Automating the amino acid identification in elliptical dichroism spectrometer with Machine Learning

In order to train an ML algorithm to detect the right amino acid type in the EDS automatically, we require the absorption measurements from the EDS for the known amino acid samples. To collect such information, amino acids in lyophilized powder form were dissolved in phosphate-buffered saline (PBS, pH 7.0) to a final concentration of 1 milligram per millilitre. 200 microliters of the sample solution were loaded into a clean cuvette with a stopper lid and inserted into the cuvette holder in the EDS. Before beginning measurements, all electrical components of the EDS were turned on for at least 30 minutes to ensure stable readout. The predominant source of noise is the cost-effective photodetector, selected for its simplicity and clinical applicability, with optical components securely mounted to mitigate mechanical perturbations. In the EDS operating mode, absorbance measurements are taken in 5 degree increments through a full 360 degree range. This produces absorbance profiles known as absorbance sweep graphs, which are then used for further preprocessing techniques and ML. Additionally, blank measurements conducted under identical experimental conditions ensure robust normalization, thereby isolating intrinsic amino acid spectral features irrespective of solvent or pH variations. Also, post-normalization noise levels remain competitive, with the ED spectrometer exhibiting a consistent signal-to-noise ratio (SNR ? 14), as demonstrated in S1 Fig.

We collected about 628 absorption profiles of 17 soluble amino acid types in this way. Table 1 shows the distribution of the data collected across the 17 amino acid categories. It is evident from this table that the data distribution is skewed, with L-Alanine and L-Arginine having the highest number of absorption measurements. Each absorption profile contains 72 absorbance measurements captured at varying angles ranging from 0° to 355° with an interval of 5°. It is worth noting that the absorption values making these absorption profiles are derived from the logarithmic ratio of light intensities and are dimensionless quantities.

After the necessary data preprocessing procedures, we used the preprocessed absorption profiles of known amino acid samples to investigate traditional ML algorithms like Random Forest [21], Decision Tree [22] and XgBoost [23]. We also delved into designing neural networks optimized for performance. This paper’s Results and Discussion section elaborates on this neural network algorithm’s design.

As each of the absorption profiles is associated with an amino acid category, which acts as a class label that is essential for training a classification ML model to predict the correct class for new, unseen data, proper encoding techniques were applied to ensure that these class labels were converted into a format that ML algorithms understand. For this purpose, one-hot encoding of class labels was performed to train neural networks and label encoding of class labels was performed to train traditional ML algorithms.

One-hot encoding, implemented using Pandas 1.5.1, transforms each class label into a binary vector representation. For instance, to classify between 17 amino acid types in our investigation, we created a binary vector of length 17 for each absorption profile. The element corresponding to the class label of the absorption profile was set to 1, while all other elements were set to 0. This element, represented with a 1, indicates the presence of the associated amino acid category. On the other hand, label encoding, implemented using Scikit-Learn 1.1.3, transforms each class label into a numerical format. Each unique class label was assigned a numerical value ranging from 1 up to 17, denoting 17 amino acid types. Therefore, for each absorption profile belonging to a specific amino acid category, the corresponding numerical value of that category was assigned as its class label.

Table 2 summarizes our results in classifying all 17 amino acid types across different ML algorithms. Accuracy (Acc), rounded off to two decimals, obtained across the three data splits (train, test and validation datasets) is reported in this table. In ML, Acc is a metric used to evaluate the performance of a classification model, and it represents the ratio of correctly predicted instances to the total number of instances in the dataset. Additionally, since the neural network we investigated in our study contains dropout, the reported train Acc is the accuracy obtained in the training dataset with the saved best model used to report validation and test Acc. While training, the train Acc obtained was 0.08 (rounded off to two decimals).

The lower performance observed in our study can be attributed to the limited availability of experimental data, especially for certain amino acid categories like L-Tryptophan. This lack of adequate data poses a significant challenge for ML algorithms to effectively learn discriminative features and generalize well to unseen data. Despite setting class weights to address class imbalances, the dataset’s small size severely constrains the models’ learning capacity. It undermines their ability to extract meaningful patterns and relationships from the data.

Therefore, to study if utilizing ML for automatic amino acid identification is feasible, our next goal was to classify amino acids, L-Alanine and L-Arginine, which had the highest data count among all the 17 amino acid types. Moreover, although the overall performance was low, the investigated algorithms showcased higher performance in identifying these two amino acid categories, owing to the availability of relatively higher amounts of data acquired for these amino acid groups. Additionally, since there was inadequate data in most of the amino acid categories, using different neural network models with varying hyperparameters didn’t improve the results of classifying all 17 amino acid types. Therefore, the best neural network model we obtained in classifying L-Arginine and L-Alanine was used to report the result in Table 2. The architectural details of this model are elaborated in the subsequent sections.

Table 3 summarizes our results (in terms of accuracy rounded off to two decimals) in classifying L-Arginine and L-Alanine across different algorithms, and Table 4 summarizes the individual amino acid accuracy (rounded off to two decimals) obtained. As stated earlier, since the neural network we investigated in our study contains dropout, the reported train Acc is the accuracy obtained in the training dataset with the saved best model used to report validation and test Acc. While training, the train Acc obtained was 0.79 (rounded off to two decimals). Fig 4 illustrates the configuration of the DL model utilized in our study. The model was structured with specific hyperparameters tailored to optimize model performance. Furthermore, this experiment did not incorporate class weights as L-Alanine and L-Arginine do not showcase class imbalance and contain an almost equal number of experimentally collected absorption profiles.

As depicted in the Fig 4, we utilized a combination of popular DL operations such as Conv1D, BatchNorm, ReLU, Maxpool1D, Flatten, Dropout and Dense layers to design the neural network model used to classify between L-Arginine and L-Alanine. Conv1D layers are specialized neural network layers designed to extract features from one-dimensional sequential data, such as time series or text. They utilize a learnable sliding 1D matrix to scan the input sequence to find relevant patterns. BatchNorm layer is used to improve the stability and performance of neural networks by normalizing the outputs of each layer. ReLU is an activation function commonly used to introduce non-linearity into the model. MaxPool1D is a downsampling operation commonly used to reduce the spatial dimensions of feature maps (or inputs) while retaining the most salient information. A Flatten layer is used to reshape a tensor to a 1D vector suitable to input to Dense Layers. Dropout Layer is used to prevent the model from overfitting, which is a scenario in which the model memorizes the train data and performs poorly on unseen test data. Dense layers are typically employed in the final stages, transforming the high-level features learned by earlier layers into predictions or classifications. Additionally, the Input Layer is where the absorption profiles are passed as input and the Output Layer is where the model outputs a prediction (or the amino acid type) for a given input absorption profile.

We manually tuned the hyperparameters such as the number of filters (#), activation functions, stride (s), kernel size and number of units in dense layer to incur the highest possible performance in distinguishing L-Alanine and L-Arginine. Additionally, we observed that decreasing or increasing the number of dense layer in combination with increasing or decreasing the number of convolution layer set—Conv1D, BatchNorm, ReLU and Maxpool1D, didn’t result in increase in performance. Furthermore, we observed that increasing the number of these layers or the layer set just spiked the total number of parameters and resulted in the overfit of the model. Thus, the neural network model used to report the results in Tables 3 and 4 was chosen as the optimal model in our investigation.

From the results reported in Tables 3 and 4, it is clear that the traditional ML algorithms demonstrate overfitting. Overfitting occurs when a model learns to memorize the training data rather than capturing underlying patterns that generalize well to unseen data. It is evident from these tables that the traditional ML algorithms have higher Acc on the training dataset in comparison to the unseen test and validation datasets. However, this is not observed in the reported performances across these three datasets (train, test, and validation) of the DL model. This demonstrates that traditional ML algorithms, such as Random Forest, Decision Tree, and XGBoost, are susceptible to overfitting when trained on datasets with limited size or noisy features. Additionally, the good performance of the DL model in classifying L-Arginine and L-Alanine proves that with adequate data, DL can be leveraged to classify all 17 amino acid types in the future.

As we successfully classified between L-Alanine and L-Arginine, having the highest number of acquired absorption profiles in the dataset, with high performance (of about 80 percent test accuracy), we wanted to investigate the classification of L-Alanine, L-Arginine, and L-Lysine, as L-Lysine had the next highest count of absorption profiles experimentally collected. Thus, this section discusses our results in classifying these three amino acid types.

Table 5 summarizes our results (in terms of accuracy rounded off to two decimals) in classifying L-Arginine, L-Alanine and L-Lysine across different algorithms, and Table 6 summarizes the individual amino acid accuracy (rounded off to two decimals) obtained across these algorithms. Additionally, since the neural network we investigated in our study contains dropout, the reported train Acc is the accuracy obtained in the training dataset with the saved best model used to report validation and test Acc. While training, the train Acc obtained was 0.64 (rounded off to two decimals). In this investigation, we incorporated class weights during training as the L-Lysine amino acid type had only one-third of the experimentally collected absorption profiles compared to the other two amino acid groups. Each amino acid group in this experiment was assigned a class weight, which was calculated as the total number of experimental records/absorption profiles in the training dataset divided by the number of records in that class.

The results show that ML algorithms perform poorly in identifying L-Lysine compared to L-Alanine and L-Arginine. The ML algorithms’ inability to learn features specific to L-Lysine can be attributed to the inadequate absorption profiles collected for this amino acid type compared to L-arginine and L-alanine. The other amino acid groups, L-Alanine and L-Arginine, contain almost thrice the number of training data as L-Lysine. This further shows that ML algorithms’ performance in identifying amino acids can be improved with adequate data.

We further investigated if the DL model trained to classify between L-Alanine and L-Arginine can effectively learn features of L-Lysine through a transfer learning approach. Usually, in ML, when the size of the acquired dataset to train a DL model for a particular automation task is small, giving rise to poor performance in the task, we leverage the capabilities of good-performing state-of-the-art DL models trained with ample amounts of data for related ML tasks, through a process called Transfer Learning [24].

In the transfer learning approach, we freeze the weights of the architecture, except for the last few layers, of the good-performing state-of-the-art DL model, trained for an ML task with ample amounts of relevant data. Then, we tune the unfrozen layers to perform well on a related ML task, even if the amount of data acquired for this task is less. As the frozen weights of the architecture are optimized to capture features of a dataset specific to an ML task and have been trained with ample amounts of relevant data for high performance, it transfers its knowledge to improve the performance of other related ML tasks even though the size of the acquired data for these tasks are small.

In section 3.2, we demonstrated the architecture and the results of a DL model trained for optimal performance to classify between L-Alanine and L-Arginine. By showcasing high performance in distinguishing these two amino acid types, this DL model has learnt features specific to identifying amino acids. Therefore, in this section, we investigate if the knowledge that this DL model acquired can be transferred to improve the performance in identifying L-Lysine, which has the least amount of data compared to L-Alanine and L-Arginine. Furthermore, we hypothesized that if this approach works well in identifying L-lysine, we could apply the same technique to identify other amino acid types with relatively fewer absorption profiles than L-Lysine.

To study the capability of utilizing transfer learning to improve the performance of L-Lysine identification, we froze the weights of all the layers in the DL model (used for classifying L-Alanine and L-Arginine) except for the output layer. Then, we set the size of this output layer to 3 (to output a class between L-Alanine, L-Arginine and L-Lysine) and trained it with the acquired data. This enables the output layer to tune its weights with respect to the weights of the frozen layer, which are already tuned to learn the features related to identifying amino acid categories. Tables 7 and 8 demonstrates the performance (in terms of accuracy rounded off to two decimals) of this model in classifying L-Alanine, L-Arginine and L-Lysine. Class weights were incorporated while training, where class weights for each class were calculated as elaborated in the section 3.3. Additionally, since the neural network from which the layers’ weights were freezer in our study contains dropout. Therefore, the reported train Acc is the accuracy obtained in the training dataset with the saved best model used to report validation and test Acc. While training, the train Acc obtained was 0.57 (rounded off to two decimals).

It is evident from Table 7 that the model has generalized well to unseen test data. However, its performance in identifying L-Lysine is poor, as demonstrated in Table 8. This behaviour is clearly due to a lack of data. The accuracy scores in the Table 7 clearly demonstrate that the amount of data available for training, testing and validating L-Lysine is insignificant as compared to the other two dominant amino acid categories, which heavily contributed to the higher and generalized accuracy scores. This further reinforces that we could classify other amino acid types if we acquire more data, and the approach to automate the amino acid identification using ML is feasible.