本文约3400字，建议阅读5分钟
本文介绍了使用机器学习筛选喷墨打印生物墨水的技术。Pharmaceutical inkjet printing is a highly flexible and intelligent way of making pharmaceuticals. According to relevant report statistics, the market size in this field will show exponential growth in the near future. In the past, methods for screening suitable bioinks were time-consuming and labor-intensive, and thus constituted one of the major challenges in the field of pharmaceutical inkjet printing. To solve this problem, the International Journal of Pharmaceutics: X published a research result that used a machine learning model to predict ink printability, with a prediction accuracy rate of 97.22%.
Key words: inkjet printing 3D printing random forest
According to a report released by Xinhua News Agency in 2022, the drug development cycle usually lasts 10-15 years, and the capital investment is about 1-2 billion US dollars. Its technological progress and iteration are very slow, especially for the most mature dosage forms in the pharmaceutical field. Take a solid dosage form as an example. Over the past 100 years, there has been no disruptive technology, and its production and marketing still face key obstacles such as the stability of active pharmaceutical ingredients, release kinetics and bioavailability.
Although traditional pharmaceutical methods are suitable for large-scale production of single preparations, in early clinical trials, dose-escalation studies are usually carried out on drugs to determine the best and safest patient doses. Therefore, for smaller batches of experimental drugs, Traditional pharmaceutical methods do not apply.
Due to its high flexibility and digital and continuous production process, 3D printing can slow down or even overcome the above obstacles to a certain extent in the design, manufacture and use of drugs.
Among them, pharmaceutical inkjet printing not only provides the ability to digitally control printing, but also has the advantages of dose control and free design, making it possible for personalized drug delivery. In pharmaceutical inkjet printing, the optimization of ink characteristics and printing effect have always been the focus of research. In the past, researchers used the Oneszog number to predict the printability of inks. However, this traditional prediction method is often inaccurate .
Recently, researchers such as Paola Carou-Senra from Universidade de Santiago de Compostela and Jun Jie Ong from University College London pioneered the application of machine learning models to predict biological ink printability, and successfully improved the prediction rate. The research has been published in the "International Journal of Pharmaceutics: X" journal, titled "Predicting pharmaceutical inkjet printing outcomes using machine learning".
Figure 1: The research results have been published in "International Journal of Pharmaceuticals: X"
Paper address:
https://www.sciencedirect.com/science/article/pii/S2590156723000257
Experiment overview
The Ohnesorge number is a dimensionless number used in fluid mechanics to measure the relationship between viscous force, inertial force and surface tension, and is mainly used to predict the printability of ink. In a printable formulation, an ink is generally considered printable when 0.1 < Ohnesorge < 1, ie 1 < Z < 10 (Z value is the reciprocal of Ohnesorge). In many exceptional cases, however, inks with Z > 10 are printable.
Figure 2: Histogram and boxplot of printable formulation Z and Ohnesorge
As shown in Figure 1, the Z values of the printable ink formulations in this study ranged from 1 to 62.2, and there were 68 sets of ink formulations with Z values greater than 10. It can be seen that it is not accurate to predict printability based on Z value alone. To improve the accuracy of printability predictions, the researchers used a machine learning model and compared the performance of several different models.
experiment procedure
data set
This research data set contains the research results of 75 English-language literature published between May 2000 and February 2022 collected from Google Scholar, PubMed, Web of Science, PubChem, the Handbook of Pharmaceutical Excipients (9th ed.), Also added 2 types of internal recipe information. Finally, the dataset has a total of 687 recipes. Among them, there are 636 kinds of formulas that can be printed, accounting for 92.6% of the total formulas, and 51 kinds of formulas that cannot be printed, accounting for 7.4%.
The data set includes parameter information related to the inkjet printing process, and the specific variables are shown in the following figure:
Table 1: Variables during printing
model development
In this study, the researchers developed three machine learning models: artificial neural networks (ANN), support vector machines (SVM), random forests (random forests, RF), and calculated their Cohen's kappa coefficient (the Cohen's kappa coefficient, kappa), coefficient of determination (coefficient of determination, R²) and mean absolute error (Mean Absolute Deviation, MAE) were compared to obtain the best prediction model. At the same time, during the experiment, the researchers also explored the combination of models, feature sets, and hyperparameters.
* kappa: kappa is an indicator used to test consistency, and can also be used to measure the effect of classification. It can detect whether the model prediction results are consistent with the actual classification. Its value ranges from -1 to 1, usually greater than 0, where 1 represents complete consistency, 0 represents random consistency, and -1 represents complete inconsistency.
* R² : Coefficient of determination, also known as coefficient of determination, coefficient of determination, this indicator is based on the decomposition of the sum of squares of the total deviation, and is used to illustrate a measure of the degree of fitting of the regression equation to the observed data . The higher the coefficient of determination, the better the fit to the observed data, and the smaller the coefficient of determination, the worse the fit.
* MAE: Mean Absolute Error, also known as Mean Absolute Deviation, represents the average value of the absolute error between the predicted value and the observed value, so it can accurately reflect the size of the actual forecast error. Its value range is [0,+∞), and it is equal to 0 when the predicted value is completely consistent with the real value, that is, the perfect model; the larger the value, the larger the error.
Hyperparameter Tuning
Although printable ink can be jetted, it may also produce satellite droplets. This shape will cause printing inaccuracy and is an important indicator for evaluating ink quality. At the same time, if the medicine in the ink is insufficient, the printed medicine cannot achieve the therapeutic effect, so the content of the medicine in the ink is also an important indicator for evaluating the quality. Therefore, in addition to predicting the printability of the ink, the model in this study was also used to predict the quality of the printable ink and the drug dosage in the ink.
In the experiment of predicting these two different aspects, the R&D personnel optimized the hyperparameters of the model. Among them, the default hyperparameters of the model in the initial state are as follows:
Table 2: Model default hyperparameters
Hyperparameters are some parameters that need to be manually set in machine learning algorithms. Usually, the value of hyperparameters has a great impact on model performance and generalization ability. Hyperparameter tuning (or hyperparameter optimization) is the process of automatically finding the optimal parameter combination through a certain method. The researchers first pre-defined a set of possible values for each model, then performed a 5-fold cross-validation grid search on the training set to determine the best hyperparameter values, and finally the optimized machine learning model was applied to the test set.
Experimental results
For predicting ink printability, the best predictive model is the RF model. The researchers noted that the optimized RF model had an accuracy rate of 97.22 percent and a kappa coefficient of 0.854, indicating that the model was highly accurate and reliable in predicting printability.
In terms of predicting ink quality, i.e. predicting whether a printable ink will produce satellite droplets or not, the best predictive model is the ANN model. The researchers pointed out that the optimized ANN model has an accuracy rate of 97.14% and a kappa coefficient of 0.74. Here they also emphasize that the kappa coefficient takes into account the possibility of getting the correct prediction by chance, so the data set (most of the data set is printable formula, and the proportion of non-printable formula is very small) will be relatively unbalanced, but even so, the model obtains A score and accuracy of 1 indicate that the prediction is reliable.
For pharmaceutical inkjet printing, the best predictive model for predicting drug dose is the RF model. Here, the researchers pointed out that if the feature set of predefined hyperparameters and material names is used, the RF model performs best; if the minimum threshold of the residual data set is removed, the best prediction model is the ANN model. The following figure shows the training of two models different datasets.
As shown in Figure 3, the distributions of measured drug doses were similar for both datasets, but the dataset used to train the RF model was larger in size and had a relatively high proportion of data with drug doses between 2.5–5.0 mg (drug injection typical range of drug doses used in ink printing). Therefore, although the ANN model performs slightly better than the RF model, the RF model is more suitable for optimization. The R² of the RF model after optimization was 0.800 and the MAE was 0.291, indicating that it could predict the drug dose within ±0.291 mg.
Figure 3: Plot of measured drug doses in training RF and ANN model datasets
In summary, the machine learning model can predict the printability and print quality of the printable ink, and can also predict the dosage of the drug. Among them, different algorithms and feature sets can also have different effects on the prediction results.
3D Printing of Drugs: Driving the Industry to Digital
3D printing provides greater flexibility in the early stages of drug development. Just by adjusting a few parameters, it is easier to change the shape, size, dose, release, etc. of the drug, speeding up the progress of clinical trials and shortening the time for new drugs . listing cycle. At the same time, it can also provide personalized medical treatment by tailoring the precise dosage for the patient.
In July 2015, American pharmaceutical company Aprecia launched the prescription drug SPRITAM (levetiracetam) instant tablets prepared by 3D printing technology for the treatment of epilepsy. This is the world's first 3D printed drug approved by the US Food and Drug Administration (FDA), marking that 3D printing of drugs has become a reality, and it also set off a wave of research on 3D printed drugs. Since then, Aprecia has successfully transformed into a pharmaceutical preparation technology platform company based on its own advantages, and actively cooperates with scientific research institutions such as the Purdue University School of Pharmacy and large multinational pharmaceutical companies to promote new drug research and development.
Looking at China, Nanjing Triassic Medicine, established in 2015, can be regarded as the leader in the field of 3D printing drugs in China. Triassic Medicine was co-founded by Dr. Cheng Senping and Professor Xiaoling Li, an expert and educator in the American pharmaceutical industry. The chain's proprietary 3D printing technology platform, the three 3D printing drugs developed by it - T19 (for rheumatoid arthritis), T20 (reducing the risk of stroke and systemic embolism in patients with non-valvular atrial fibrillation, etc.), T21 (for the treatment of ulcerative colitis), all of which have been approved by the US FDA for clinical trials and have entered the clinical stage.
Undoubtedly, 3D printing of drugs has built the foundation of digital pharmacy, and its long-term market demand is large and the prospect is promising. According to a report released by Grandview Research, the global pharmaceutical 3D printing market may show exponential growth in the near future. It is estimated that by 2030, the global pharmaceutical 3D printing market will reach 269.74 million US dollars, and the intensification of aging will undoubtedly create The giant window to personalized medicine. It is believed that in the near future, drug 3D printing will reshape the business landscape of biopharmaceuticals.
Reference link:
[1]http://www.news.cn/mrdx/2022-09/13/c_1310662292.htm
[2]https://www.pudong.gov.cn/019010005/20221026/741820.html
Editor: Wang Jing
