, the multi-modal dataset comprises: pre- and post- administration of Huangjing behavioral time series data, Huangjing administration information, blood concentration or pharmacokinetic curve, end-point pathology quantification, and individual metadata;
[0057] The pre- and post- administration of Huangjing behavioral time series data comprises escape latency, crossing platform times, target quadrant residence time, swimming trajectory and speed;
[0058] The Huangjing administration information comprises dose, administration frequency, administration duration, and preparation form;
[0059] The end-point pathology quantification comprises Aβ, tau, and inflammatory markers;
[0060] The individual metadata comprises gender, age, and batch;
[0061] Further, all available information related to the test subjects is collected and formatted into a sample set that can be used for subsequent analysis according to a unified timeline and identification. This includes but is not limited to behavioral time series observations (e.g. multi-day learning / memory test results, activity trajectory summaries), administration information (dose, administration frequency, administration start and end time, administration method), available pharmacokinetic assays (blood concentration sampling points and times), end-point biological markers (pathology quantification, immune indicators, omics assays), and necessary individual metadata (gender, age, batch, etc.). A multi-modal record is obtained for each test individual.
[0062] In step S2, the multi-modal dataset is filtered, missing values are interpolated, time series are aligned, and synthetic samples are generated based on Huangjing pharmacokinetic and pathology models under biological constraints, obtaining a trainable time series multi-modal feature sequence;
[0063] In step S2, the behavioral time series is filtered and scaled, and missing time points are interpolated or statistically completed;
[0064] Image-based pathology data is color-normalized and patch-divided;
[0065] Omic data is batch effect corrected;
[0066] And based on the one-compartment, two-compartment pharmacokinetic model or pathology progression model of Huangjing, limited simulation samples are generated to expand the training set;
[0067] Further, the original multi-modal data is cleaned and converted into a feature sequence suitable for modeling. The behavioral time series is denoised and standardized, and measurement errors and outliers are processed; reasonable interpolation or statistical completion methods are used for missing time points to maintain time continuity; necessary scale transformation and batch correction are performed on different modalities (time series, images, omics, etc.) to eliminate measurement batch effects.
[0068] To alleviate the lack of sample size or improve the robustness of the model, a simulation generation strategy based on pharmacokinetic model or pathological progression model can be introduced under the constraint of biological rationality to generate synthetic samples consistent with the statistical characteristics of real data. Finally, a standardized, aligned and directly input feature sequence set is obtained for representation learning or supervised training.
[0069] S3, performing intra-modal representation extraction on the time-series multi-modal feature sequence and aligning through cross-modal comparison to obtain a sample-level cross-modal semantic vector;
[0070] In step S3, a time convolution network or Transformer is used for future window prediction task to extract time-series representation for behavior time series;
[0071] A contrastive learning or autoencoder is used to extract visual representation for images;
[0072] A variational autoencoder is used to learn latent representation for omics or biochemical scalar;
[0073] An InfoNCE or contrastive loss is used for cross-modal vector alignment to obtain the sample-level cross-modal semantic vector;
[0074] Further, self-supervised learning is used to extract representation from each modality to reduce dependence on labeled data and improve generalization ability of downstream prediction tasks. For time-series data, future window prediction, time-series reconstruction or time-series contrastive task can be used to learn time correlation; for image data, contrastive learning or autoencoder reconstruction can be used to extract local-global visual features;
[0075] For omics or scalar biological indicators, variational autoencoder or self-supervised regression / contrastive task is used to capture latent structure. Then, through cross-modal contrastive loss or projection alignment strategy, the representations of different modalities can be mapped and fused in the semantic space to generate a joint semantic vector for each sample.
[0076] S4, coupling the administration information and pharmacokinetic parameters with the sample-level cross-modal semantic vector, and introducing dose consistency constraint in training to obtain model learning dose response explainability;
[0077] In step S4, the plasma concentration-time curve of Huangjing or the pharmacokinetic parameters obtained by fitting are spliced or attention fused with the sample-level cross-modal semantic vector; and a dose consistency regularization term is added to the training objective, so that samples with similar Huangjing doses or similar PK characteristics have smaller distance in the representation space, thereby obtaining the model learning dose response explainability;
[0078] Further, the dosing related dynamic information (e.g. dose, duration of dosing or pharmacokinetic parameters obtained by fitting) is coupled with the joint semantic vector, thus explicitly preserving the dose-response relationship at the representation level. At implementation level, the pharmacokinetic curve or its parameters can be encoded as a vector and concatenated with the joint semantic vector or fused through attention mechanism, thus forming a dose-sensitive representation in the representation space.
[0079] To ensure the biological consistency of the representation, a "dose consistency constraint" is adopted, i.e. in the training objective or regularization term, samples with similar doses or similar PK characteristics are encouraged to be close to each other or have consistent relative response trends in the representation space. The output of this step is a representation that contains dose-response information and is easy to interpret.
[0080] S5, regress or classify the dose-response interpretable representation learned by the model using the comprehensive efficacy endpoint and output the efficacy prediction distribution to obtain the efficacy prediction result with confidence interval;
[0081] In step S5, the model is trained using a joint loss function (including regression or classification loss and uncertainty-related terms) with the comprehensive efficacy score or single endpoint as the supervised target; and the output is estimated for uncertainty through Bayesian regression head, MC-Dropout or deep Gaussian process, finally outputting the efficacy prediction distribution and corresponding confidence interval;
[0082] Further, the prediction model is trained under the guidance of the supervised signal, and the model learns to map the dose-response interpretable representation to the efficacy output. The supervised target can be a single endpoint (such as a certain index of behavior) or a synthetic comprehensive efficacy score (which is a weighted summary of multiple endpoints). To improve the credibility and decision support ability of the model, in addition to minimizing the point estimate loss, the prediction uncertainty also needs to be modeled (such as using Bayesian neural network, MC-Dropout, deep Gaussian process or prediction distribution learning), so as to output the prediction distribution and confidence interval. Appropriate cross-validation and time segmentation validation strategies should be used for training evaluation to ensure the stability and generalization ability of the model on time series and different batches of data. The efficacy prediction result with confidence interval is: the efficacy prediction distribution (including point estimate and uncertainty measure) for each input sample.
[0083] S6, explain the efficacy prediction result with confidence interval for interpretability and simulate the dosing scenario based on counterfactuals to obtain the corresponding individualized dose duration recommendation;
[0084] In step S6, attention-based or SHAP feature importance analysis is applied to a single sample to identify key driving factors related to the efficacy of Huangjing;
[0085] And by changing the dose of Huangqi in the input or the length of administration (generate counterfactual PK curves) to input the training model to estimate the efficacy increment and uncertainty under different administration scenarios, and output individualized dose length recommendations with confidence intervals;
[0086] Further, for result interpretation and decision support, first, based on the internal weights of the model, attention distribution, or based on post-hoc interpretation methods (such as SHAP, LIME or gradient-based sensitivity analysis), identify the key features driving the prediction (e.g. behavior indicators at a certain time point, certain biomarkers or PK characteristics). Second, through counterfactual simulation - artificially changing the dose of administration / length of administration or pharmacokinetic parameters (within a biologically reasonable range) at the input layer of the model and re-inferencing the efficacy prediction results with confidence intervals - evaluate the efficacy changes and uncertainties under different administration scenarios, thereby generating several comparable individualized dose-length recommendations, and reporting the predicted benefits and confidence levels for each recommendation.
[0087] The method further comprises performance evaluation of the efficacy prediction model, and the evaluation indicators include: RMSE, MAE, R 2 in the regression case, AUROC and accuracy in the classification case, and prediction calibration indicators (such as prediction interval coverage or CRPS), and use multiple batches or external validation to prove the robustness and transferability of the Huangqi-based efficacy prediction model.
[0088] As can be seen from the above, by integrating the behavior time series data before and after the administration of Huangqi (such as escape latency, swimming trajectory), pharmacokinetic parameters (blood concentration curve), pathological quantitative indicators (Aβ, tau protein) and individual metadata (age, gender), a unified time-labeled multi-modal data set is constructed. Through filtering, time series alignment and synthetic sample generation under biological constraints (such as simulation data augmentation based on pharmacokinetic models), the problems of noise interference, missing values and sample scarcity in real data are effectively solved. The collaborative modeling of multi-modal features (such as time convolution network to extract behavior time series features, contrastive learning to process pathological images) significantly improves the model's ability to capture the complex mechanisms of the disease, and its prediction effect is better than that of single modal model.
[0089] Embodiment Two:
[0090] End-to-end test based on behavior data for efficacy prediction:
[0091] Objective: To use existing behavior data and administration labels to construct and verify a deep learning-based efficacy prediction model as a feasibility verification of the model.
[0092] Grouping and dosage: blank group (CON), model group (MOD), positive drug group (DPZ, 1 mg / kg / d), high-dose group of Rhizoma Polygonati (HJ-H, 300 mg / kg / d), low-dose group of Rhizoma Polygonati (PSPL, 150 mg / kg / d).
[0093] Main available endpoints: escape latency of Morris water maze (5-day time series), number of crossings, residence time in target quadrant (percentage), swimming trajectory / movement distance.
[0094] Data preparation: multi-source data collection;
[0095] Raw samples: behavioral daily sequence data (5-day escape latency / daily trajectory summary) of 10 mice per group and grouping / dose labels, a total of 5 groups x 10 = 50 samples (each sample contains 5-day time series and outcome indicators).
[0096] Data format:
[0097] Each mouse forms a record (ID, group, dose, day1_latency, …, day5_latency, total_crossings, target_quadrant_time_pct, swim_distance_mean).
[0098] train / val / test split: in mouse units, 70% training (35), 15% validation (7), 15% testing (8). If the sample size is insufficient, stratified resampling is used to ensure the representativeness of each group in each set.
[0099] Preprocessing and augmentation: time series preprocessing and data augmentation;
[0100] Time series processing: Z-Score normalization is performed on the escape latency sequence; if individual time points are missing, linear interpolation or spline interpolation is used for completion.
[0101] Trajectory / speed: convert trajectory summary into average speed and turning entropy features for dimensionality reduction.
[0102] Data augmentation (small sample relief): use time series segmentation and reorganization (sliding windows) and a small amount of Gaussian noise injection; at the same time, use intra-dose group comparison learning to generate slightly perturbed samples (simulation) while preserving biological consistency.
[0103] Character learning and model: self-supervised representation learning & pharmacodynamic-pharmacokinetic coupling;
[0104] Self-supervised pre-training: do future window prediction task (Transformer encoder-decoder) with all time series in training set, learn time series representation (embedding dim = 128).
[0105] Fusion input: final model input is time series representation vector + dose scalar (one-hot + scalar value).
[0106] Pharmacokinetic simplification coupling: plasma concentration, by "dose value" and "dosing day (30 d)" as pharmacokinetic proxy variables, add "dose consistency" regularization (make the same dose samples closer in representation space) in model training.
[0107] Model structure (end-to-end): time series Encoder (Transformer: 4 layers, head=4, hidden=256) → representation concatenation (plus dose) → MLP regression head (3 layers, hidden units 256 / 128 / 64) → output as continuous efficacy score (synthetic efficacy score = weighted sum of normalized escape latency reduction and crossing times).
[0108] Training details: end-to-end supervised training and uncertainty quantification;
[0109] Loss function: mainly mean square error (MSE) for regression; additional dose consistency term (lambda = 0.1).
[0110] Uncertainty estimation: MC-Dropout (dropout = 0.2) in MLP regression head and estimate prediction mean and confidence interval (95% CI) by multiple forward passes (T = 50).
[0111] Hyperparameters: learning rate 1e-4 (Adam), batch = 8, training epochs = 200, early stopping (stop if validation loss does not decrease for 20 rounds).
[0112] Validation strategy: 5-fold cross-validation (stratify by group) and finally evaluate RMSE, MAE, R² and prediction interval coverage on test set.
[0113] Evaluation and output: interpretable posterior analysis and counterfactual dose optimization;
[0114] Output: predicted efficacy score and 95% confidence interval for each mouse; and estimate efficacy change by simple counterfactual (dose from 150 -> 300 or -> 0) input model (used to guide whether dose increase can bring significant efficacy benefit).
[0115] Interpretability: SHAP values are used to calculate the contribution of each input time point and dose to the prediction, and the key time points are output (e.g., day3 latency contributes the most to the final efficacy).
[0116] For HJ-H and PSPL groups in the test set, compare the model's predicted efficacy with the actual average efficacy (escape latency reduction, crossing frequency increase) to assess model consistency and error range.
[0117] Expected results:
[0118] The model is well trained and is expected to achieve an acceptable RMSE range (depending on data variability) in the test set; and can identify that the HJ-H (300 mg / kg / d) group has a higher predicted efficacy score than the PSPL (150 mg / kg / d) group.
[0119] As can be seen from the above, by coupling pharmacokinetic parameters with cross-modal semantic vectors and introducing a dose-consistent regularization term, the model is forced to learn an interpretable explanation of the dose-response relationship. Specifically, samples with similar doses or similar pharmacokinetic characteristics are constrained to be adjacent in the representation space, so that the model's prediction not only focuses on the efficacy result, but also explicitly reflects the influence mechanism of drug dosage. At the same time, by analyzing the feature importance of individual samples through SHAP values or attention mechanisms, key driving factors related to efficacy (such as Aβ deposition in specific brain regions or peak blood drug concentration) can be identified. This design not only meets the strict requirements of the medical field for model transparency (such as FDA's regulations on the interpretability of AI / ML medical devices), but also provides clinicians with verifiable decision-making basis.
[0120] Example Three:
[0121] Introduce pathological quantification and PK simulation to achieve higher interpretability of dose-response prediction:
[0122] Objective: Based on Example Two, expand the data modalities to improve the model's biological relevance and preclinical translational value. Suitable for further patent protection and subsequent scale-up experiments.
[0123] Based on the dosing regimen (30-day gavage, dose grouping) as the design basis. To further modeling, this example adds in similar experiments: blood samples for simple pharmacokinetic (blood drug concentration Tmax / Cmax) determination; endpoint Aβ / tau quantification or immunohistochemical scoring; optional inflammatory factor (such as IL-1β, TNF-α) ELISA values. Existing behavioral studies have shown that Huangjing can improve learning and memory, and the expansion of biomarkers helps to establish a biological logic chain for dose-response.
[0124] Data Preparation:
[0125] Each mouse record contains: behavioral time series (5 days), dosing schedule, blood concentration series (sampling points like 0.5h, 1h, 4h, 24h post-dose at week 1 and week 4), endpoint Aβ / tau quantification, inflammatory factors and possibly pathology images (light microscope photos or scans).
[0126] Sample size: at least 12-15 per group, to ensure adequate statistical power; if budget allows, increase the number of replicates (n=2 batches) to enhance extrapolation.
[0127] Time series and PK pre-processing:
[0128] PK modeling: fit a simplified one- or two-compartment pharmacokinetic model (parameter estimation: clearance Cl, volume of distribution Vd, half-life t1 / 2) from blood concentration points. Use the fitted parameters to generate individual blood concentration-time curves (continuously estimated) as input for the pharmacodynamic model.
[0129] Image processing: digitize the pathology slides, perform color normalization, extract patches (224x224), and annotate Aβ deposition areas (if manually annotated).
[0130] Representation learning:
[0131] Behavioral time series encoding: Transformer (or TCN) to extract sequence representation (dim=128).
[0132] PK representation: sample the fitted PK curves into equidistant vectors and encode using a small CNN or 1D-Conv (dim=64).
[0133] Image representation: ResNet50 encoder (pretrained optional) outputs dim=128.
[0134] Cross-modality alignment: align the three types of representations using contrastive learning (InfoNCE) with projection heads, resulting in joint semantic vectors.
[0135] Pharmacodynamic-coupled learning:
[0136] Concatenate the PK encoding, dose values, and joint semantic vectors; use multi-head attention at the fusion layer, where the attention heads can be interpreted as “which modality / time point is most affected by the dose”.
[0137] Introduce physical constraints: add “dose-PK consistency loss” (e.g., a regularization term that penalizes samples with similar PK parameters being close in representation space) and “biomarker consistency term” (samples with high predicted efficacy should show lower Aβ or lower inflammatory markers, if consistent with the data) to the model loss.
[0138] End-to-end training and uncertainty:
[0139] Supervised objective: can be defined as a multi-endpoint composite metric (behavioral score standardization + inverse weighting of Aβ / tau + magnitude of inflammation reduction) to capture the overall efficacy.
[0140] Uncertainty: use deep Gaussian processes or Bayesian neural networks (if computational resources are limited, use MC-Dropout) to obtain predictive distributions.
[0141] Training strategy: stratified cross-validation + batch effect correction; also perform time-series leave-one-out validation to test long-term generalization (e.g., use the first batch for training, the second batch for external validation).
[0142] Posterior explanation and counterfactual optimization:
[0143] Interpretability: attention weights combined with SHAP to identify the contribution of "PK peak / curve features" or "specific pathological metrics" to the efficacy.
[0144] Counterfactual experiments: systemically change dose / pharmacokinetic parameters at the input layer of the model (e.g., halve the clearance rate to simulate a slow-clearance individual) to observe changes in efficacy and provide individualized dose recommendations (while reporting uncertainty). This provides direct suggestions for subsequent in vivo validation and preclinical dose design.
[0145] Output format: for each experimental animal, generate (a) predicted efficacy score with 95% CI, (b) list of key driving factors, (c) several feasible dose-duration combinations and their predicted gains and risks (uncertainty).
[0146] Experimental validation and evaluation:
[0147] In extended embodiments, two validations are performed: one is internal consistency validation (cross-validation on existing batches); the other is external repetition validation (on independent batches or mice of different ages). In addition to RMSE / MAE / R², biological consistency tests are added (e.g., samples with high predicted efficacy have significantly lower average Aβ levels than those with low predicted efficacy, statistical test p<0.05).
[0148] Additional notes:
[0149] Implemented using PyTorch or TensorFlow; training uses single or multi-GPU (e.g., NVIDIA RTX 3090 or A100); save random seeds to ensure reproducibility.
[0150] Hyperparameters (can be used as a starting point and need to be tuned): Transformer layers = 4, embed = 128, batch = 8-16, lr = 1e-4, epochs = 100-300 (early stopping), dropout = 0.2.
[0151] 30 d continuous gavage with dose setting directly used for model of drug administration label and pharmacokinetic simulation baseline.
[0152] Behavioral endpoints (escape latency, crossing times, target quadrant residence time) as the core supervision target and evaluation index of the model.
[0153] As can be seen from the above, the method changes the dose or administration time in the input, uses the trained model to simulate the efficacy increment and uncertainty under different administration conditions, and outputs individualized dose recommendations with confidence intervals. This counterfactual reasoning capability breaks through the limitations of traditional group average effect analysis and can dynamically adjust treatment strategies according to patient individual characteristics (such as metabolic rate, pathological progression stage). For example, in the treatment of dry eye, counterfactual analysis has successfully helped patients evaluate the expected effect of different eye drops and avoid potential allergic risks. In addition, the program estimates the uncertainty of the prediction results through Bayesian regression or deep Gaussian process, quantifies the reliability of the model in different scenarios, and further reduces the decision-making risk in clinical application.
[0154] Example Four:
[0155] Pharmacological effects of Huangjing on 3xTg-AD mice:
[0156] ①Grouping and administration:
[0157] 6-month-old 3xTg-AD mice were randomly divided into 10 groups, 10 in each group, half male and half female, and 10 C57BL / 6J mice of the same age were used as a blank group, as shown in Table 1. Except for the Huangjing administration group, the blank group and the model group were given the same volume of normal saline, and the positive drug donepezil hydrochloride was calculated according to the instructions, once a day, for 30 days of continuous gavage administration (all gavage drugs were dissolved in 5‰ CMC).
[0158] Table 1 Grouping and administration dose:
[0159] Full name of group Abbreviation of group Dose of administration Blank group (C57BL / 6J mice) CON - Model group (3xTg-AD mice) MOD - Positive drug group (donepezil hydrochloride) DPZ 1 mg / kg / d High-dose Huangjing group HJ-H 300 mg / kg / d Low-dose Huangjing group PSPL 150 mg / kg / d
[0160] ②Morris water maze test behavior experiment:
[0161] The results of the place navigation experiment showed that Figure 2 , Figure 3): The model group mice walked around the periphery, while the trajectories of the other groups were mostly directional, and the movement distance and escape latency were significantly shorter. Compared with the blank group, the escape latency of the model group mice was longer (P<0.01), indicating that the learning and memory ability of the model group mice was significantly impaired; compared with the model group, the movement distance and escape latency of the positive drug group and the Huangjing administration group were shortened (P<0.05), indicating that the Huangjing extract can improve the learning and memory ability of AD mice.
[0162] The results of the spatial exploration experiment show that Figure 4 ) Compared with the blank group, the number of times of crossing the platform and the time of staying in the target quadrant of the model group mice were significantly reduced (P<0.01); compared with the model group, the number of times of crossing the platform and the time of staying in the target quadrant of the AD mice in the positive drug group and the Huangjing administration group were significantly increased (P<0.01), indicating that Huangjing has the ability to improve memory.
[0163] Example Five:
[0164] As Figure 5 shown, in one embodiment, the present application also provides a computer deep learning-based efficacy prediction system for treating neurodegenerative diseases by using Huangjing extract, comprising:
[0165] A multi-source data acquisition module is used to collect time series and endpoint observation data related to Huangjing extract administration and test subjects, and obtain a unified time-labeled multi-modal data set;
[0166] A time series preprocessing and data enhancement module is used to filter, impute missing values, align time series, and generate synthetic samples based on Huangjing pharmacokinetics and pathological models under biological constraints, and obtain trainable time series multi-modal feature sequences;
[0167] A self-supervised representation learning module is used to extract intra-modal representations from the time series multi-modal feature sequences and align them through cross-modal comparison, and obtain sample-level cross-modal semantic vectors;
[0168] A pharmacodynamic and pharmacokinetic coupling representation learning module is used to couple the administration information and pharmacokinetic parameters with the sample-level cross-modal semantic vectors, and introduce a dose consistency constraint in the training, and obtain a model learning dose response interpretable explanation;
[0169] An uncertainty quantification module is used to regress or classify the model learning dose response interpretable explanation with respect to the comprehensive efficacy endpoint and output an efficacy prediction distribution, and obtain an efficacy prediction result with a confidence interval;
[0170] A counterfactual dose optimization module is used to explain the interpretability of the efficacy prediction result with a confidence interval and simulate the administration scenario based on counterfactuals, and obtain the corresponding individualized dose length recommendation.
[0171] The beneficial effects are the same as those of the technical effects of the computer deep learning-based prediction method for the curative effect of the Rhizoma Polygonati extract on neurodegenerative diseases, which will not be repeated here.
[0172] Although embodiments of the present application have been shown and described, it is to be understood that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A computer deep learning-based method for predicting the efficacy of a polygonatum extract in treating a neurodegenerative disease, characterized by, The method comprises the following steps: Collecting time series and endpoint observation data related to the administration of the polygonatum extract and the subject, to obtain a unified time-labeled multi-modal data set; Filtering, missing value imputation, time series alignment, and generating synthetic samples based on the polygonatum pharmacokinetic and pathological models under biological constraints on the multi-modal data set, to obtain time series multi-modal feature sequences; Extracting intra-modal representations from the time series multi-modal feature sequences and aligning them through cross-modal comparison, to obtain sample-level cross-modal semantic vectors; Coupling the administration information and pharmacokinetic parameters with the sample-level cross-modal semantic vectors, and introducing a dose consistency constraint in the training, to obtain a model learning dose response interpretable explanation; Performing regression or classification training on the model learning dose response interpretable explanation based on the comprehensive efficacy endpoint, and outputting an efficacy prediction distribution, to obtain an efficacy prediction result with a confidence interval; Performing interpretable explanation on the efficacy prediction result with a confidence interval and simulating the administration scenario based on counterfactuals, to obtain a corresponding individualized dose duration recommendation.
2. The computer deep learning-based method for predicting the therapeutic effect of the Rhizoma Polygonati extract on neurodegenerative diseases according to claim 1, characterized in that, The multi-modal data set comprises: pre- and post-administration of polygonatum behavioral time series data, polygonatum administration information, blood concentration or pharmacokinetic curves, endpoint pathological quantification, and individual metadata; The pre- and post-administration of polygonatum behavioral time series data comprises escape latency, platform crossing times, target quadrant residence time, swimming trajectory and speed; The polygonatum administration information comprises dose, administration frequency, administration duration, and preparation form; The endpoint pathological quantification comprises Aβ, tau, and inflammatory markers; The individual metadata comprises gender, age, and batch. 3.The computer deep learning-based method for predicting the therapeutic effect of the Rhizoma Polygonati extract on neurodegenerative diseases according to claim 1, characterized in that, The filtering, missing value imputation, time series alignment, and generating synthetic samples based on the polygonatum pharmacokinetic and pathological models under biological constraints on the multi-modal data set, to obtain time series multi-modal feature sequences, comprises: Filtering and scale normalization for behavioral time series, and interpolation or statistical completion for missing time points; Color normalization and patch segmentation for image-based pathological data; Batch effect correction for omics data; And generating restricted simulation samples based on the one-compartment, two-compartment pharmacokinetic model or the pathological progression model of polygonatum to expand the training set. 4.The computer deep learning-based method for predicting the therapeutic effect of the Rhizoma Polygonati extract on neurodegenerative diseases according to claim 1, characterized in that, The intra-modal representation extraction from the time series multi-modal feature sequences and the cross-modal comparison alignment, to obtain sample-level cross-modal semantic vectors, comprises: Using time convolution network or Transformer for future window prediction task to extract time series representation for behavioral time series; Using contrastive learning or autoencoder to extract visual representation for images; Using variational autoencoder to learn latent representation for omics or biochemical scalar; And aligning the cross-modal vectors through InfoNCE or contrastive loss, to obtain the sample-level cross-modal semantic vectors. 5.The computer deep learning-based method for predicting the therapeutic effect of the Rhizoma Polygonati extract on neurodegenerative diseases according to claim 1, characterized in that, The coupling of the administration information and pharmacokinetic parameters with the sample-level cross-modal semantic vectors, and the introduction of a dose consistency constraint in the training, to obtain a model learning dose response interpretable explanation, comprises: The blood concentration-time curve of Huangjing or the fitted pharmacokinetic parameters are spliced or attention-fused with the sample-level cross-modal semantic vector, and a dose consistency regularization term is added in the training target, so that samples with similar Huangjing doses or similar PK characteristics have smaller distances in the representation space, thereby enabling the model to learn dose response interpretable explanation. 6.The computer deep learning-based method for predicting the therapeutic effect of the Rhizoma Polygonati extract on neurodegenerative diseases according to claim 1, characterized in that, The model learns dose response interpretable explanation by regression or classification training based on the comprehensive efficacy endpoint and outputs efficacy prediction distribution, to obtain efficacy prediction results with confidence intervals, including: The model is trained using a joint loss function with comprehensive efficacy scores or single endpoints as supervision targets, and the output is uncertainty estimated through Bayesian regression head, MC-Dropout or deep Gaussian process, finally outputting efficacy prediction distribution and corresponding confidence intervals. 7.The computer deep learning-based method for predicting the therapeutic effect of the Rhizoma Polygonati extract on neurodegenerative diseases according to claim 1, characterized in that, The model learns dose response interpretable explanation by regression or classification training based on the comprehensive efficacy endpoint and outputs efficacy prediction distribution, to obtain efficacy prediction results with confidence intervals, including: Attention-based or SHAP feature importance analysis is applied to single samples to identify key driving factors related to Huangjing efficacy; The training model is changed to estimate the efficacy increment and uncertainty under different dosing scenarios by changing the Huangjing dose or dosing duration input, thereby outputting individualized dose duration recommendations with confidence intervals. 8.The computer deep learning-based method for predicting the therapeutic effect of the Rhizoma Polygonati extract on neurodegenerative diseases according to claim 1, characterized in that, The method further comprises performance evaluation of the therapeutic effect prediction model, the evaluation indexes comprising: RMSE, MAE, R 2 in the regression case, AUROC and accuracy in the classification case, and prediction calibration indexes, and multiple batches or external verification is used to prove the robustness and transferability of the therapeutic effect prediction model based on Rhizoma Polygonati.
9. A computer deep learning-based system for predicting the efficacy of a polygonatum extract in treating a neurodegenerative disease, characterized by, Including: A multi-source data collection module is used to collect time-series and endpoint observation data related to Huangjing extract administration and subjects, and to obtain a unified time-labeled multi-modal dataset; A time series preprocessing and data enhancement module is used to filter, impute missing values, align time series, and generate synthetic samples based on Huangjing pharmacokinetics and pathology models under biological constraints, to obtain time-series multi-modal feature sequences; A self-supervised representation learning module is used to extract intra-modal representations from time-series multi-modal feature sequences and align them through cross-modal comparison, to obtain sample-level cross-modal semantic vectors; A pharmacodynamic and pharmacokinetic coupling representation learning module is used to couple dosing information and pharmacokinetic parameters with the sample-level cross-modal semantic vector, and introduce dose consistency constraints in training, to obtain model learning dose response interpretable explanation; An uncertainty quantification module is used to regress or classify the model learning dose response interpretable explanation based on the comprehensive efficacy endpoint and output efficacy prediction distribution, to obtain efficacy prediction results with confidence intervals; An counterfactual dose optimization module is used to explain the interpretability of the efficacy prediction results with confidence intervals and simulate the dosing scenarios based on counterfactuals, to obtain corresponding individualized dose duration recommendations.