TensorFlow models for Aptamers

Running these models

See the playbook (playbook.md) for instructions.

Binned and SuperBin Labels

As described in the manuscript the binned model used a ternary label to represent each sequence for each subexperiment based on the positive bead fraction. The ternary labels with fraction cutoffs were:

• 0 = [0, .1)
• 1 = [.1, .9]
• 2 = (.9, 1.0]

For the three subexperiments in PD round 2 these were labeled as 'low_3bins', 'med_3bins', 'high_3bins'.

These bin labels were then passed into the super_bin which created a label of either:

• -1: ambiguous
• 0: all 0
• 1: all 0 except low_3bins = 1
• 2: all 0 except low_3bins = 2 or low_3bins = 1, med_3bins = 1
• 3: low_3bins= 2, med_3bins=1
• 4: low_3bins=2, med_3bins=2 or low_3bins=2, med_3bins=1, high_3bins=1
• 5: low_3bins=2, med_3bins=2, high_3bins=1
• 6: low_3bins=2, med_3bins=2, high_3bins=2

For clarity, an ipynb 'create_binned_and_super_bin_labels.ipynb' has been included in this folder which can go from the provided pd_clustered_input_data_manuscript.csv to both the binned and super_bin labels.

Labeled axes

We use LabeledTensor[https://github.com/tensorflow/tensorflow/tree/v1.15.0/tensorflow/contrib/labeled_tensor] to keep track of meaningful axis name and labels on our tensors. The LabeledTensor objects defined in this module share a common set of labeled_tensor.Axis objects:

• batch_axis: an axis with name 'batch' and no axis labels denoting batches of training/evaluation examples.
• input_position_axis: an axis with name 'input_position' and labels giving the integer number of nucleic acid offsets from the 5-prime end (namely, [0, 1, 2, ..., sequencing_length - 1]).
• input_channel_axis: an axis with the name 'input_channel' and labels for the channels corresponding to each input feature (e.g., ['A', 'T', 'G', 'C'] if not using secondary structure).
• output_axis: an axis with the name 'output' and labels correpsonding to the name of each physical measurement/assay that could be used for either training or validation. By construction, labels match SequencingReads.name fields from the selection_pb2.Experiment proto in aptamers_paper1/selection.proto.
• logit_axis: an axis used to label each distinct prediction from the feedforward network part of the model. The name of this axis and its labels depend on the particular output layer model.
• target_axis: an axis with the name 'target' used to label loss targets for training. Labels for loss targets should be a subset of the measurements provided by output_axis.

batch_axis, input_position_axis, input_channel_axis and output_axis are produced by preprocessing routines in aptamers_paper1/learning/data.py. logit_axis and target_axis are created by output layer models in aptamers_paper1/learning/output_layers.py.