Our previous proposal to CACHE3 was ranked joint 6th out of 33 entries by the CACHE3 peer review process. Here we describe a substantially enhanced version. CODASS4 relies on the data fusion concept of consensus to significantly boost the reliability of both its small molecule binding mode (herein referred to as “pose”) predictions, and its predictions of ligand binding likelihoods.CODASS4 is a sophisticated process, and is therefore best understood when visualised as a workflow diagram:
(please click on the link, or copy-paste to a web browser).
The improvements in CODASS4 include:
Consensus Docking - also known as consensus posing or multi-docking (and not to be confused with consensus scoring), consensus docking leverages multiple docking programs to substantially boost the reliability of the predicted poses that are fed into the downstream scoring schemes. Rigorous evaluation by multiple research groups (including ours, which made the initial discovery) has shown this to hold true under a variety of conditions; see our publications and the following, which all reference our Houston & Walkinshaw 2013 publication that originally reported the concept:
In addition to our standard CODASS usage of Autodock-GPU, Vina-GPU+ and GWOVina, three new pose prediction methods have been added to our consensus docking scheme:
SILCS-MC: This newly adopted MD approach is substantially cheaper computationally than conventional MD for large numbers of compound evaluations. The consideration of local target backbone flexibility and atom-level target-water interactions are significant advantages of this method over docking tools. https://doi.org/10.1021/acs.jcim.9b00210
DiffDock: Like SILCS-MC, the pose prediction accuracy of DiffDock has been shown to be agnostic of local variations in protein backbone as well as side chains within the target’s active ligand binding site pocket. https://arxiv.org/abs/2210.01776v2
DeepDock: The geometric deep learning approach of DeepDock learns a potential that is specific for each ligand-target complex. Thus, it can be retrained using the valuable information represented by the 895 ligands with known IC50s, producing a target-specific version of this algorithm. See more below on the advantages of training target-specific approaches. https://doi.org/10.1038/s42256-021-00409-9
SAIYAN: We will be exploring the use of SAIYAN, our latest DL high-throughput, structure-based binding prediction software. SAIYAN uses an exclusive approach allowing it to leverage over an order of magnitude more training data than previous ML docking-based tools, improving its accuracy and generality. Its architecture leverages geometric deep learning and GPU parallelism, enabling ultra-fast inference across multi-billion chemical libraries.
Classical Scoring Functions: a battery of methods based on force-field, knowledge-based, and machine learning approaches, with a proven track record in reliability (see our publications). An additional open-source method currently under late-stage in-house development in-house, RFRanker, will be added to this consensus scoring scheme.
Generic DL Scoring Function: An evolution of the DL-based Scoring Function SCORCH (which we recently published, see relevant publications), SCORCH 2.0 is itself a consensus method as it combines the ML/DL methods of its predecessor SCORCH 1.0 (GBDT using XGBoost, a FF NN, and a W&D NN) in a new way, namely by implementing a consensus model by average prediction. The result is superior screening, and ranking power as well as a higher throughput.
Target-specific DL Scoring Function: Target-specific Scoring Functions are generally much more predictive than generic SFs. http://dx.doi.org/10.1016/j.ddtec.2020.09.001; https://doi.org/10.1016/j.cbpa.2021.04.009. One example in the literature is a prospective study that used a Scoring Function based on a docking-based QSAR model trained with as few as 47 target-docked actives: https://doi.org/10.1016/j.ejmech.2014.01.019. The 895 known ligands of the CBLB enable the training of a target-specific Scoring Function, a trivial task for our group, given our experience with training four different DL SFs (see above), particularly as there are now tools to facilitate the generation and validation of such SFs:
Final consensus scores are calculated in two ways: 1. The conventional “Rank-by-rank” scheme originally reported to be superior by https://doi.org/10.1021/ci010025x and 2. The new “Exponential consensus ranking” method reported by https://doi.org/10.1038/s41598-019-41594-3
The second crucial element of CODASS is its Similarity Search component, which enables the screening of much larger virtual chemical libraries than structure-based methods alone. In addition, it enables the use of the 895 known ligands (with predicted binding poses supplied by our consensus docking scheme) to search for compounds with similar pharmacophores or protein-ligand interaction patterns but different chemical templates.
Free FEature POint PharmacophoreS (OpenFEPOPS): Our own open-source implementation of the FEPOPS scaffold-hopping molecular similarity evaluation method (https://doi.org/10.1016/j.jmgm.2007.02.005) has been integrated into CODASS4, complementing our existing 2D (FP2+Tanimoto), Graph (USRCAT) and 3D-based (Autodock-SS) similarity techniques for querying large chemical databases. This allows the identification of potential binders in accurate pose prediction regimes which would have been missed using other similarity matching techniques. Autodock-SS, USRCAT and OpenFEPOPS can all be considered “scaffold-hopping” methods, as they do not rely on molecular templates to perform their similarity evaluations and are thus suited to identify novel templates.