Lecture 11b Machine Learning and Bioinformatics

Introduction to Machine Learning and Bioinformatics

• Bioinformatics utilizes machine learning techniques to analyze and interpret biological data.
• Machine learning (ML) is a subset of artificial intelligence (AI) focused on enabling computers to learn from experiences.
• Deep learning is a subset of machine learning that makes the computation of multi-layer neural networks feasible.

Artificial Intelligence

• Artificial intelligence encompasses any technique enabling computers to mimic human behavior.

Machine Learning

• Machine learning is the study and application of algorithms to enable computers to learn from data without explicit rules.
• It employs a paradigm of providing training data to the computer system followed by testing with unseen data.
  • Supervised learning builds a model to transform input data into known outputs to make predictions. This includes classification and regression.
  • Unsupervised learning builds a model of data without known outputs. This includes clustering, pattern discovery, dimensionality reduction, and feature learning.
• Specific supervised learning methods include:
  • Support Vector Machine (SVM)
  • Linear Regression
  • Logistic Regression
• Specific unsupervised learning methods include:
  • K-Means Clustering
  • Hierarchical Clustering
  • Principle Component Analysis (PCA)

Deep Learning

• Deep learning utilizes artificial neurons to build layered models for data analysis.
• Applicable to both supervised and unsupervised learning.
• Deep learning models utilize various neural network architectures, including fully connected, convolutional, and recurrent networks.
• Current advancements in hardware, along with increased access to large datasets, have boosted the popularity of deep learning.
• Deep learning models often do not require significant feature engineering.

Artificial Neuron

• An artificial neuron takes multiple inputs.
• Inputs are multiplied by weights.
• Bias may be added to weighted sum.
• Weighted sum is processed by an activation function to propagate the signal.

Multilayer Perceptron (MLP)

• A type of neural network containing interconnected layers of artificial neurons.
  • The input layer receives input data.
  • The hidden layers process the data.
  • The output layer provides the final prediction/classification.
  • The loss layer measures the difference between predicted and actual values.

Activation Functions

• Activation functions introduce non-linearity to neural networks, allowing them to approximate any function.
• Specific activation functions include:
  • Sigmoid/Logistic - Output values between 0 and 1, centered on 0.5
  • Tanh - Output values between -1 and 1, centered at zero
  • ReLU - Output is either the input or zero, whichever is greater.

Feedforward Neural Networks

• A type of artificial neural network with no cycles within the network graph.
• Includes single-layer perceptrons and multilayer perceptrons.
• Employs back-propagation.

Backpropagation

• The most popular method for training deep neural networks.
• The difference between known results and model output is calculated as an error term.
• Error propagates backward through the network to adjust weights.

Fully Connected Neural Networks

• A neural network design where each neuron in a layer is connected to every neuron in the subsequent layer. The input layer receives data as tensors.
• Often used in preliminary stages of training before introducing more robust models such as convolutional networks.

Convolutional Neural Networks (CNNs)

• A popular class for processing image data.
• Composed of convolutional layers, pooling layers, and fully connected layers.
• Convolutional layers employ kernels (filters) of fixed size to analyze the input image.
• These kernels learn image features as a convolution with the image.
• The final layers learn classifications.
• CNNs excel at preserving spatial relationships of image data unlike fully connected networks.

CNN Learned Filters

• Convolutional neural networks learn filters that detect specific features in images.
• These filters are visualized as grayscale image examples.

CNN Classifier

• CNNs can classify images, such as identifying objects within a scene composed of multiple objects.

CNN for Medical Imaging

• CNNs are applicable to medical imaging data.
• They involve the use of convolutional layers (filters) to process images like MRI, X-ray imaging.

Transformers

• Introduced in 2017, Transformers are a state-of-the-art neural network architecture.
• Utilizes 'self-attention' to analyze data.
• Enables parallel execution or encoding of positional data.
• Replacing standard recurrent models in NLP and other tasks.

Transformers and Attention

• Analyzing input data, focusing on specific parts within data, and predicting outputs.

Receiver Operating Characteristic (ROC) Curve

• A graph visualizing the performance of a binary classifier system in terms of true positive rate against false positive rate.

Confusion Matrix

• A table visualizing the performance of a classification model summarizing correct and incorrect predictions into four buckets: true positives, true negatives, false positives, and false negatives.

Applications in Bioinformatics

• Deep learning is a growing field used in bioinformatics, capable of handling large datasets.
• Applications include DNA/RNA binding motif determination, SNP genotyping, and phenotype/genotype estimations.

AlphaFold2

• A deep learning model for protein structure prediction.
• Employs a sequence-based approach to predict protein structures.
• Accurately predicts protein structures from amino acid sequences.
• Used in protein structure determination.

AlphaFold3

• The latest iteration of AlphaFold, improving upon the previous version by incorporating and potentially refining previous prediction models and techniques.
• Improved accuracy in predicting protein structures to handle various situations in the target data itself.