Excited to share @Nature: How does naloxone (Narcan) stop an opioid overdose? We determined the first GDP-bound μ-opioid receptor–G protein (wt) structures and found naloxone traps a novel "latent” state, preventing GDP release and G protein activation. 🧵 https://t.co/E2fZXw2Pom

Big thanks for all the support from @VanderbiltAIPD @VanderbiltCSB @VandyPharm @NIDAnews

I started my lab in April 2024, and I think we are starting to build some momentum. Hopefully in the next few months I will be able to share some other stuff we are working on.

Finally, I want to note that the peer review process improved this manuscript. While I understand the benefits of pre-prints and the limitations of peer review, this was an instance where it was genuinely constructive and elevated the final paper.

Also, apologies for being slow with it, but I'll add more scripts, examples, a better UI, etc. to the GitHub soon. I had to freeze a lot of the project a while back for purposes of benchmarks and picking a stopping point for the manuscript.

The new CORDIAL model will also be trained on substantially more synthetic null data covering broader chemical and structural perturbations. We will upload the weights for these new models, too.

Co-folding models are likely overtrained on pairs of sequences and chemical substructures, but for generating plausible structures for affinity prediction with CORDIAL, they are probably a better version of what I tried to do with the MCS maps and refinement.

Speaking of improvements, we're extending the training set with the new SAIR dataset from @SandboxAQ . Our original augmentation mimicked known poses via MCS mapping.

The code is open - feel free to explore.

Anyway, CORDIAL generalized pretty well, but perhaps not unexpectedly it did not yield dramatic performance improvements over Vina. There are clear ways to increase CORDIAL's expressivity - learning atom-pair embeddings/weights, incorporating additional geometric information, etc

This first pass at the CATH-LSO benchmark was useful, but in subsequent iterations I'll be tweaking it to make it more challenging. I hope this work encourages more dialogue on best practices for retrospective validation. I'm open to better strategies if folks have suggestions.

So, how do we evaluate generalizability? I tried to set it up to mimic screening against a member of a novel, unseen protein superfamily. I hold out a protein superfamily and its associated chemistry, train on the remainder, and test on the held-out set. Thank you CATH team.

This doesn't completely eliminate bias, but it reduces it and makes it more predictable. For example, signal magnitudes in feature columns can differ between train/test sets. Consequently, BatchNorm1d or something similar is required to prevent the model from over-training.

The approach here was to use a task-specific architecture. Instead of guiding the model to focus on interactions, we restrict its learning space to them. The model is constrained to view the problem only through distance-dependent physicochemical pairings.

The challenge is that the model needs a massive amount of data to guide it to learning the problem how we want. With a broad inductive bias, a model can easily learn non-causal correlations from training set artifacts instead of the generalizable principles we intend.

Often we have an idea of what we want the model to learn, and it is easy to assume that the network will tend to learn the problem the way that we consider it.

This manuscript is an exploration of learning spaces. In my lab, we think a lot about the spaces of things. A model's architecture defines the manifold on which learning occurs.

To be clear, I am not selling a model, I do not believe I have solved this problem, I am not suggesting you should scrap your existing tools to just use this. The paper introduces a model, CORDIAL, but it's not really about the model itself. So, what is this manuscript about?

Manuscript:

My first manuscript as an independent PI, and my first single-author research article, is now published in @PNASNews. It's an attempt to contribute to the dialogue on generalizability in structure-based protein-small molecule affinity prediction with neural networks.