jrieke / lstm-biology

Neural networks with LSTM to classify and predict biological cell movement

lstm-biology

Code and report for "Applying LSTM neural networks to biological cell movement" (project at the Biophysics Group, University of Erlangen-Nuremberg).

Abstract: Neural networks with Long Short-Term Memory (LSTM) were used on scientific time series data. Each time series contains the positions of a biological cell while it moves through one of three environments (a collagen network, a plastic surface, or a plastic surface coated with fibronectin). The networks were used for two tasks: 1) Classifying the movement trajectories based on the cell environment. Several networks of increasing complexity were trained on parts of the trajectories, using softmax classification. The best networks achieved an accuracy of ~95 % (on test data) and generalized well to longer trajectories. 2) Generating new movement trajectories by predicting one step of a time series after another. For this purpose, LSTM was combined with the idea of a mixture density network (MDN): It does not predict the values of the next time step directly, but outputs the parameters of a mixture distribution, from which they can be sampled. The generated trajectories replicated the shape as well as the rough statistics of the original dataset.

Requirements: keras (v0.3.2), Theano (v0.8.0.dev0), numpy, matplotlib, jupyter (optional for computing statistics of generated trajectories: bayesloop, seaborn, scipy)