Helpful, but unless there's an ML/AI form of docking scoring function that's also new the resulting bound conformations are helpful but not really useful for predicting activity. Enough to get started - but can this technology find docked poses for structures or pockets that do not have (high?) similarity to the training set?
As has been said somewhat sadly at various conferences, "Scoring functions suck". They can be trained within (close?) families, they can show reasonable poses when enthalpies predominate but non-bonded/grease interactions were and probably still aren't a strength. It's a start until a crystal structure is generated and will help if/when the sar goes to hell and another is required.
>... Existing methods are computationally expensive as they rely on heavy candidate sampling coupled with scoring, ranking, and fine-tuning steps. We challenge this paradigm with EquiBind, an SE(3)-equivariant geometric deep learning model performing direct-shot prediction of both i) the receptor binding location (blind docking) and ii) the ligand's bound pose and orientation. ...
I think they used many nice words and in the right order, but the impact this will have on the rate of good drugs being approved is close to zero.
That's both because the technology itself doesn't really produce good leads, but because even if you find good leads, there are many other high probability reasons your drug will fail that have nothing to do with its safety and efficacy, and no ML model that exists today can address that.
Remember the dream of modeling - get to market 3 months sooner and you make 1-2B USD. Or, defend a patent or help evade/break one. For blockbuster drugs, 90 days off the multi-year drug development process is serious coin. And we'll need such speed when paxlovid starts encountering resistance.
I work in pharma and used to work on DOCK (the UCSF code). Nobody in pharma believes that speeding up docking is making a difference to get to market. Well, not nobody, but certainly not the decision makers. We all wish there was "Press Button, Spend Money, Get Leads Quick".
It's useful - it's doing the docking calculation as a statistical fit. There are a number of ML models such as fitting small-molecule structures to high-level theory calculations to make conformer search go faster. However, the pose is just a prediction - the scoring function has to estimate how good the pose is and hopefully the numbers are valid comparing structures and poses against each other (the dreaded what do I make next question). It can provide insight into what's interacting in the binding, which suggests changes to the medchemist. It might also help with understanding why curious results might appear.
6 comments
[ 3.4 ms ] story [ 28.9 ms ] threadAs has been said somewhat sadly at various conferences, "Scoring functions suck". They can be trained within (close?) families, they can show reasonable poses when enthalpies predominate but non-bonded/grease interactions were and probably still aren't a strength. It's a start until a crystal structure is generated and will help if/when the sar goes to hell and another is required.
>... Existing methods are computationally expensive as they rely on heavy candidate sampling coupled with scoring, ranking, and fine-tuning steps. We challenge this paradigm with EquiBind, an SE(3)-equivariant geometric deep learning model performing direct-shot prediction of both i) the receptor binding location (blind docking) and ii) the ligand's bound pose and orientation. ...
That's both because the technology itself doesn't really produce good leads, but because even if you find good leads, there are many other high probability reasons your drug will fail that have nothing to do with its safety and efficacy, and no ML model that exists today can address that.