Immunogenicity is one of the most serious problems facing the design of new biologics. In a nutshell, immunogenicity is when the patients immune system mounts a response against the introduced biologic. There are a number of reasons this can happen, such as contamination of the biologic (for example, deamidation can create isoaspartic acid "residues" on biologic, triggering an immune response). See this detailed FDA review:

https://www.fda.gov/downloads/Drugs/DevelopmentApprovalProcess/SmallBusinessAssistance/UCM408709.pdf

It seems likely that deep learning could help model immunogenicity. Are there any public reports of data available?

https://www.nature.com/articles/s41598-017-04439-5

A cool paper that shows how a simple heuristic can find disease motifs in a dataset of B-cell receptor sequences drawn from (roughly) 50 healthy patients and 50 unhealthy patients.

A simple deep-learned version of such a model might use a 1D conv-net to predict from sequence to patient healthy/unhealthy state.

Would it be possible to get access to the data as a public resource?

http://www.pnas.org/content/114/5/944.full

Contains a large (by standards of field :-)) dataset of all antibodies tested in phase-2 or phase-3 clinical trials measured "criteria regarding the feasibility of their manufacture, stability in storage, and absence of off-target stickiness."

This could be a valuable dataset for benchmarking antibody algorithmic techniques.

http://opig.stats.ox.ac.uk/webapps/sabdab-sabpred/Welcome.php

A useful overlay over PDB pulling out all antibody structures. Could be a useful source for building learning algorithms on antibodies.

Linking discussion from deepchem/deepchem#616

Regular graph convolutions only draw in information from an atom's direct neighbors. Atrous graph convolutions could draw in information from an atom's bond-distance 2 or 3 neighbors. This allows the receptive fields of graph convolutions to grow much larger. Note that a 4 layer deep graph convolution has visibility 4 bonds out, while a 4 layer deep 2-atrous graph convolution has visibilty 2^4=16 bonds out. This long range visibility could be a key enabler for working with biologics, which are much bigger molecules (protein antibodies typically) than small molecules.

PRs for atrous graph convolutions either here or in deepchem/deepchem welcome!

http://onlinelibrary.wiley.com/doi/10.1002/prot.24756/full

Would be worth reimplementing, but looks like data uses proprietary dataset.