Physics-Enhanced QUantum Quantitative ADME
Pequqa was developed in response to the OpenADMET2025 ExpansionRX Leaderboard challenge.
https://huggingface.co/spaces/openadmet/OpenADMET-ExpansionRx-Challenge
What’s in Pequqa?
- DFT Features from Promethium by QC Ware
Molecules are processed using Promethium’s GPU-accelerated DFT engine (http://promethium.qcware.com) via 3D conformation search, followed by geometry optimizations, and single-point calculations performed in various solvation states to generate descriptive features such as partition energy, hydration energy, polarizability, HOMO-LUMO gap etc.
This level of throughput is made possible due to the high-level of GPU-acceleration in the electronic structure engine. In most cases the DFT calculations complete in under a minute.
- Auxiliary Features
Auxiliary features were generated using publicly available datasets for BBB (https://github.com/theochem/B3DB) , Aqueous Solubility (https://github.com/mcsorkun/AqSolDB) and LogD (ChemBLdb29). Scikit-learn Gradient Boost was used to generalize these features from Morgan, Avalon and RDKit-2D descriptors (Mordred fingerprints were also investigated but didn’t improve the auxiliary models and were also very slow to calculate). Chemeleon (https://github.com/JacksonBurns/chemeleon) message-passing weights were also retrieved and PCA was used to reduce the dimensionality.
- Advanced Neural Network Models
After experimenting with many models from home-made NNs to Chemprop, TabPFN (Prior Labs https://github.com/automl/tabpfn) was selected for the final iteration of the model as it works very well without needing any additional tuning. 30 replicates were trained using the raw training data set with a randomized 70:30 train:test split for each replicate.
Pequqa v6, powered by Promethium and Chemprop with Chemeleon:
Chemprop models using the Chemeleon foundation acting on the multi-target set worked reasonably well for LogD as well as MP, MB and MG: just as good as the TabPFN models. However, MLM, HLM, and Caco-2 PE/PP were not well described in the multi-target mode. MLM, KSOL and HLM worked well in multi-target mode once the Log scale was used for those variables. Pequqa v6.1 uses these with approach with all variables transformed to Log scale
Pequqa v7, powered by Promethium and TabPFN with Chemeleon:
TabPFN immediately impressed with getting very high R2 on training and test for most variables. Caco-2, HLM and MLM were still very difficult to describe, but were better with TabPFN than Chemprop. The main variables tuned were linear or log scaling of the targets, and the depth of the PCA reduction of the Chemeleon message passing weights. Pk scaling of MG, MP, MB was also investigated. The difference between 7.0a and 7.1a is (a) HLM CLint, KSOL, MLM Clint and Caco-2 Perm. Papp A>B are log-scaled in 7.0a, whereas all variables are in raw form (as downloaded) in 7.1a, and (b) 7.1a uses 70:30 train:test splits for the TabPFN.
What matters to Pequqa?
To determine whether or not the DFT features were significant to the model performance, a permutation importance feature analysis was performed. The feature analysis was performed on a lower dimensional model (Pequqa v7.0 with PCA-8) to expedite the slow process of the feature analysis. We find that the DFT contributed descriptors play a significant role in the model performance when assessed with permutation importance sampling.
| Target | Top 5 Features (most important first) |
|---|---|
| Log D | Auxiliary LogD, Chemeleon/PC3, Polarizability (DFT), Fraction CSP3 (RDkit), Chemeleon/PC4 |
| KSOL | Polarizability (DFT), Chemeleon/PC3, Auxiliary LogD, Fraction CSP3 (RDkit), Volume (DFT) |
| HLM CLint | Volume (DFT), Chemeleon/PC4, Chemeleon/PC2, Fraction CSP3 (RDkit), Chemeleon/PC0 |
| MLM CLint | Volume (DFT), Chemeleon/PC4, Chemeleon/PC2, Fraction CSP3 (RDkit), Chemeleon/PC0 |
| Caco-2 Perm. Efflux | Partition Energy (DFT), Hydration Energy (DFT), Number of H Donors (RDkit), Auxiliary LogD, Chemeleon/PC7 |
| Caco-2 Perm Papp A>B | Polarizability (DFT), Number of H Donors (RDkit), Auxiliary LogD, Partition Energy (DFT), Chemeleon/PC0 |
| MPPB | Polarizability (DFT), Fraction CSP3 (RDkit), Auxiliary LogD, Chemeleon/PC4, Chemeleon/PC5 |
| MBPB | Polarizability (DFT), Fraction CSP3 (RDkit), Chemeleon/PC0, Auxiliary LogD, Volume (DFT) |
| MGMB | Polarizability (DFT), Volume (DFT), Auxiliary LogD, Fraction CSP3 (RDkit), Chemeleon/PC0 |
