The ExCAPE project is scaling up machine learning techniques to solve challenging problems on HPC machines. Our driving application is compound-activity prediction for drug discovery. We are preparing data, developing the state of the art for machine learning algorithms, and researching programming models to implement them. Our main achievements so far include:

• Public release of a data set that resembles industry data in terms of size and distribution of hits and misses, and experiments showing the potential of novel compound descriptors for doing multi-target predictions at scale
• Open source release of HyperLoom, a programming model and task execution system that runs on HPC systems and is designed to cope with machine learning
• Open source release of SMURFF, a matrix factorization package allowing sophisticated combinations of techniques such as GFA and Macau
• Exploration of the use of and hardware implications of sparse matrix techniques to deal with large sparse feature vectors, and the impact on deep learning
• Benchmarking of ML techniques on large scale data sets using HyperLoom
• Novel algorithms to improve scalability of matrix factorisation on large machines
• Demonstrating the computational requirements of conformal prediction
• Novel clustering implementations for pre-processing compound data

Project areas of international collaboration

• Experts in scheduling machine learning tasks on HPC systems
• Pharma companies to compare scalability of learning approaches
• Machine learning experts to explore scalability of different classes of algorithms
• Other users and developers of large scale multi-target learning