A Graph Convolutional Neural Network Approach to Antibiotic Discovery | Hacker News Reader

A Graph Convolutional Neural Network Approach to Antibiotic Discovery

welcometothejungle.com

70 pointswickwavy6 comments

This post is nicely written. We are from the Pande Lab @ Stanford and are very supportive of this work being done by Barzilay et al at MIT. We cite each others' papers for good reason.

If this type of research interests anyone here, Genesis Therapeutics -- Stanford spinout and a16z-backed -- is hiring!

Here is our HN "who's hiring?" post. Join us in using deep learning to discover medicines for patients suffering from devastating diseases.

Genesis Therapeutics | South San Francisco, CA | Onsite | Full-time | https://www.genesistherapeutics.ai

We’re a hybrid AI + biotech company, developing novel neural networks to predict molecular properties and using them to accelerate the development of new medicines.

- Core deep learning tech was invented by co-founder + CEO Evan Feinberg during his PhD at Stanford’s Pande lab (the lab that did Folding@Home). See the peer-reviewed PotentialNet paper: https://pubs.acs.org/doi/10.1021/acscentsci.8b00507

- Raised $4.1m seed round led by Andreessen Horowitz

- Our platform was validated in collaboration with a top-five pharma company, in a public paper: https://arxiv.org/abs/1903.11789

- In addition to strong software + AI talent, our small team has top drug discovery scientists who have collectively discovered eleven FDA-approved treatments

Here are our open roles:

- Software Engineer: https://jobs.genesistherapeutics.ai/apply/ILdjypU87G/Softwar....

- AI Engineer: https://jobs.genesistherapeutics.ai/apply/NAtx0CsIIL/AI-Engi....

Tech Stack: python, pytorch, postgres, docker, kubernetes, various computational chemistry libraries + tools

Please apply online, or email your resume to: ben@genesistherapeutics.ai

Graph convolutions are really powerful for handing structured data like chemical compositions. With the right corpus, I think this area is ripe for unsupervised feature representation learning approaches like what we've seen with BERT-like approaches and how they've dominated NLP in the past few years.

Side note: I worked with Kyle a few years ago on the MIT-MGH Deep Learning for Mammography project. I'm glad to see his work + brilliance being recognized.

Unrelated question for the AI experts here: What's the potential for AutoML solving problems with just X and Y. I understand that you'd need to search a large subspace but at what stage is it currently and can it solve problems as complex as the one here?

If with AutoML you mean substantial architecture search (as your question seems to suggest) then the answer is it’s feasible only for the largest players in the field. It’s possible to use some tricks but even then your standard cluster with a few hundred GPUs isn’t going to cut it for large scale problems. That said, any problem that e.g. google deems valuable enough can be meta-optimized to oblivion. Same holds if you have a huge pile of cash and think a few percent performance boost is worth it.

So AutoML is not smart yet? Its just about throwing more compute?

Nothing is smart yet. What people understand under automl is different but I’m referring to meta learning here, be it the architecture, the optimizer, the learning rate etc. Optimizing these things requires an outer loop in addition to training the architecture and for architecture search in particular this boils down to training lots of models and evaluating all of them. There are ways of making this search a bit better than random search with techniques such as regularized evolution (essentially evolution with a bias for younger individuals) or other tricks but all of them require huge compute resources.

Reply on news.ycombinator.com