Last update: 5th August 2022

Noteworthy blogs to follow:

  1. Patrick Walters Blog on Cheminformatics
  2. Is Life Worth Living

  3. Andrew White’s ML for Molecules and Materials Online Book

  4. Cheminformia

  5. Depth-First

  6. DrugDiscovery.NET - Andreas Bender

  7. DrugHunter - Dennis Hu

  8. Practical Fragments

Online resources


Best practices

Nice account outlining guidelines for evaluating different AI/ML methodologies in molecular science. They propose a checklist of tests and best practices to assess the practicality and importance of different methodologies thereby providing a framework on how to evaluate plethora of ML workflows being proposed in different areas of chemical science. The basis for not overlooking the older non-ML method when evaluating the ‘new’ learning-based method, emphasis on model interpretation to translate the corrleation to chemical causality and finally

Set of rules, considerations, and caveats to keep in mind when designing ML model for chemical science. The authors propose a checklist when evaluating ML models, while intuitive at first, when lot of the new ML papers are scanned through that lens, you can identify the shortcommings of the proposed model. This checklist is especially helpful for those entering just entering the field.


Pedagogical account of various machine learning techniques, models, representation schemes from perspective of synthetic chemistry. Covers different applications of machine learning in synthesis planning, property prediction, molecular design, and reactivity prediction

Paper outlining good practices for interpretating QSAR (Quantative Structure-Property Prediction) models. Good set of heuristics and comparison in the paper in terms of model interpretability. Create 6 synthetic datasets with varying complexity for QSAR tasks. The authors compare interpretability of graph-based methods to conventional QSAR methods. In regards to performance graph-based models show low interpretation compared to conventional QSAR method.

Recent review summarising the state of the molecular property prediction and structure generation research. In spite of exciting recent advances in the modeling efforts, there is a need to generate better (realistic) training data, assess model prediction confidence, and metrics to quantify molecular generation performance.

In-depth account of the machine learning and computational methods used in material science and small molecules. Nice introduction to the mathematics and theory behind first-principles based methods.

Review from Aspuru-Guzik and Allen’s group discussing how ML can be leveraged for various tasks in drug formulation tasks.

Industry-focused drug discovery reviews

Overview of methods and scope of computational methods used in the drug development process.

Special Journal Issues

  1. Nice collection of recent papers in Nature Communications on ML application and modeling

  2. Data Science Meets Chemistry

This issue includes contributions that demonstrate the profound impact data science techniques have had in chemistry including chemical and materials synthesis, catalyst and materials design, and overhauling the models used in traditional theoretical or computational chemistry.

  1. Journal of Medicinal Chemistry compendium of AI in Drug discovery issue

  2. Account of Chemical Research Special Issue on advances in data-driven chemistry research

  3. Special Issue on Reaction Informatics and Chemical Space, Journal of Chemical Information and Modeling (2022)

Meeting notes

Specific areas of interest

Catalog of recent reviews and manuscripts I have found useful when learning more about the state-of-the-art in Cheminformatics. I’ve tried to categorize them roughly based on their area of application:




Comparative study of descriptor-based and graph-based models using public data set. Used descriptor-based models (XGBoost, RF, SVM, using ECFP) and compared them to graph-based models (GCN, GAT, AttentiveFP, MPNN). They show descriptor-based models outperform the graph-based models in terms of prediction accuracy and computational efficiency with SVM having best predictions. Graph-based methods are good for multi-task learning.

Predictive modeling

Self-supervised learning using special type of GNN architecture (GeoGNN) that includes molecule geometric / spatial information. Geometry-enhanced molecular representation learning method (GEM). The model achieves SOTA performance on 14 of 15 public classification and regression datasets.

Benchmark property prediction models on 19 public and 16 proprietary industrial data sets spanning a wide variety of chemical end points. Introduce a modeling framework (Chemprop) that consistently matches or outperforms models using fixed molecular descriptors as well as previous graph neural architectures on both public and proprietary data sets.

Combine structure (Graph-networks) and descriptor based features (QM-derived) to predict activation energies (E2/SN2 barrier height prediction) and regioselectivity. Incorporating QM and structure leads to better overall accuracy and generalizability even in low data regions. Atom and bond level features derived using QM and used in the model generation with a smaller dataset.

QSAR benchmarks

Enumeration of chemical space

Looks at biosteric replacements for the phenyl rings in the lead optimization phase. Phenyl rings results in improve potency but have poor solubility and lipophilicitty. Find biosteres can be used to improve them.

Analyze the nature of rings which appear in bioactive compounds. Ring systems are systematically extracted from one billion molecules and are analyzed to discover a structure or correlation in the bioactivity and type of rings. No simple set of structural descriptors separating active and inactive rings could be identified, the separation is best described by a neural network model taking into account a complex combination of many substructure features.

Authors propose an algorithmic approach called as SpaceCompare to calculate overlap and diversity of the ultra-large combinatorial chemical libraries. The tool uses topological fragment spaces to capture the subtlties of the reaction having same product but different reactant substructures.

Organizing the chemical space of ChEMBL, and ZINC to compare its overlap with natural products through COCONUT. Generative Topological Mapping is used for the clustering and analysis. Helpful overview of the method with its application to drug discovery can be found here

Explainable/Interpretable Machine Learning



Uncertainty quantification

Benchmark different models and uncertainty metrics for molecular property prediction.

Train network to output the parameters of an evidential distribution. One forward-pass to find the uncertainty as opposed to dropout or ensemble - principled incorporation of uncertainties

Conduct a global multi-objective optimization with expected improvement criterion. Find transition metal complex redox couples for Redox flow batteries that address stability, solubility, and redox potential metric. Use distance of a point from a training data in latent space as a metric to quantify uncertainty.

Distance from available data in NN latent space is used as a variable for low-cost, quantitative uncertainty metric that works for both inorganic and organic chemistry. Introduce a technique to calibrate latent distances enabling conversion of distance-based metric to error estimates in units of predicted property

Active Learning

Active learning provides strategies for efficient screening of subsets of the library. In many cases, we can identify a large portion of the most promising molecules with a fraction of the compute cost.

Train property prediction model to output a distribution statistics in single pass that describes the uncertainty. This is in contrast to using ensemble models like MC dropout. Interesting way to estimate the epistemic (due to / from model) uncertainty in the prediction. Use this approach on antibiotic search problem of Stokes et. al. Compare Chemprop and SchNet models on different tasks.

Transfer Learning



Transfer learning by training a network to DFT data and then retrain on a dataset of gold standard QM calculations (CCSD(T)/CBS) that optimally spans chemical space. The resulting potential is broadly applicable to materials science, biology, and chemistry, and billions of times faster than CCSD(T)/CBS calculations.

Meta Learning

Authors demonstrate how one-shot learning can be used to signifinicantly lower the amount of data required to make predictions in drug discovery tasks. LSTM combined with GCNNs is shown to improve learning capabilities of the model. In the simplest one-shot learning formalism these continuous vectors are then fed into a simple nearest-neighbor classifier that labels new examples by distance-weighted combination of support set labels

Use CheMBL dataset to train a gated graph neural network (GGNN) for prediction and classification tasks using meta learning protocols. Show appreciable model performance even with just approx. 256 datapoints.

Federated Learning

Consortia comprising of leading resarch labs and companies working on decentralized datasets and predictive modeling of biochemical and cellular activity.

Generative models



Test SOTA language models and representation performance against graph-based methods (CGVAE, JTVAE) for ‘challenging’ generative modeling tasks - generate a molecule - property distribution as a function of synthetic feasiblity. Graph models faced chanllenge in generating large molcules (> 100 HAs). Selfies provided advantage here. All of the models seem to generate novel molecules - how practical each of these novel molecules are is yet an open question.

Propose a platform to deploy and compare state-of-the-art generative models for exploring molecular space on same dataset. In addition the authors also propose list of metrics to evaluate the quality and diversity of the generated structures.

Evaluation framework from BenevolentAI to compare different de-novo design models.

Interesting analysis from team at AstraZeneca R&D. They look at the chemical space coverage accounted by the SOTA generative models. Proposes a metric for evaluating space coverage, and thereby comparing different SOTA models, using a reference data (GDB-13 in this case). The new metric computes how much of the GDB-13 dataset can be recovered by a model that is trained on small GDB subset. Generative models were trained on same 1M data points and 1B molecules were then sampled from each model. It was seen that at most 39% of the molecules in the parent dataset were sampled / generated by the model. Most models sampled the same compounds atleast twice. It was observed that graph-based model sampled much diverse molecules than string-based methods. Besides, the coverage of GAN-based models was worse compared to Language and Graph models.

This paper looks at different ways of integrating synthesizability criteria into generative models.

Bechmark work from AstraZeneca/MIT AI team to document different graph architecture schemes and algorithms for generative models.

Language models:

One of the first implementation of a variation auto-encoder for molecule generation

Representation using SELFIES proposed to make it much more powerful

SMILES-based language model that generates molecules from scaffolds and can be trained from any arbitrary molecular set. Uses randomized SMILES to improve final prediction validity.


Reinforcement learning-based generative model whici is an update on point cloud approach by the same group to now incorporate ‘grammar’ for building molecules in form of functional groups in 3D space.

Junction tree based decoding. Define a grammar for the small molecule and find sub-units based on that grammar to construct a molecule. The molecule is generated in two-steps: first being generating the scaffold or backbone of the molelcule, then the nodes are added with molecular substructure as identified from the ‘molecular grammar’.

Introduce a graph generation model by building a Message Passing Neural Network (MPNNs) into the encoder and decoder of a VAE (MPGVAE).

Algorithm to predict 3D conforms from molecular graphs.

GraphINVENT uses a tiered deep neural network architecture to probabilistically generate new molecules a single bond at a time.


Generative adversarial network for finding small molecules using graph networks, quite interesting. Avoids issues arising from node ordering that are associated with likelihood based methods by using an adversarial loss instead (GAN)

Molecular generation strategy is described which combines an autoencoder and a GAN. Generator and discriminator network do not use SMILES strings as input, but instead n-dimensional vectors derived from the code-layer of an autoencoder trained as a SMILES heteroencoder that way syntax issues are expected to be addressed.


Team at Novartis and Microsoft propose MoLeR, graph based model to generate molecule using scaffold as a seed. Scaffold based SAR speed up shown.

Scoring functions

Method that combines GNNs with CNNs to incorporate 3D pharmacophoric constraints into molecular generation.

Extension to the fragment-based reinforcement learning methods for generating novel compounds. Comparison of 3D molecular fragments to aid in identifying bioactive conformations.

Using quantum chemistry attributes calculated on-the-fly as scoring functions for sampling the generative model chemical space. Active learning strategy is deployed to explore the area of space where the properties of the molecules are unknown.

Computer Aided Synthesis Planning (CASP)


Perspective on the current SOTA of synthesis planning, automation, and reaction optimization in drug discovery and development phases using AI and ML.

Perspective on ML for organic chemistry reactivity prediction. Group uses DFT-derived physical features of the reaction molecules and conditions for representation. Small data set plus HTE experimentation dataset for yield estimation.

Perspective article summarising their position on the current state of research and future considerations on developing better reaction network models. Break down the analysis of reaction networks as into 3 classes (1) Front Open End: exploration of products from reactants (2) Backward Open Start: Know the product and explore potential reactants (3) Start to End: Product and reactant known, explore the likely intermediates.

Nice summary of potential challenges in the field:

Technical details of various algorithms being implemented for reaction mechanism discovery at the time of writing the review.

Best practices

Article from Varnek group on best practices on processing data for reaction informatics.


Benchmarking framework for comparing different multi-step retrosynthesis methods from researchers at AstraZeneca R&D. Provides 10k synthetic routes which can be used as a validation set for different methodologies, providing a platform for systematic comparison of different methods being proposed in the community.

Classifying chemical reactions:

Using scrapped US Patent data to classify chemical reactions and deploy various fingerprints and ML models for classification.

Transformer-based model for reaction classification. Compared it with BERT. Besides classification, the work also formalizes the reaction fingerprint generation using the learned representations. The reaction fingerprints are visualized using TMAPS.

Reaction classifiction prediction using atom-mapped reaction that are used to generate condensed reaction graphs and passed through a GCN-variant as implemented in chemprop.

Atom mapping:

Comparative analysis of different atom-mapping schemes for generating atom-mapped reaction features. Comments on the state of the art methods and their performance on a curated reaction database.

Data-driven atom mapping schemes which uses transformers for learning the context of the chemical reaction. Researchers at IBM trained a flavor of language model based on Transformer architecture and used it to find reaction centers and maps atoms. Shown to be robust compared to other SOTA methods.

Predicting reaction outcomes:

Template-free prediction of organic reaction outcomes using graph convolutional neural networks

Retrosynthetic routes:

Interesting work on de-novo design of molecules wherein, the molecules being created are made up from the fragments that is known to exist and are available to the user. New molecules are generated based on the fragmented (synthons) made available in the dataset.

Hybrid neural-symbolic approach for both retrosynthesis and reaction prediction that can be trained with large reaction sets from databases. Template extraction from known reaction datasets to classify new reaction to known reaction classes.

In template-based retrosynthesis predictions, templates with few examples are excluded from training. This works talks on methods to augment the current set of data to account for the cases where examples for training are few.

Introduce a template-based single-step retrosynthesis model based on Modern Hopfield Networks, which learn an encoding of both molecules and reaction templates in order to predict the relevance of templates for a given molecule. The model does not consider templates as distinct categories, but can leverage structural information about the template. The retrieval approach enables generalization across templates, which makes zero-shot learning possible and improves few-shot learning. On the single-step retrosynthesis benchmark USPTO-50k, the MHN model reaction reaches the state-of-the-art at top-k accuracy for k ≥ 3.

Graph2SMILES, a template-free retrosynthesis model to predict reaction outcomes and retrosynthesis routes. This model eliminates the need for any input-side SMILES augmentation, while achieving noticeable improvements over Transformer baselines (especially for top-1 accuracy).

Generate reaction networks:

Newest version of RMG (v3) is updated to Python v3. It has ability to generate heterogeneous catalyst models, uncertainty analysis to conduct first order sensitivity analysis. RMG dataset for the thermochemical and kinetic parameters have been expanded.

Presents an algorithmic improvement to the reaction network prediction task through their YARP (Yet Another Reaction Program) methodology. Shown to reduce computational cost of optimization while improving the diversity of identified products and reaction pathways.

Look at exploration of reaction space rather than compound space. SOAP kernel for representing the moelcules. Estimate atomization energy for the molecules using ML. Calculate the d(AE) for different ML-estimated AEs. Reaction energies (RE) are estimated and uncertainty propogation is used to estimate the errors. Uncorrelated constant error propogation. 30,000 bond breaking reaction steps Rad-6-RE network used. RE prediction is not as good as AE.

Estimate molecular synthesizability

The idea of estimating whether a molecule is ‘synthesizable’ can be thought of from two areas:

  1. Complexity based - compare the fragments in the molecule to the known fragments in the chemical space
  2. Full retrosynthesis based - entire route is considered for molecule generation. Reactant complexity drives route complexity.

Synthetic Accessbility score (SA_Score) is a popular heuristic score for quantifying synthesizability. It computes a score using a fragment-contribution approach, where rarer fragments (as judged by their abundance in the PubChem database of 1mil representative cmpds) are taken as an indication of lower synthesizability.

SCScore is a learned synthetic complexity score computed as a neural network model trained on reaction data from the Reaxys database. It was designed with synthesis planning in mind to operate on molecules resembling not just drug-like products but intermediates and simpler building blocks as well.

RetroGNN is a graph neural network based model to predict outcome of a synthesis planner given the target molecule. Shown to better perform than SAScore. Code is yet to be released.

Data-driven chemistry modeling and reaction optimization



Experimental design using Bayesian Optimization. Look at 3 rxn class with multiple reaction parameters - temp solvent ligand. Algorithm identifies the optimal conditions. Variables looked into: ligands, bases, solvents, temperatures, concentrations. Algorithm arrived at 99% yields consistently - which was possible by using unusual ligand not known to work well (cognitive bias).

Multi-objective optimization of catalytic reactions that employ chiral bisphosphine ligands. Optimization of 2 sequential reactions in asymmetric synthesis of API. Classification method identify active catalysts – 5% yield (user provided) cutoff for binary classification. Linear regression to model reaction selectivity. DFT-derived descriptor dataset of >550 bisphosphine ligands. Develop an interpretable chemical space mapping tool using PCA. Look at the domain of applicability with the euclidean distance in chemical space.

Bayesian optimization (BO) to improve the experimental measured activity as a direct function of compositional variables without educating physical knowledge to the machine. We applied BO in screening spinel CraMnbFecCodNieCufZn3–a–b–c–d–e–fO4 for the decomposition of nitric oxide into environmentally friendly nitrogen.


Automated chemistry workflows

Account of Eli Lilly and Company’s ASL (Automated Synthesis Lab)

DNA-encoded Libraries

New form of storing huge amounts of molecule related data using DNA. Made partially possible by low cost of DNA sequencing. Each molecule in the storage is attached with a DNA strand which encode information about its recipe.

DNA encodings for discovery of novel small-molecule protein inhibitors. Outline a process for building a ML model using DEL. Compare graph convolutions to random forest for classification tasks with application to protein target binding. Graph models seemed to achieve high hit rate comapred to random forest. Apply diversity, logistical, structural filtering to search for novel candidates. First work to use GCN for hit searching.

Code / Packages:

Automates the selection of decision threshold for imbalanced classification task. The assumption for this method to work is the similar characteristics (like imbalance ratio) of training and test data.

Benchmarking platform to implement molecular generative models. It also provides a set of metrics to evaluate the quality and diversity of the generated molecules. A benchmark dataset (subset of ZINC) is provided for training the models.

Production-ready tool for de novo design from Astra Zeneca. It can be effectively applied on drug discovery projects that are striving to resolve either exploration or exploitation problems while navigating the chemical space. Language model with SMILE output and trained by “randomizing” the SMILES representation of the input data. Implement reinforcement-leraning for directing the model towards relevant area of interest.

DeepChem aims to provide a high quality open-source toolchain that democratizes the use of deep-learning in drug discovery, materials science, quantum chemistry, and biology - from Github

Github repository for implmenting message passing neural networks for molecular property prediction as described in the paper Analyzing Learned Molecular Representations for Property Prediction by Yang et. al.

“Chainer Chemistry is a deep learning framework (based on Chainer) with applications in Biology and Chemistry. It supports various state-of-the-art models (especially GCNN - Graph Convolutional Neural Network) for chemical property prediction” - from their Github repo introduction

Tool to generate chemical reaction networks. Includes Arkane, package for calculating thermodynamics from quantum mechanical calculations.

Active learning approach to efficiently and confidently identify the Pareto front with any regression model that can output a mean and a standard deviation.

Github repository to generate chemical reaction fingerprints from reaction SMILES.

Interactive chemical viewer for small molecules (RDKit wrapper)

Spotfire like capabilities to jupyter notebook.

Datasets & Chemical libraries

Molecule datasets

  • PubChem: public sourced molecules

  • ChEMBL: bioactive molecules (most synthetic)

  • ZINC: collection of synthetic molecules (not all are bioactive)

  • QM 7/8/9: small molecules having not more than 7/8/9 heavy atoms

  • Papyrus

  • COCONUT: NP 400k there are some which are not NP

  • Mcule: Used in DEL enumerations

  • DrugBank

  • QMugs

QMugs (Quantum mechanical properties of drug-like molecules) collection comprises quantum mechanical properties of more than 665 k biologically and pharmacologically relevant molecules extracted from the ChEMBL database, totaling 2M conformers.

Reaction Datasets

  • Pistachio
  • Reaxys
  • Open Reaction Database

Commericial (building block) vendors

  • eMolecules building blocks

  • Enamine REAL Space

  • WuXi GalaXi space

  • Otava’s CHEMriya

Helpful utilities: