A method of predicting chemical reactions

Abstract

The application discloses a kind of prediction methods of chemical reaction product, including based on using original data set D0 training original training model, the reaction product of different reactions is predicted and the reaction with the confidence lower than threshold value is calculated, these data are screened and constitute first data set D1.Second data set D2 is provided, and similar reaction with the chemical reaction in first data set D1 is screened as third data set D3.D3 data is merged in original data set or D3 is used separately, and model training is carried out again.The method of the application can improve the relationship between the confidence and the real accuracy of prediction, so that high confidence prediction has high accuracy, and finally improves the accuracy of reaction prediction, and the method also has the advantages of small data volume and short time.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a chemical reaction product prediction method and device. BACKGROUND

[0002] In the field of pharmaceutical chemistry applications, the organic synthesis of new chemical molecules requires relevant prediction and judgment of chemical reactions conceived by organic chemists or virtually generated by computer algorithms to avoid losses and waste caused by experimental failures.

[0003] The prediction accuracy of existing reaction prediction models is highly dependent on training data, and the model performance may be limited due to incomplete reaction data. Simply retraining the model with additional reaction data cannot effectively solve the key reactions focused on in specific application fields. For example, in the design of organic synthesis, increasing reaction data is not always better, and indiscriminate supplementation cannot effectively improve this category, and even may decrease the performance of such reactions. SUMMARY

[0004] Therefore, it is necessary to provide a chemical reaction product prediction method and device to solve the problem of low prediction accuracy of the existing reaction prediction model.

[0005] In one aspect, the present application discloses a chemical reaction product prediction method, which comprises:

[0006] Step 1: obtaining one or more machine models capable of generating reaction prediction and outputting the prediction reliability thereof, calculating the reliability of the predicted product corresponding to each model in the original data set through a given chemical reaction, statistically analyzing the overall reliability of all models, screening reaction data with a reliability less than a threshold value, and obtaining a first data set D1; wherein the threshold value is any number in the range of 0.3 to 0.9, preferably 0.4 to 0.8, and more preferably 0.5 to 0.7, such as about 0.5, 0.6 or 0.7.

[0007] Step 2: providing a second data set D2, calculating the similarity sim(w, v) between the chemical reaction W in D2 and the chemical reaction V in D1, screening similar reaction data in D2 with sim(w, v) greater than or equal to a threshold value, and obtaining a third data set D3 in a set manner, wherein the threshold value is any value in the range of 0.1 to 1, preferably 0.3 to 0.8, and more preferably 0.5 to 0.8, such as about 0.6, 0.7 or 0.8;

[0008] Step 3: merging D3 data in the original data set or using D3 data to retrain the model.

[0009] In one embodiment, in step 1, K≥1 model features of machine models that can generate reaction prediction and output the credibility of their prediction are characterized by model parameters θt, where t = 1, 2, …, K represents. Where t represents the tth model snapshot, and K is the number of collected model snapshots.

[0010] In the case where the original training data D0 is available, the machine translation transformer (Transformer) is selected as the original training model, and in other embodiments, the machine model can be replaced by other deep neural network-based models.

[0011] In one embodiment, in step 1, when the product information is known, confidence = p(Y|X, θt); when the reaction product information is unknown, Y max = arg max i (confidence = p(Yi|X, θt)) gets (X, Y max ), i is the ith output prediction that the model can provide, preferably i≤10. t represents the tth model snapshot, K is the number of collected model snapshots, X represents the reactants of the chemical reaction, Y represents the products of the reaction, p represents the probability of the model output Y when X and θt are known, Y max is the predicted product of the model, and arg max represents the maximum value of all Y i probabilities.

[0012] In one embodiment, in step 1, the overall credibility can be characterized as the average mean(confidence(X, θt)), or the maximum max(confidence(X, θt)), or other statistical operations that can be easily mastered by skilled persons in the field.

[0013] In one embodiment, in step 2, for any reaction W∈D1, the similarity sim(w, v) between it and any reaction V∈D2 under the model parameters θt is calculated, where sim(w, v) = sim(w = encoding(W), v = encoding(V)); where w, v are the encodings of input reactions V, W by model θt, respectively.

[0014] In one embodiment, in step 2, the amount of data D3 is less than the original training data D0, and for each reaction in D1, the number of reactions supplemented by D3 can be controlled within one hundred, preferably |D3|≤|D0|, |D3|≤50×|D1|.

[0015] In one embodiment, in step 3, R times the data amount of D3 is randomly sampled from the original training data D0, and D3 is merged to generate a new data set, and the reinitialized machine model parameters are retrained; R can be selected from the range of [0.5, max(1, |D0| / |D3|)].

[0016] In one embodiment, in step 3, a new data set is generated using D3, and the reinitialized machine model parameters are retrained. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 A flowchart illustrating a method of predicting chemical reaction products is shown.

[0018] Figure 2 Seven incorrect reaction similar neighbor reaction numbers are shown.

[0019] Figure 3 Twelve correct reaction similar neighbor reaction numbers are shown. DETAILED DESCRIPTION

[0020] The present application will be described in detail below with reference to embodiments and in conjunction with the accompanying drawings. The above aspects of the present application and other aspects of the present application will be apparent from the following detailed description. The scope of the present application is not limited to the following embodiments.

[0021] As shown in Figure 1 , the present application discloses a method for predicting chemical reaction products, the method comprising:

[0022] Step 1: Based on the original training model, predict the reaction products of different reactions and calculate the "under-learning" reactions with a confidence lower than a threshold value, screen these data and form a first data set D1.

[0023] Step 2: Screen similar reactions to the "under-learning" chemical reactions as the third data set D3.

[0024] Step 3: Merge D3 data into the original data set and retrain the model.

[0025] In step 1, one or more machine models capable of generating reaction predictions and outputting their prediction confidence are first obtained, then the prediction confidence corresponding to the predicted products in each model is calculated through a given chemical reaction in the original data set, the overall confidence of all models is counted, and finally the reaction data with a confidence less than a threshold value, for example, 0.5, is screened to obtain the first data set D1.

[0026] In step 1, K≥1 model features of machine models that can generate reaction prediction and output the credibility of their prediction are represented by model parameters θt, where t = 1, 2, …, K. When the product information is known, confidence = p(Y|X, θt); when the reaction product information is unknown, Y max = arg max i (confidence = p(Yi|X, θt)) gets (X, Y max ), i is the ith output prediction that the model can provide, i≤10. The overall credibility can be represented as the average mean(confidence(X, θt)), or the maximum max(confidence(X, θt)), or other statistical operations that can be easily mastered by those skilled in the art.

[0027] The specific calculation method of the credibility Confidence can be that the X part of the reaction data is calculated by the weights of each layer of the multi-layer neural network of the trained machine model Transformer, and the original weights zi (>0) of all possible M element symbols in the output product are obtained at the output layer of the model, i = 1, 2, …, M, and the normalized probability is calculated by Softmax as the confidence of each character i, and the element symbol sequence with the maximum probability is output as the prediction Y.

[0028]

[0029] In other embodiments, the machine model can be replaced by other deep neural network-based machine models, which use the same Softmax to calculate the output layer, except that the form of the output element symbol is changed.

[0030] In step 2, first provide the second data set D2, then for each chemical reaction W in D2, calculate its similarity sim(w,v) with each chemical reaction V in D1, and finally filter the similar reactions in D2 with sim(w,v) greater than or equal to a threshold value, and obtain the third data set D3 by taking the union of the filtered data. The threshold value can be any value in the range of 0.1 to 1. For any reaction W in D1, calculate its similarity sim(w,v) with any reaction V in D2 under the model parameter θt, where sim(w,v) = sim(w = encoding(W), v = encoding(V)); where w and v are the encodings of input reactions V and W, respectively, under the model θt. The amount of data in D3 is less than the original training data D0, and for each reaction in D1, the number of reactions supplemented by D3 can be controlled within one hundred, preferably |D3| ≤ |D0| and |D3| ≤ 50 × |D1|. In a specific embodiment, for reaction W, w = f(W, θt) = [w1, w2, … wn], where f(W, θt) is the vector representation of the reaction W input to the model θt through the parameters of each layer before the output prediction element, and n is a parameter representing the length of the vector. Similarly, for each reaction V in D2, v = f(V, θt) = [v1, v2, … vn]; n can be selected in the range of 2 6 = 64 to 2 12 = 4096, preferably n = 256.

[0031] sim(w,v) can be realized as the normalized inverse of its Euclidean distance (+1 to avoid division by zero), or other normalized similarity that can be mastered by those skilled in the art:

[0032]

[0033] In step 3, merge D3 data into the original data set and retrain the model. First, randomly sample R times the amount of data in D3 from the original training data D0, and merge it with D3 to produce a new data set, then retrain the reinitialized machine model parameters; R can be selected from the range [0.5, max(1, |D0| / |D3|)].

[0034] Or use D3 for fine-tuning learning, i.e. retrain the model θt on D3 data for F ≥ 1 iterations to continuously update the model parameters.

[0035] Example 1: Method for predicting chemical reaction products

[0036] 1. Based on the original training model trained using the original data set D0, predict the reaction products of different reactions and calculate the reactions with a confidence level below a threshold value, filter these data and form the first data set D1

[0037] This example is in the case where the original training data D0 is available, D0 comes from 400,000 training data of the public dataset United States Patent Data Office (USPTO), and the machine translation converter (Transformer) (Philippe Schwaller et al. Molecular transformer: A model for uncertainty-calibrated chemical reaction prediction, 2019 Sep 25; 5(9): 1572-1583; Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pp. 5998-6008, 2017) is selected as the original training model. In other examples, the machine model can be replaced by other deep neural network-based models (Coley, Connor W., et al. A graph-convolutional neural network model for the prediction of chemical reactivity. Chemical science 10.2 (2019): 370-377.; John Bradshaw, Matt J. Kusner, Brooks Paige, Marwin H. S. Segler, José Miguel Hernández-Lobato, A Generative Model For Electron Paths, https: / / arxiv.org / abs / 1805.10970), only the output element symbol form is changed. And record K>=1 model snapshots in the training iteration process using D0; the features of the model snapshots can be characterized by the model parameters θt, t = 1, 2, …, K, where t represents the tth model snapshot, and K is the number of collected model snapshots.

[0038] The selection of θt,t = 1,2,…,K is selected according to different iterations of model training; each time a model is updated according to the parameters θt of each sample of the training data, it is called an epoch iteration; K can be selected as the total number of epochs, and θt corresponds to the model of the tth epoch iteration. When K = 1, the model trained to the last epoch iteration is selected. In other embodiments, an epoch can also be set as a certain number of iteration intervals, such as every 1000 iterations as an epoch.

[0039] Given the chemical reaction data (X, Y) to be analyzed, X represents the reactants of the chemical reaction, and Y represents the products of the reaction. The confidence can be calculated by the parameter θt of the model snapshot, p represents the probability of the model output Y when X and θt are known; such as only given X, the model can output Y max = arg max i (confidence = p(Y i |X, θ)) gets (X, Y max ), Y max is the predicted product of the model, and arg max represents the maximum value of all Y i probability. i is the ith highest output prediction obtained by the model through beam-search, i ≤ 10.

[0040] The specific calculation method of confidence can be that the X part of the reaction data is calculated through the weights of each layer of the multi-layer neural network of the trained machine model Transformer, and the original weights zi (> 0) of all possible M element symbols in the output layer of the model are obtained, i = 1,2,…,M, and the normalized probability is calculated by Softmax as the confidence of each character i, and the element symbol sequence with the maximum probability is output as the prediction Y.

[0041]

[0042] For the above chemical reaction data set to be analyzed, the “under-learning” reaction data set D1 is screened, and th represents the confidence threshold, th is selected from 0.5.

[0043] In other embodiments, the confidence threshold can be in the range of 0.3 to 0.9, preferably 0.4 to 0.8, more preferably 0.5-0.7, for example 0.5, 0.6 or 0.7.

[0044] 2. Provide a second data set D2, and screen similar reactions to the chemical reactions in the first data set D1 as a third data set D3

[0045] A candidate supplementary response dataset D2 = {(X',Y')} is provided for screening responses. D2 is derived from the USPTO supplementary database, which contains approximately 1 million responses on the USPTO Stereo. For any response W ∈ D1, this invention calculates its similarity to any response V ∈ D2 under model parameters θ: sim(W,V) = sim(w = encoding(W), v = encoding(V)); where w and v are the encoding functions of model θ for the input responses V and W, respectively.

[0046] The calculation of similarity sim(w,v) is explained below. Taking reaction W as an example, its encoding vector representation is w=f(W,θ)=[w1,w2,…w n f(W,θ) is specifically the vector representation of the reaction W input to the model θ, calculated using parameters from each layer, and preceding the output prediction element. Here, n is a parameter representing the pre-defined length of the representation vector. Similarly, for each reaction V in D2, the encoding vector v = f(V,θ) = [v1, v2, ... v2] can be obtained. n ]; n can be 2 6 =64 to 2 12 The length is selected from the range of 4096, and in this embodiment, n=256 is selected.

[0047] sim(w,v) can be implemented as the normalized reciprocal of its Euclidean distance (+1 to avoid the divisor being 0), or a normalized similarity that can be grasped by those skilled in other fields.

[0048]

[0049] Set the relevant similarity threshold th2, which is set in the range [0.1,1].

[0050] We filter similar response supplementary datasets that meet the criteria sim(encoding(W), encoding(V))≥th2, where th2∈[0.1,1] is the similarity threshold; we obtain the set D3 of similar response supplementary data by taking a set. In this example, we use th2=0.6 and analyze the relevant supplementary result samples of 0.7 and 0.8.

[0051] 3. Combine the D3 data with the original dataset or use D3 separately to retrain the model.

[0052] In one of the experiments in this embodiment, model θ is fine-tuned using D3, which means that model θ is trained on D3 data for F>=1 iterations to continuously update the model parameters. This scheme is used in embodiment 2.

[0053] In another variant of the embodiment, when both the model θ and its corresponding original training data (denoted as D0) exist, the reaction data of the two sets is selected in the ratio of |D0|: |D3| using D0 and D3, i.e. directly merging D0 and D3 as a new data set as option 1, and retraining the reinitialized machine model.

[0054] Optionally, the variant of the embodiment can take the results of fine-tuning and retraining as N = 2 option models, and select the final model by selecting the option model with the largest accuracy improvement in the test reaction data set D1 (or other additional provided test reaction data set), and the largest number of prediction / verification reactions in the category with a confidence of > 0.9, or more strictly, a confidence of > 0.99.

[0055] Accuracy detection of embodiment 2

[0056] The following shows the experimental results based on the embodiment. According to the steps in embodiment 1, the machine learning model θ uses a Transformer, the encoding vector dimension is selected as n = 256, the number of training iterations is 500,000, and a small batch of 4096 characters (tokens) is processed at each iteration. The trained reaction data is a public data set of 400,000 training data of the USPTO (USPTO), and θ takes the model output at the last iteration. The 400,000 USPTO training data is denoted as D0.

[0057] The reaction data set D1 for the under-learning analysis test is about 1381 reactions extracted by internal chemists from basic organic chemistry books. The candidate reaction data set D2 to be supplemented is from the USPTO data set that does not overlap with D0, and there are a total of 400,000 reactions; it is worth noting that the backend of the reaction data commercial service contains a large number of reactions, but such services only provide a small amount of data query and cannot obtain a large amount of data. For example, the Reaxys www.reaxys.com page introduces that it contains more than 55 million reactions, but only about ten reactions can be seen on a page of query results.

[0058] In the verification experiment, the baseline model trained only by D0 and the confidence improvement model obtained by fine-tuning about 14000 reactions with a similarity of ≥0.6 from the D2 data set containing 400,000 candidate supplementary reactions in D1 are compared.

[0059] Top-k accuracy represents the highest confidence of the model's prediction of k different possible products, one of which exactly matches the true product. Top-1 accuracy is the proportion of all reactions in which the model's prediction of the most likely product exactly matches the true product.

[0060] As shown in Table 1, the experimental results prove that using the screening method of the present application, about ten similar reactions are supplemented for each reaction of D1, which can significantly improve the prediction effect on the original model. Whether it is the overall Top-1 accuracy or the coverage of correct prediction with high confidence, there is a significant improvement. The Top-1 accuracy is improved by 22.6%, the coverage of Confidence>0.9 is increased by 20.86%, and the prediction in this confidence interval reaches a Top-1 accuracy of 93.9%.

[0061] Table 1

[0062]

[0063] In further experiments, for the 200 test reactions screened for low confidence Confidence<0.5, the Top-1 accuracy of the reaction prediction was only 8.5% before using the invention for screening and improvement. This set of reactions is verified as "under-learning" reactions. After screening by the invention and fine-tuning the baseline model, the average Confidence of the test is improved from 0.378 to 0.796, and the Top-1 accuracy is improved to 60.5%, verifying the improvement of the invention method on the accuracy and confidence of reaction prediction.

[0064] Table 2

[0065] Confidence <0.5 screen 200 test reactions Baseline Model Confidence Improvement Model Average Confidence 0.378 0.796 Top-1 Accuracy 8.5％ 60.5％

[0066] On the other hand, for another 100 randomly selected error test reactions, the original accuracy is 0, of which 33 have Confidence>0.9 and 16 have Confidence>0.8. The experiment supplements data without distinction, i.e., without similar reaction supplementation according to the reaction prediction Confidence threshold. The Top-1 accuracy of this part of test reactions after supplementation is 14%, i.e., without supplementation according to the Confidence threshold and similarity, the accuracy improvement is limited.

[0067] Table 3

[0068]

[0069] For the reactions with Confidence>0.9 after non-discriminatory supplementation, Figure 2The 7 reactions with high conf) and still predicted wrong were analyzed for the reason of lacking of similar training or supplementary reaction data, i.e. the number of similar neighbor reactions with sim_threshold >= 0.6 or 0.7 or 0.8 was very small Figure 2 On the contrary, the 12 reactions with Confidence > 0.9 and predicted correctly had significantly more number of similar neighbor reactions Figure 3 The experimental results further demonstrated the necessity of combining confidence and screening similar reactions to improve the prediction results of reactions.

[0070] Those skilled in the art will appreciate that while the application has been described in terms of specific embodiments, the application can be practiced with modifications without departing from its spirit and scope. The particular embodiments and examples of the application are not provided to limit the application. The application is limited only by the claims. All documents cited are incorporated by reference in their entirety.

Claims

1. A method for predicting chemical reaction products, comprising the following steps: Step 1: predicting reaction products of different reactions and calculating the reactions with a confidence lower than a threshold value based on an original training model trained using an original dataset D0, screening these data and composing a first dataset D1; Step 2: providing a second dataset D2, calculating the similarity sim(w, v) of a chemical reaction W in D2 to a chemical reaction V in D1, screening the similar reaction supplement data in D2 with sim(w, v) greater than or equal to a threshold value, and obtaining a third dataset D3 in a set manner; Step 3: merging D3 data in the original dataset or using D3 alone to retrain the model.

2. The method of claim 1, wherein the step 1 comprises: obtaining one or more machine models capable of predicting reaction products and outputting the confidence of the prediction; calculating the confidence of the predicted product corresponding to each model in the original dataset by a given chemical reaction, and calculating the overall confidence of all models; screening the reaction data with a confidence less than a threshold value to obtain a first dataset D1; wherein the threshold value in the step 1 is any number in 0.3 to 0.

9.

3. The method of claim 2, wherein, The threshold value in the step 1 is any number in 0.4 to 0.

8.

4. The method of claim 2, wherein, The threshold value in the step 1 is any number in 0.5 to 0.

7.

5. The method of claim 2, wherein, The threshold value in the step 1 is 0.

5.

6. The method of claim 2, wherein, When product information is known, confidence = p(Y | X, θt); when reaction product information is unknown, Ymax = arg max i (confidence = p(Yi | X, θt)) gets (X, Ymax), i is the ith output prediction that the model can provide, t represents the tth model snapshot, t = 1, 2, …, K, K is the number of collected model snapshots, X represents the reactants of the chemical reaction, Y represents the products of the reaction, p represents the probability of the model output Y when X and θt are known, Y max is the predicted product of the model, and arg max represents the maximum value of all Y i probabilities.

7. The method of claim 1, wherein, The threshold value in the step 2 is any number in 0.1 to 1.

8. The method of claim 7, wherein, The threshold value in the step 2 is any number in 0.3 to 0.

8.

9. The method of claim 7, wherein, The threshold value in the step 2 is any number in 0.5 to 0.

8.

10. The method of claim 7, wherein, The threshold value in the step 2 is 0.6, 0.7 or 0.

8.

11. The method of claim 7, wherein sim(w, v) = sim(w=encoding(W), v=encoding(V)); wherein w, v are the encodings of input reactions V, W by model θt, respectively.

12. The method of claim 7, wherein w = f (W, θt) = [w1, w2, … wn], f(W, θt) is specifically the vector representation of a layer before the output prediction element is calculated by inputting the reaction W into the model θt through the parameters of each layer, wherein n belongs to the parameters of the model preset to represent the vector length; and the reaction V obtains v = f (V, θt) = [v1, v2, … vn].

13. The method of claim 12, wherein, n can be any number in the range of 2 6 =64 to 2 12 =4096.

14. The method of claim 1, wherein, The step 3 comprises: randomly sampling R times the data amount of D3 in the original training data D0, and merging D3 to generate a new dataset, and then retraining the reinitialized machine model parameters; or using D3 for fine-tuning learning, i.e., retraining the model θt on the D3 data for F≥1 iterations to continuously update the model parameters.

15. The method of claim 14, wherein, R can be selected from any number in 0.5 to max(1, |D0| / |D3|).

16. A device for predicting chemical reaction products, the device comprising: a first prediction module configured to predict reaction products of different reactions based on the original training model and calculate the reactions with a confidence lower than a threshold value, filter these data and form a first data set D1; a second prediction module configured to provide a second data set D2, for a chemical reaction W in D2, calculate the similarity sim(w, v) of W with a chemical reaction V in D1, filter the similar reaction data in D2 with sim(w, v) greater than or equal to a threshold value, and obtain a third data set D3 in a set manner; a third prediction module configured to merge D3 data in the original data set or use D3 alone to retrain the model.

17. The apparatus of claim 16, wherein, The first prediction module is specifically configured to: obtain one or more machine models capable of predicting reaction products and outputting the prediction confidence; calculate the confidence of the predicted products in each model through a given chemical reaction in the original data set, and statistically analyze the overall confidence of all models; filter the reaction data with a confidence lower than a threshold value to obtain a first data set D1; The threshold value in the first prediction module is any number in the range of 0.3 to 0.

9.

18. The apparatus of claim 17, wherein, The threshold value in the first prediction module is any number in the range of 0.4 to 0.

8.

19. The apparatus of claim 17, wherein, The threshold value in the first prediction module is any number in the range of 0.5 to 0.

7.

20. The apparatus of claim 17, wherein, The threshold value in the first prediction module is 0.

5.

21. The apparatus of claim 17, wherein, When the product information is known, confidence = p(Y | X, θt); when the reaction product information is unknown, Ymax = arg max, (confidence = p(Yi | X, θt)) gets (X, Ymax), i is the ith output prediction that the model can provide, t represents the tth model snapshot, t = 1, 2, …, K, K is the number of collected model snapshots, X represents the reactant of the chemical reaction, Y represents the product of the reaction, p represents the probability of the model output Y when X and θt are known, Y max is the predicted product of the model, and arg max represents taking the maximum value of all Y i probabilities.

22. The apparatus of claim 16, wherein, The threshold value in the second prediction module is any number in the range of 0.1 to 1.

23. The apparatus of claim 22, wherein, The threshold value in the second prediction module is any number in the range of 0.3 to 0.

8.

24. The apparatus of claim 22, wherein, The threshold value in the second prediction module is any number in the range of 0.5 to 0.

8.

25. The apparatus of claim 22, wherein, The threshold value in the second prediction module is 0.6, 0.7, or 0.

8.

26. The apparatus of claim 22, wherein, sim(w, v) = sim(w=encoding(W), v=encoding(V)); wherein w, v are the encodings of input reactions V, W by model θt.

27. The apparatus of claim 26, wherein, w = f (W, θt) = [w1, w2, … wn], f(W, θt) is specifically the vector representation of a layer before the output prediction element of reaction W input to model θt through parameter calculation of each layer, wherein n belongs to the parameter of the preset representation vector length of the model; reaction V obtains v = f (V, θt) = [v1, v2, … vn].

28. The apparatus of claim 26, wherein, n can be any number in the range of 2 6 = 64 to 2 12 = 4096.

29. The apparatus of claim 16, wherein, The third prediction module is specifically configured to: randomly sample R times the data amount of D3 in the original training data D0, merge D3, generate a new data set, and then retrain the reinitialized machine model parameters; or fine-tune learning using D3, i.e., performing F≥1 iterations of training on D3 data to continuously update the model parameters.

30. The method of claim 29, wherein, R can be selected from any number in the range of 0.5 to max(1, |D0| / |D3|).

31. An apparatus comprising a processor and a memory, the memory being configured to store a computer program, and the processor being configured to execute the compound reaction product prediction method according to any one of claims 1-15 according to the computer program.

32. A computer readable storage medium for storing a computer program for performing the compound reaction product prediction method of any one of claims 1-15.

Images