Guillaume Godin

Igor V. Tetko, Ruud van Deursen, Guillaume Godin

Hyperparameter optimization is very frequently employed in machine learning. However, an optimization of a large space of parameters could result in overfitting of models. In recent studies on solubility prediction the authors collected seven thermodynamic and kinetic solubility datasets from different data sources. They used state-of-the-art graph-based methods and compared models developed for each dataset using different data cleaning protocols and hyperparameter optimization. In our study we showed that hyperparameter optimization did not always result in better models, possibly due to overfitting when using the same statistical measures. Similar results could be calculated using pre-set hyperparameters, reducing the computational effort by around 10,000 times. We also extended the previous analysis by adding a representation learning method based on Natural Language Processing of smiles called Transformer CNN. We show that across all analyzed sets using exactly the same protocol, Transformer CNN provided better results than graph-based methods for 26 out of 28 pairwise comparisons by using only a tiny fraction of time as compared to other methods. Last but not least we stressed the importance of comparing calculation results using exactly the same statistical measures.

Guillaume Godin

Residual connections are central to modern deep neural networks, enabling stable optimization and efficient information flow across depth. In this work, we propose SCORE (Skip-Connection ODE Recurrent Embedding), a discrete recurrent alternative to classical layer stacking. Instead of composing multiple independent layers, SCORE iteratively applies a single shared neural block using an ODE (Ordinary Differential Equation)-inspired contractive update: ht+1 = (1 - dt) * ht + dt * F(ht) This formulation can be interpreted as a depth-by-iteration refinement process, where the step size dt explicitly controls stability and update magnitude. Unlike continuous Neural ODE approaches, SCORE uses a fixed number of discrete iterations and standard backpropagation without requiring ODE solvers or adjoint methods. We evaluate SCORE across graph neural networks (ESOL molecular solubility), multilayer perceptrons, and Transformer-based language models (nanoGPT). Across architectures, SCORE generally improves convergence speed and often accelerates training. SCORE is reducing parameter count through shared weights. In practice, simple Euler integration provides the best trade-off between computational cost and performance, while higher-order integrators yield marginal gains at increased compute. These results suggest that controlled recurrent depth with contractive residual updates offers a lightweight and effective alternative to classical stacking in deep neural networks.

Pavel Karpov, Guillaume Godin, Igor V. Tetko

We present SMILES-embeddings derived from the internal encoder state of a Transformer [1] model trained to canonize SMILES as a Seq2Seq problem. Using a CharNN [2] architecture upon the embeddings results in higher quality interpretable QSAR/QSPR models on diverse benchmark datasets including regression and classification tasks. The proposed Transformer-CNN method uses SMILES augmentation for training and inference, and thus the prognosis is based on an internal consensus. That both the augmentation and transfer learning are based on embeddings allows the method to provide good results for small datasets. We discuss the reasons for such effectiveness and draft future directions for the development of the method. The source code and the embeddings needed to train a QSAR model are available on https://github.com/bigchem/transformer-cnn. The repository also has a standalone program for QSAR prognosis which calculates individual atoms contributions, thus interpreting the model's result. OCHEM [3] environment (https://ochem.eu) hosts the on-line implementation of the method proposed.

Guillaume Godin

Bond Centered FingerPrint (BCFP) are a complementary, bond-centric alternative to Extended-Connectivity Fingerprints (ECFP). We introduce a static BCFP that mirrors the bond-convolution used by directed message-passing GNNs like ChemProp, and evaluate it with a fast rapid Random Forest model on Brain-Blood Barrier Penetration (BBBP) classification task. Across stratified cross-validation, concatenating ECFP with BCFP consistently improves AUROC and AUPRC over either descriptor alone, as confirmed by Turkey HSD multiple-comparison analysis. Among radii, r = 1 performs best; r = 2 does not yield statistically separable gains under the same test. We further propose BCFP-Sort&Slice, a simple feature-combination scheme that preserves the out-of-vocabulary (OOV) count information native to ECFP count vectors while enabling compact unhashed concatenation of BCFP variants. We also outperform the MGTP prediction on our BBBP evaluation, using such composite new features bond and atom features. These results show that lightweight, bond-centered descriptors can complement atom-centered circular fingerprints and provide strong, fast baselines for BBBP prediction.

Guillaume Godin

Outliers introduce significant training challenges in neural networks by propagating erroneous gradients, which can degrade model performance and generalization. We propose the Z-Error Loss, a statistically principled approach that minimizes outlier influence during training by masking the contribution of data points identified as out-of-distribution within each batch. This method leverages batch-level statistics to automatically detect and exclude anomalous samples, allowing the model to focus its learning on the true underlying data structure. Our approach is robust, adaptive to data quality, and provides valuable diagnostics for data curation and cleaning.

Guillaume Godin

We introduce molFTP (molecular fragment-target prevalence), a compact representation that delivers strong predictive performance. To prevent feature leakage across cross-validation folds, we implement a dummy-masking procedure that removes information about fragments present in the held-out molecules. We further show that key leave-one-out (key-loo) closely approximates true molecule-level leave-one-out (LOO), with deviation below 8% on our datasets. This enables near full data training while preserving unbiased cross-validation estimates of model performance. Overall, molFTP provides a fast, leakage-resistant fragment-target prevalence vectorization with practical safeguards (dummy masking or key-LOO) that approximate LOO at a fraction of its cost.

Guillaume Godin

Injecting rule-based models like Random Forests into differentiable neural network frameworks remains an open challenge in machine learning. Recent advancements have demonstrated that pretrained models can generate efficient molecular embeddings. However, these approaches often require extensive pretraining and additional techniques, such as incorporating posterior probabilities, to boost performance. In our study, we propose a novel strategy that jointly trains a single Graph Transformer neural network on both sparse multitask molecular property experimental targets and synthetic targets derived from XGBoost models trained on Osmordred molecular descriptors. These synthetic tasks serve as independent auxiliary tasks. Our results show consistent and significant performance improvement across all 19 molecular property prediction tasks. For 16 out of 19 targets, the multitask Graph Transformer outperforms the XGBoost single-task learner. This demonstrates that synthetic task augmentation is an effective method for enhancing neural model performance in multitask molecular property prediction without the need for feature injection or pretraining.

Ruud van Deursen, Igor V. Tetko, Guillaume Godin

In the present paper we evaluated efficiency of the recent Transformer-CNN models to predict target properties based on the augmented stereochemical SMILES. We selected a well-known Cliff activity dataset as well as a Dipole moment dataset and compared the effect of three representations for R/S stereochemistry in SMILES. The considered representations were SMILES without stereochemistry (noChiSMI), classical relative stereochemistry encoding (RelChiSMI) and an alternative version with absolute stereochemistry encoding (AbsChiSMI). The inclusion of R/S in SMILES representation allowed simplify the assignment of the respective information based on SMILES representation, but did not always show advantages on regression or classification tasks. Interestingly, we did not see degradation of the performance of Transformer-CNN models when the stereochemical information was not present in SMILES. Moreover, these models showed higher or similar performance compared to descriptor-based models based on 3D structures. These observations are an important step in NLP modeling of 3D chemical tasks. An open challenge remains whether Transformer-CNN can efficiently embed 3D knowledge from SMILES input and whether a better representation could further increase the accuracy of this approach.

Igor V. Tetko, Pavel Karpov, Ruud Van Deursen et al.

We investigated the effect of different training scenarios on predicting the (retro)synthesis of chemical compounds using a text-like representation of chemical reactions (SMILES) and Natural Language Processing neural network Transformer architecture. We showed that data augmentation, which is a powerful method used in image processing, eliminated the effect of data memorization by neural networks, and improved their performance for the prediction of new sequences. This effect was observed when augmentation was used simultaneously for input and the target data simultaneously. The top-5 accuracy was 84.8% for the prediction of the largest fragment (thus identifying principal transformation for classical retro-synthesis) for the USPTO-50k test dataset and was achieved by a combination of SMILES augmentation and a beam search algorithm. The same approach provided significantly better results for the prediction of direct reactions from the single-step USPTO-MIT test set. Our model achieved 90.6% top-1 and 96.1% top-5 accuracy for its challenging mixed set and 97% top-5 accuracy for the USPTO-MIT separated set. It also significantly improved results for USPTO-full set single-step retrosynthesis for both top-1 and top-10 accuracies. The appearance frequency of the most abundantly generated SMILES was well correlated with the prediction outcome and can be used as a measure of the quality of reaction prediction.

Fabio Capela, Vincent Nouchi, Ruud Van Deursen et al.

Prediction of molecular properties, including physico-chemical properties, is a challenging task in chemistry. Herein we present a new state-of-the-art multitask prediction method based on existing graph neural network models. We have used different architectures for our models and the results clearly demonstrate that multitask learning can improve model performance. Additionally, a significant reduction of variance in the models has been observed. Most importantly, datasets with a small amount of data points reach better results without the need of augmentation.

Ruud van Deursen, Guillaume Godin

Graphs and networks are a key research tool for a variety of science fields, most notably chemistry, biology, engineering and social sciences. Modeling and generation of graphs with efficient sampling is a key challenge for graphs. In particular, the non-uniqueness, high dimensionality of the vertices and local dependencies of the edges may render the task challenging. We apply our recently introduced method, Generative Examination Networks (GENs) to create the first text-based generative graph models using one-line text formats as graph representation. In our GEN, a RNN-generative model for a one-line text format learns autonomously to predict the next available character. The training is stopped by an examination mechanism checking validating the percentage of valid graphs generated. We achieved moderate to high validity using dense g6 strings (random 67.8 +/- 0.6, canonical 99.1 +/- 0.2). Based on these results we have adapted the widely used SMILES representation for molecules to a new input format, which we call linear graph input (LGI). Apart from the benefits of a short compressible text-format, a major advantage include the possibility to randomize and augment the format. The generative models are evaluated for overall performance and for reconstruction of the property space. The results show that LGI strings are very well suited for machine-learning and that augmentation is essential for the performance of the model in terms of validity, uniqueness and novelty. Lastly, the format can address smaller and larger dataset of graphs and the format can be easily adapted to define another meaning of the characters used in the LGI-string and can address sparse graph problems in used in other fields of science.

Ruud van Deursen, Peter Ertl, Igor V. Tetko et al.

Recurrent neural networks have been widely used to generate millions of de novo molecules in a known chemical space. These deep generative models are typically setup with LSTM or GRU units and trained with canonical SMILEs. In this study, we introduce a new robust architecture, Generative Examination Networks GEN, based on bidirectional RNNs with concatenated sub-models to learn and generate molecular SMILES with a trained target space. GENs autonomously learn the target space in a few epochs while being subjected to an independent online examination mechanism to measure the quality of the generated set. Here we have used online statistical quality control (SQC) on the percentage of valid molecules SMILES as an examination measure to select the earliest available stable model weights. Very high levels of valid SMILES (95-98%) can be generated using multiple parallel encoding layers in combination with SMILES augmentation using unrestricted SMILES randomization. Our architecture combines an excellent novelty rate (85-90%) while generating SMILES with a strong conservation of the property space (95-99%). Our flexible examination mechanism is open to other quality criteria.

Talia B. Kimber, Sebastian Engelke, Igor V. Tetko et al.

In our study, we demonstrate the synergy effect between convolutional neural networks and the multiplicity of SMILES. The model we propose, the so-called Convolutional Neural Fingerprint (CNF) model, reaches the accuracy of traditional descriptors such as Dragon (Mauri et al. [22]), RDKit (Landrum [18]), CDK2 (Willighagen et al. [43]) and PyDescriptor (Masand and Rastija [20]). Moreover the CNF model generally performs better than highly fine-tuned traditional descriptors, especially on small data sets, which is of great interest for the chemical field where data sets are generally small due to experimental costs, the availability of molecules or accessibility to private databases. We evaluate the CNF model along with SMILES augmentation during both training and testing. To the best of our knowledge, this is the first time that such a methodology is presented. We show that using the multiplicity of SMILES during training acts as a regulariser and therefore avoids overfitting and can be seen as ensemble learning when considered for testing.