Machine Learning is an emerging field which draws on and embraces research in artificial intelligence, robotics, statistical learning, data driven analysis and modelling. Our machine learning consulting services are focused on the techniques that are related to drug discovery research and biotech. Machine learning can be an option for answering the following questions:
- How can I get an overview of this large amount of data?
- How can I make predictions with new data?
- Are there any non-linear or non-obvious correlations in the data?
- Which features of the available data are most important?
- Is there a correlation in my data I have overlooked?
Benefits of using Machine Learning
Implementation of machine learning can help businesses with better utilization of the available data. Data amounts and production are increasing. Modern projects accumulate more data than ever before, which is often archived in databases. Machine learning can simplify complex datasets and make it possible to focus on the bigger picture. Patterns can be discovered in increasingly large and diverse datasets. With machine learning techniques, the complexity of the data can be reduced and the provided overview leads to better decisions.
Resources can be saved when new project data highlights similarities to a previous projects. Those teachings and results can be included in the new projects.
Machine learning can save time and lead to faster decisions and feedback cycles. Previous processes can be automated and results of the analysis will instantly become available for strategic decision making. Machine learning can build predictive models and thus suggest the most likely outcomes for the given new data. This leads to better decisions because routine non-machine evaluations processes and procedures may not always take the whole accumulated knowledge into consideration.
When the structure of the data is better understood, machine learning can lead to more innovative decisions and ideas. The models can find non-obvious and non-linear correlations in multidimensional data. Analysis of the machine learning models and feedback about their decision processes can highlight these correlations.
If you suspect that machine learning could be an interesting option for your project or business, contact Wildcard Pharmaceutical Consulting for a free and confidential chat about the possibilities.
Benefits of using External Machine Learning Consulting
- External Input and Experience: Get fresh perspectives to your project and data
- Get Scalable: Flexible resources can be scaled to project needs and the expertise is thus accessible also for smaller research organizations.
- Optimize your investment: Lower overhead costs by avoiding buying and maintaining hardware and software, as well as maintaining expensive in-house expertise and training.
A use case: Active learning in vHTS
An example use case for machine learning in a drug discovery setting is to use active learning in a screening campaign. The goal for all screening campaigns is to find high number of actives with as few tested compounds as possible. In a more traditional set up, the dataset to screen may be selected by a number of filters and-or an activity model trained on the activity data available from the beginning of the screening project. With active learning the computer gets a say in what compounds should be tested in the next batch of testing, and this opens op an exploration option, where the initial goal may not be to find the maximum number of compounds, but rather to improve the underlying model as much as possible. This leads to another screening dynamic, where more new scaffolds are found, rather than just finding compounds more or less similar to the initial known actives. This opens up larger possibilities for scaffold hopping and thus circumventing a bias to the known actives which could already be patented. This approach has been tested in a retrospective study with very promising results.
Machine Learning in more depth
Machine learning is a technique for getting computers to act and make decisions without being explicitly programmed. It is a data driven approach for making predictive models or finding structure in vast amounts of data. Therefore it is much used in conjunction with large amounts of data and Big Data.
Supervised and Unsupervised Machine Learning Algorithms
There are two fundamentally different classes of machine learning.
In supervised learning the machine learning application is trained to predict an already known outcome from the parameters available in the data. The machine learning program is trained on compounds with known toxicology. For instance the application has to predict whether the molecule is toxic or not from available physico-chemical data and structural features of the molecule.
Another machine learning example is from the field of chemometrics. There the concentration of a given compound is predicted from spectroscopic data. In this application the machine learning algorithm can overcome matrix effect or use relative non-analytical chemical information, as the overtone regions in near infrared spectroscopy.
The other main class of machine learning applications is unsupervised learning, where outcome is not known. The computer is trained to find patterns in the data presented to it. The outcome can be used to find correlations and trends in the data. This may not be directly visible. …. and make plots of multidimensional data after dimensional reduction to two latent variables as is done in principal components analysis.
The supervised learners can then further be subdivided into discriminant or classification algorithms which tries to assign discrete classes to the outcome in contrast with the regression based machine learning models which tries to predict a continuous variable. An example of a classification example could be as in the example before whether a compound is predicted to be toxic or not, where an example of a regression application could be to make the computer predict the affinity of a molecule towards a given protein receptor.
Linear/Non-linear machine learning models
Many machine learning models can also overcome the inherent limitations of the linear thinking that humans are taught in schools. It is thus also a fundamental feature of the algorithm if its linear or non-linear. Many of the linear models can be made to handle non-linear data through the use of a kernel trick. The choice of model has a huge impact on the usability and applicability of the developed machine learning algorithm.
Wildcard Pharmaceutical Consulting can help with consulting about what machine model fits the nature, quality and amount of your data
Preprocessing of data for machine learning applications
The efficiency of the machine learning algorithm is heavily influenced by the way data is presented. With data preprocessing this can be optimized. It can be simple scaling operations to ensure that each parameter gets presented with an equal weight to the algorithm. Calculating cross terms between parameters or applying non-linear transformations can help with transforming a non-linear problem into a linear one.
Preprocessing is often domain specific, such as the transformation and alignment done in chemometrical handling of spectroscopic data, or the choice of chemical fingerprint or physico-chemical properties of molecular structures when developing QSAR models with machine learning algorithms. This part is where the experience and knowledge of the human developer really makes a difference to the outcome.
Wildcard Pharmaceutical Consulting has special insight into the chemical, biological and pharmaceutical research domain. We can consult and apply the most appropriate treatment of your data.
Neural networks deserve individual attention. They are modeled based on the same principles as biological neurons and are processing information by interacting with each other. Initially the idea was to create artificial intelligence by mimicking the way the brain works. Neutral networks has found wide usage in visual and image analysis, which are good generalist non-linear approximators. The models work by combining multiple neurons together into layers, where all neurons interact with all neurons in the next layer. The training of the network is then progressed and the weights of the interactions between one neuron to the next and the sign of the interaction (inhibitory or excitatory), can be fit so the neural networks performs best for the optimization goal at hand. The choice of network (number of layers and number of neurons) and the way they interact can be used to tune the complexity of the network. The number of parameters in the model rises exponentially with the number of neurons. Lot of data or a simple network is usually needed to limit over-fitting.
Deep learning is an emerging buzz word, covering the idea to use hierarchical layers of learners to provide higher and higher abstractions. Most often it is used in conjunction with special architectures and layouts of neural networks.The first layers can be trained in an unsupervised way using a large general dataset. The output from the first layers provides an abstraction of the features in the general dataset and can subsequently be used as input for training layers in a supervised fashion. The dataset for the supervised part can be much smaller than the dataset used for the unsupervised deeper layers. This results in a much more efficient use of the data available. In drug discovery we see a similar situation with a large amount of known molecules, but often rather limited datasets available with information about bioactivity.
Computers as Creative Tools
A special case of neural net is the recurrent neural network, which learns from sequences of data and stores a model of sequences that are more likely to appear. They can after training be used to generate new likely sequences that are similar to the ones found in the training data. It is spooky to watch how a recurrent neural network with an long short-term memory (LSTM) architecture can write something that resembles Shakespeare in style and format, but without making any sense at all . In a drug discovery setting this could be used to generate molecules resembling the ones in the dataset and thus be a source for computational creativity. By tuning the generation temperature, the model output can be set to be more or less conservative in the generation, from being almost similar to the training set to being complete random nonsense.
Recommended Introductory Course
If you want to dive into developing machine learning programs yourself, Wildcard Pharmaceutical Consulting recommends the excellent Machine Learning introductory course from Stanford University by Professor Andrew Ng.
Contact Wildcard Pharmaceutical Consulting today for an informal chat about how you can leverage the power of machine learning for optimizing your business data usage.