Small molecule drug discovery resources

Compendium of recent articles, resources, and blogs in the area of drug discovery and AI
chemical-science
machine-learning
resources
Published

January 10, 2024

Last update: 1st May 2024

Noteworthy blogs to follow:

Cheminformatics

  1. Patrick Walters Blog on Cheminformatics

  2. Is Life Worth Living

  3. Andrew White’s Deep Learning for Molecule and Material

  4. Cheminformia

  5. Depth-First

Fragment-based drug dicovery

  1. Practical Fragments

Medicinal chemistry

  1. Darryl B McConnell’s Medchem blogpost

General field

  1. DrugDiscovery.NET - Andreas Bender

  2. DrugHunter

  3. Derek Lowe’s In the Pipeline

Online resources

Books

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.

Pharma R&D Business

Overviews and Reviews

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.

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

Commentary

Correspondence on assessing the impact of AI on medicinal chemistry. It is a well written account on practical implication of generative design on pharmaceutical research.They outline two recent cases of ‘success’ of AI generative design in drug discovery and give more context and propose best practices for furthering the development of algorithms and drug discovery pipelines.

Very good post outlining the focus on the good practices and lack thereof for consistent datasets for comparing different ML algorithms.

Industry-focused drug discovery reviews

Latest review of the field and application of AI technologies to various functions of drug discovery. In addition to providing a quick review of the main technology the author list different case studies where AI has proposed clinical candidates across different therapeutic areas. Yet, the need for better data, novelty estimation, and validation in wet lab limit the full scale deployment and accuracy of AI pipelines in drug discovery. In closing they also hint at the limitation of ML model training a single property with single structure, QSAR, while in reality the molecule can exist in different conformers something multi-instance learning (MIL) can address. ‘It is reasonable to assume that user expertise, bias, and time constraints play a significant role in early drug discovery, often more so than AI.’

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

Special Journal Issues

  1. 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

Chemical modalities

Overview of different chemical modalities currently at work to address different disease targets. The article addresses the small molecule medicinal chemists and how they can expand their outlook of small molecules to include other molecular entities when considering the angle of attack for different target engagement strategies. The authors offer a nice set of tools and thought process when selecting possible drug modalities for different target classes and what questions should be asked when zeroing in a possible mode of action.

Meta themes on optimizing small molecules

Retrospective analysis on factors influencing the bioavailability of drug candidates. Authors find rotatable bonds and polar surface area or hydrogen bond count (sum of donor and accpetors) found to be important predictors of good oral bioavailability. Compounds having <10 rotatable bonds and <140 A (or < 12 hydrogen bonds) have good chances of being orally bioavailable.

AB-MPS calculated using cLogD, the number of aromatic rings (nAr), and the number of rotatable bonds (nRotB) according to the formula AB-MPS = Abs(cLogD −3) + nAr + nRotB. The lower the AB-MPS score, the more likely the compound is to be absorbed, and a value of ≤14 is reported to predict a higher probability of oral absorption.

Hueristics for oral bioavailability of molecules that are violating the rule of 5. MW may reach up to approximately 1000 Da provided that TPSA increases proportionally up to 250 Å2. In contrast, cLogP and HBDs must be carefully controlled at high MW. Our lack of ability to predict compound conformations and flexibility is currently a hurdle that is critical to overcome to enable further prospective design in oral bRo5 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.

Longitudinal analysis of physico-chemical properties for approved drugs in the clinic. They show that most of the drugs flout most of the Lipinski’s rule of 5 except the HBD which is always consistently less 4. In addition, they show that in recent times, by categorizing the drugs in different time-bound classes, the mean MW and HBA has increased but mean HBD has constantly stayed less than 2.

Synthesis Chemistry

Catalog of recent research articles that look at synthesis chemistry from a point of view of computational workflows, how traditional synthetic chemistry methods can be combined with informatics to augment drug discovery and synthesis processes.

Curated set of substrates to quickly assess the practicality of synthetic methods with the complete capture of success and failure, that can optimize reaction conditions with a broader scope with respect to relevant applications.

Large chemical libraries and Virtual Screening

Over the past few years several entites offering ultra-large ensembles of chemical libraries which can be made on-demand or purchased immediately have emerged. The existence of such services has reinvigorated the field of virtual screening and combinatorial library design. In addition, research groups have devised novel ways to navigate these libraries, more efficiently and also understand the differences in the chemical space these library cover. Following are some of the key papers in the field.

SpaceCompare: calculation of the overlap of large, nonenumerable combinatorial fragment spaces, utilizes topological fingerprints and the combinatorial character of these chemical spaces. Enamine’s REAL Space, WuXi’s GalaXi Space, and Otava’s CHEMriya. The overlap of the commercial make-on-demand catalogs is only in the low single-digit percent range, despite their large overall size.

PathFinder uses retrosynthetic analysis followed by combinatorial synthesis to generate novel compounds in synthetically accessible chemical space.

New version of ZINC with two major new features: billions of new molecules and new methods to search them. As a fully enumerated database, ZINC can be searched precisely using explicit atomic-level graph-based methods. Over 97% of the core Bemis–Murcko scaffolds in make-on-demand libraries are unavailable from “in-stock” collections. Correspondingly, the number of new Bemis–Murcko scaffolds is rising almost as a linear fraction of the elaborated molecules. Thus, an 88-fold increase in the number of molecules in the make-on-demand versus the in-stock sets is built upon a 16-fold increase in the number of Bemis–Murcko scaffolds. The make-on-demand library is also more structurally diverse than physical libraries

The authors examine the composition of the recently published and, so far, biggest chemical space, “eXplore”, which comprises approximately 2.8 trillion virtual product molecules. The utility of eXplore to retrieve interesting chemistry around approved drugs and common Bemis Murcko scaffolds has been assessed with several methods (FTrees, SpaceLight, SpaceMACS). Further, the overlap between several vendor chemical spaces and a physicochemical property distribution analysis has been performed. Despite the straightforward chemical reactions underlying its setup, eXplore is demonstrated to provide relevant and, most importantly, easily accessible molecules for drug discovery campaigns.

This paper proposes and studies Bloom filters for testing if a molecule is present in a set using either string or fingerprint representations. Bloom filters are small enough to hold billions of molecules in just a few GB of memory and check membership in sub-milliseconds. The authors found string representations can have a false positive rate below 1% and require significantly less storage than using fingerprints. Canonical SMILES with Bloom filters with the simple FNV hashing function provide fast and accurate membership tests with small memory requirements. They provide a general implementation and specific filters for detecting if a molecule is purchasable, patented, or a natural product according to existing databases at https://github.com/whitead/molbloom.

Virtual screeening

First paper to showcase deployment of quantum-based generative models with virtual screening workflow on a target-based compound discovery. Chemistry42 is used for reward function implementation. The authors show that molecule generated from this combined effort are ‘better’ quality-wise than traditional workflow (LSTM and Genetic algorithm) and the modeling success downstream is roughly correlated with number of qubits employed. This is exciting more from technical standpoint of combining quantum + traditional workflows.

Nice review on virtual screening workflow for streamlining drug discovery

The authors use a fragment screening approach to look at hits for protein kinase target and instead of using biophysical assay in fragment screening use crystallographic data directly to learn the conformation of the fragments. They find 4 ‘seed’ substructures which fit nicely in the protein(not affinity) and use those to inform the latter expansion which is done through the Enamine REAL dataset and known reaction classes. What I liked the most and found interesting is the high throughput binding pose and docking workflow of 200k compounds, the large scale crystallographic fragment hit analysis, and the focused curated library generation using Enamine REAL dataset. I was curious to know what seasoned experts had to say about this.

VirtualFlow as a tool for conducting virtual screening. The authors use VirtualFlow, to prepare one of the largest and freely available ready-to-dock ligand libraries, with more than 1.4 billion commercially available molecules. They screening ~1 billion compounds and identified a set of structurally diverse molecules that bind to KEAP1 with submicromolar affinity.

Researchers at UCSF looking at large scale docking for making ultra-large libraries accessible. They dock 170 million make-on-demand compounds from 130 well characterized reactions. Found new chemotypes that have interaction with 2 targets.

Protein-ligand interactions

Docking models

Conformer generators

Structure Quality Assessment

Python package for evaluating the quality of docked poses. PoseBusters performs a series of geometry checks on docked poses and also evaluates intra and inter-molecular interactions. The authors used the Astex Diverse Set and a newly developed PoseBusters benchmark set to evaluate five popular deep learning docking programs and two conventional docking approaches. The conventional docking programs dramatically outperformed the deep learning methods on both datasets. In most cases, more than half of the solutions generated by the DL docking programs failed the PoseBusters validity tests. In contrast, with the conventional docking programs, only 2-3% of the docked poses failed to validate.

Benchmarking Generated Poses: How Rational is Structure-based Drug Design with Generative Models

PoseCheck evaluates steric clashes, ligand strain energy, and intramolecular interactions to identify problematic structures. In addition, structures are redocked with AutoDock Vina to confirm the validity of the proposed binding mode. In evaluating several recently published generative models, the authors identify failure modes that will hopefully influence future work on structure-based generative design.

Binding free energetic calculations

Cheminformatics-focus

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:

Representation

Small molecules to be understood by computers and used for model training have to represented in a form amenable for optimization. In addition, this form of abstraction much capture appropriate level of chemical properties so as to imbue the data-driven models with necessary chemistry and physics for modeling. A lot of times different properties of the molecules are ‘lost in translation’ or obfuscated when converting them into machine-ready forms. Formerly the process of converting molecules from one form to another is called featurization. There are different forms, methods, theories to encode the molecules. Broadly there are as follows: * Fingerprints * Descriptors * Pharmacophores * Graph-based * Natural language-based * Shape-based

Reviews

Articles

The authors propose a new way to represent molecules, not as chemical bonds, but rather set representations. They show the set representation scheme can be alternative to SOTA graph-models and performs at par to the predictive tasks such as reaction yields and protein-ligand affinities.

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

*Reviews

Traditional ML uses the relationship between a single instance (a chemical structure) and a single label (a property). It doesn’t provide a facility for mapping multiple instances (an ensemble of conformers) to a label. There has recently been renewed interest in multiple instance learning (MIL), a technique developed over 30 years ago. MIL provides a framework that enables the mapping of conformational ensembles to properties. A recent review by Zankov from Hokkaido University and coworkers at other institutions provides an excellent overview of the challenges and opportunities associated with MIL in QSAR, genomics, and several other areas. The paper also provides links to several software packages for building MIL models.

Overview of the field and some factors that complicate current benchmarking efforts. The authors compared several molecular representations and ML algorithms in evaluating model accuracy and uncertainty. These evaluations highlighted the strengths of different QSAR modeling and ADME prediction methods. Consistent with other papers published in 2023, 2D descriptors performed best for ADME prediction, while Gaussian Process Regression with fingerprints was the method of choice when predicting biological activity.

A paper by Janela and Bajorath outlines several limitations in current benchmarking strategies. The authors used sound statistical methodologies to examine the impact of compound potency value distributions on performance metrics associated with regression models. They found that across several different ML algorithms, there was a consistent relationship between model performance and the activity range of the dataset. These findings enabled the authors to define bounds for prediction accuracy. The method used in this paper should be informative to those designing future benchmarks.

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

The paper discusses challenges in predicting thermodynamic solubility with machine learning. It reviews historical data, analyzes solubility datasets, and assesses model reliability. The authors propose a workflow for data curation and present new models and quality-assessed datasets for public use.

Deng and coworkers from Stony Brook University compared many popular ML algorithms and representations, curated new datasets, and performed statistical analysis on the results. This paper provides one of the best comparisons of ML methods published to date. The authors compare fixed representations, such as molecular fingerprints, with representations learned from SMILES strings and molecular graphs and conclude that, in most cases, the fixed representations provide the best performance. Another interesting aspect of this paper was an attempt to establish a relationship between dataset size and the performance of different molecular representations. While fixed representations performed well on smaller datasets, learned representations didn’t become competitive until between 6K and 100K datapoints were available.

A paper from Fang and coworkers at Biogen introduced several new ADME datasets. Unlike most literature benchmarks, which contain data collected from dozens of papers, these experiments were consistently performed by the same people in the same lab. The authors provided prospective comparisons of several widely used ML methods, including random forest, SVM, XGBoost, LightGBM, and message-passing neural networks (MPNNs) on several relevant endpoints, including aqueous solubility, metabolic stability, membrane permeability, and plasma protein binding.

The paper presents a novel deep learning approach to predict the developability of small molecules based on their predicted ADMET properties. The authors use a large-scale data set of compounds from the Novartis pipeline and train a neural network to rank compounds according to their potential to progress beyond in vivo PK studies. The resulting score, called bPK, outperforms other compound scoring methods and shows strong generalization ability on public data. The authors demonstrate the usefulness of bPK for series prioritization and optimization in drug discovery projects.

Account on how to treat and analyze activity cliffs in context of developing a predictive model. The authors outline best practices to probe activity cliffs. They show, using 24 DL and ML models and 30 targets, ML approaches based on molecular descriptors outperformed more complex deep learning methods. Activity cliff pairs were defined on similarity of the molecule SMILES and the bioactivity difference. Compared to most traditional machine learning approaches, deep neural networks seem to fall short at picking up subtle structural differences (and the corresponding property change) that give rise to activity cliffs.

Authors provide an evaluation of global and local models for ADME endpoint prediction. They compare the performance of global models and domain-specific local models. 10 different asays and 112 drug discovery projects were analyzed. The results showed consistent superior performance of global ADME models for property prediction. Performance improvement of global models over project-wise local models ranged from 3% to 25% in MAE. Local model improvements higher than 20% were achieved for only 7% of the assay-project pairs.

Datasets

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.

Matched molecular-pair

Enumeration of chemical space

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.

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.

Open-source tool to generate synthetically accessible chemical spaces using reaction definitions and building blocks. Virtual fragments are generated using one-step reaction and real-world building blocks - the workflow also support 2-3 steps creation.

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

Reviews/Perspectives

Articles

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.

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.

Comparisons

Traversing Chemical Space with Active Deep Learning

The authors compared six active learning approaches on three benchmark datasets and concluded that the acquisition function is critical to AL performance. When comparing molecular representations, they found that fingerprints generalized better than graph neural networks. Consistent with previous studies, they found that the choice of an initial training set had little impact on the outcome of an AL model.

Articles

Article talks about MolPAL as an active learning methodology. The team explores the application of these techniques to computational docking datasets and assess the impact of surrogate model architecture, acquisition function, and acquisition batch size on optimization performance. We observe significant reductions in computational costs; for example, using a directedmessage passing neural network we can identify 94.8% or 89.3% of the top-50 000 ligands in a 100M member library after testing only 2.4% of candidate ligands using an upper confidence bound or greedy acquisition strategy, respectively.

Article exploring different active learning strategies for looking at sampling the congeneric RBFE calculations. The paper explores the impact of several AL design choices. They show that in their case, the overall AL performance is largely insensitive to the specific ML method and acquisition functions used. The significant factor affecting the performance was the number of molecules sampled at each iteration.

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.

Multi-parameter optimization

Transfer Learning

Reviews

Articles

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 design

Reviews

Benchmarks

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.

REINVENT is a molecular design tool for de novo design, scaffold hopping, R-group replacement, linker design, molecule optimization, and other small molecule design tasks. At its heart, REINVENT uses a Reinforcement Learning (RL) algorithm to generate optimized molecules compliant with a user defined property profile defined as a multi-component score. Transfer Learning (TL) can be used to create or pre-train a model that generates molecules closer to a set of input molecules.

Language models:

Collaboration with Microsoft AI and Global Health Drug Discovery Institute. TamGen, a method that employs a GPT-like chemical language model and enables target-aware molecule generation and compound refinement.The authors identified 7 compounds showing compelling inhibitory activity against the Tuberculosis ClpP protease.

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

Reproducibility study of the STONED work from Jablonka et. al.

Representation using SELFIES proposed to make it much more powerful

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.

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

Graph-based

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’.

GANs

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)

Scaffold-retained

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

Reaction tranformation-based

Here the idea is to constraint the molecules generated by the transformations amenable to a particular platform, like automated synthesis workflow.

Authors propose a generative model to generate molecules via multi-step chemical reaction trees, each campaign first generates a reaction-tree with template transformations as breaking points.

Diffusion models

3D conformations-aware

Extension to the fragment-based generative design model (DeepFMPO) using reinforcement learning now incorporating 3D electrostatic similarity in the analysis. Ability to replace fragment with similar 3D shape and electrostatics. ESP_sim tutorial for comparison of electrostatic potential and molecule shape is used for this purpose. The authors find scaffold-hopping bioisoteres for CDK2.

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

Protein-ligand interactions aware

Linker design

Interesting work on designing linkers using conformation aware generative design algorithm. Think of it like fragment-growing.

Computer Aided Synthesis Planning (CASP)

Reviews:

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

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:

  • Validating exploration algorithms on a consistent set of reaction system.
  • Need to generate a comparative metric to benchmark different algorithms.
  • Considering effect of solvents and/or protein embeddings in the analysis

Best practices

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

Benchmarking

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:

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.

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

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

Conformational sampling and designing of chiral organic catalysts.

Retrosynthetic routes:

Transformer model evaluation for retrosynthesis from AZ folks. Template-free method.

Meta analysis on the best set of hyperparameters for retrosynthesis routines. Here the authors explore different parameters of the retrosynthesis workflow and their impact on the performance of the route scoping. They propose new set of parameters, other than the default, to assist in improving the odds of the software finding a route for diverse of set of molecules. First of its kind look into an approach to identify such a set.

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.

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.

Retro-BLEU, a statistical metric adapted from the well-established BLEU score in machine translation, to evaluate the plausibility of retrosynthesis routes based on reaction template sequences analysis. The authors use PaRoute to validate this approach.

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

Review / Perspectives

Authors discuss the design of reaction datasets in ways that are conducive to data-driven modeling, emphasizing the idea that training set diversity and model generalizability rely on the choice of molecular or reaction representation. They lay down the experimental constraints associated with generating common types of chemistry datasets and how these considerations should influence dataset design and model building.

Thoughts from industry practioners on how to label low/no yield reactions in electronic lab notebooks (eLNs). This is important when building ML model for reaction outcomes.

Articles

Latest from Abigail Doyle’s group where they use bandit optimization routine, related to Thompson sampling, to find reaction condition.

The authors propose a platform that is built for looking at photocatalytic N-heterocycle synthesis connected with HTE, automated purification, and physicochemical assays. Implement train-test split in 3 different strategies to minimize ligand overlap.

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).

Also called Kraken - a discovery platform covering monodentate organophosphorus(III) ligands providing comprehensive physicochemical descriptors based on representative conformer ensembles. Using quantum-mechanical methods, the authors calculated descriptors for 1558 ligands, including commercially available examples, and trained machine learning models to predict properties of over 300000 new ligands.

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.

Yield prediction

Generate catalysts

Use VAE and RNN to propose new catalyst for Suzuki cross-coupling reaction. The trained models are used to find catalyst’s binding energy and find high percentage of novel and valid designs.

Databases

Reaction sanitization

Reaction data extraction

Automated chemistry workflows

Reviews

This review article provides an in-depth analysis of the state-of-the-art in SDL technology, its applications across various scientific disciplines, and the potential implications for research, and industry. This review additionally provides an overview of the enabling technologies for SDLs, including their hardware, software, and integration with laboratory infrastructure. Most importantly, this review explores the diverse range of scientific domains where SDLs have made significant contributions, from drug discovery and materials science to genomics and chemistry.

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

Articles

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.

Propose a way to incoporate 3D-spatial information in the DEL read outs to denoise the data.

Set of informatic tools to look at BBs producitivity in DEL screens and guide designs for new DELs. Authors calculate joint probabilities of the BBs for its activity and find increasing binding metric for individual BBs also increases the overall binding energy. The authors then cluster these BBs using 2D and 3D tanimoto FPs (3D Tanimoto Combo) and HDBSCAN clustering. Good workflow for implementing 3D-based ROCs filtering.

Large Language Models (LLMs)

It’s a stretch to say that GPT-4 or any other LLM understands Chemistry.

At this point, LLMs seem to have two general use cases. First, summarization and information retrieval. LLMs can parse vast collections of text, which can be queried using natural language. These information retrieval capabilities have many applications, from writing computer code and collating clinical trial results to summarizing papers on a specific topic.

While there are still issues with LLMs hallucinating and providing incorrect information, tools and strategies are being developed to ensure the validity of LLM responses.

The other area where LLMs appear to be making inroads is workflow management or tools orchestration. Many activities in drug discovery, whether computational or experimental, require long sequences of steps, which can be tedious to orchestrate. These include asking questions about data, analyze results, do routine post processing for comparing with known state of the project.

While it is often possible to script the execution of these steps, scripting requires a detailed knowledge of each step. LLMs have the potential to simplify this process and carry out multi-step procedures given only a set of initial conditions and a final objective. While the amount of progress the field has made in a short time is impressive, I don’t see LLMs replacing scientists any time soon.

Previously the field has propose assistants for this job here which comprised of pre-scripted set of rules and processes. While tedious, they seem to add lot of value to project teams for quickly analyzing the SAR. The hope is LLMs might make the dream of all encompasing assistant a reality.

Total survey of the LLM landscape

Reviews

Agents

PaperQA, a Retrieval-Augmented Generation (RAG) agent for the scientific literature. PaperQA begins by constructing LLM search queries from a set of keywords. The results of these searches are aggregated into a vector database and combined with a pre-trained LLM to create a summary of the search results. In benchmark comparisons, the differences between answers provided by PaperQA and human evaluators were similar to differences between individual human evaluators. Encouragingly, unlike many other LLMs, PaperQA didn’t hallucinate citations.

ChemCrow provides software tools for performing domain-specific tasks, including web searches, file format conversions, and similarity searches. Compared with GPT-4, ChemCrow provided superior performance on tasks like synthetic route planning. The authors also point to potential misuse of LLMs and suggest mitigation strategies.

Coscientist, a set of LLMs for designing and executing organic syntheses. Coscientist consists of four components designed to search the web, write Python code, extract information from documentation, and program laboratory robotics. The authors test Coscientist using several open and closed-source LLMs and present examples of the system’s ability to plan and execute simple organic syntheses.

Predictive modeling

Authors show GPT3 based predictive models perform on-par with SOTA with lower data points. Caution is the models are purely text-based and extreme black box and sometimes, while trite, correlation doesnt mean causation might become important here. Finally the fine tuning doesnt do regression on the data in same sense as a linear regression or random forest would do.

Small molecule related tasks

BioT5+ incorporates several novel features: integration of IUPAC names for molecular understanding, inclusion of extensive bio-text and molecule data from sources like bioRxiv and PubChem, the multi-task in struction tuning for generality across tasks, and a novel numerical tokenization technique for improved processing of numerical data.

Protein design and mechanics

Mixture of LoRA Experts: Leverage the power of fine-tuned LoRA experts by employing a mixture of experts, or MoE technique.

Clinical text

Authors look at clinical summarization and implement quantitative assesments with synctactic, semantic, and conceptual NLP metrics. A clinical reader study with 10 physicians evaluated summary completeness, correctness and conciseness; in most cases, summaries from our best-adapted LLMs were deemed either equivalent (45%) or superior (36%) compared with summaries from medical experts. The research provides evidence of LLMs outperforming medical experts in clinical text summarization across multiple tasks. This suggests that integrating LLMs into clinical workflows could alleviate documentation burden, allowing clinicians to focus more on patient care.

Data curation

Hackathon to quickly fine-tune GPT to parse synthesis data and extract relevant chemistry-related information.

Material science

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. 

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. Medium article on explaining the MolPlotly. Link

Calculate similarities of shapes and electrostatic potentials between molecules. Pen has a nice blogpost on using to estimate electronic similarities of common bioisosteres. blog

Survey protein surfaces for binding hotspots can help to evaluate target tractability and guide exploration of potential ligand binding regions.

  • Morfeus Machine learning focused descriptors for small molecules with emphasis on 3D information.

  • MolPal

Active learning methodology for sampling the chemical space

Chemprop version that combines Jazzy (AZ’s workflow for predicting H-bond strength)

Datasets & Chemical libraries

Molecule datasets

  • PubChem: public sourced molecules

  • ChEMBL: bioactive molecules (most synthetic)

  • SUREChEMBL: small molecules appearing in Patents

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

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

  • GDB-11

  • Papyrus

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

  • Mcule: Used in DEL enumerations

  • DrugBank

Reaction Datasets

  • USPTO
  • Pistachio
  • Reaxys
  • Open Reaction Database

Commericial (building block) vendors

  • eMolecules building blocks

  • Enamine Fragments

  • WuXi GalaXi space

  • Otava’s CHEMriya

Helpful utilities: