Synthetic Chemistry Innovations

The sobering truth about AI drug discovery that no one wants to discuss: most of the "novel" AI-generated molecules are just expensive rehashes of existing compounds. After analyzing 71 published case studies, Xie et al. (2025) uncovered a critical blind spot in our AI-driven drug discovery revolution. While we celebrate algorithms that can navigate 10^60 possible molecules, the reality is far more humbling and costly. 🔍 The data doesn't lie: - Only 42% of AI-generated hits achieve basic novelty thresholds - Ligand-based models fail 60% of the time, creating "me-too" compounds - A mere 8.4% represent truly groundbreaking scaffolds (Tc < 0.2) This isn't just an academic curiosity, it's a multi-million dollar problem. Teams are burning through synthesis budgets on molecules that patent lawyers will tear apart as "obvious variations." Worse yet, some AI systems are literally rediscovering their own companies' shelved failures. What is the path forward? The solution isn't abandoning AI, but wielding it strategically. Structure-based approaches outperform ligand-based by 4:1 in novelty. Smart teams are building multi-metric dashboards that go beyond simple similarity scores, integrating automated patent searches and scaffold analysis from day one. The most successful paradigm? Again, it is a human-AI collaboration where computational power generates thousands of ideas, automated filters remove the derivative junk, and expert chemists make the final strategic calls. Read my full analysis based on the recent studies where I break down the root causes of "molecular déjà vu" and share the pragmatic strategies that separate the winners from the hype-chasers in AI drug discovery. Disclaimer: The opinions and information presented in this article are my own or come from the reviewed papers and do not necessarily reflect the official position of my employer. #DrugDiscovery #ArtificialIntelligence #IntellectualProperty #ChemInformatics #DrugDesign #Pharma #deeplearning #ComputationalChemisty #GenerativeAI #GenerativeChemistry #CADD

Researchers have invented a never-before-seen material called “glaphene” that combines two opposite substances—graphene and glass—into a brand-new 2D material with exotic properties that nature doesn't make on its own. Graphene is a one-atom-thick sheet of carbon known for being super strong and electrically conductive, while silica glass is an insulator. Normally, scientists stack these materials like sheets of paper, but they don’t truly bond. This time, however, an international team led by Rice University figured out how to chemically fuse them, creating a real hybrid with new behavior. Instead of stacking, the researchers chemically bonded the materials so their electrons could interact directly. This changed the way electrons move and created unique vibrations and behaviors not found in either material alone. The process involved a custom-designed setup that carefully controlled oxygen levels to first grow graphene and then form a silica layer—all in a single reaction. The result: a material that acts like both a metal and an insulator, essentially forming a new type of semiconductor. This hybrid material could pave the way for future breakthroughs in quantum computing, 3D holograms, and ultra-advanced electronics. It’s a perfect example of how combining unlikely ideas can lead to discoveries that push the boundaries of science and technology into new territory.    PMID: 40434220

I'm very pleased to report two new and very different approaches to aza-S(VI) derivatives, both separately published in ChemistryEurope journal. With Dominique Deans and collaborators we report the catalytic generation of sulfonimidamides by SuFEx reaction of sulfonimidoyl fluorides with diverse functionalised amine derivatives. The reaction performed in DMSO was suitable for automation, with reaction set-up using a liquid handling robot (Opentrons Labworks Inc.). In addition we report switchable and entirely chemoselective SuFEx vs SNAr processes with 4-fluorophenylsulfonimidoyl fluorides, which enabled the generation of further complex derivatives. https://lnkd.in/ey4fsjZ8 With Tsz-Kan Ma, Peerawat Saejong, King Long Or, Callum Begg we report a copper mediated transient C-H functionalisation reaction to form sulfides which could be converted to cyclic sulfilimines and sulfoximines. Using a catalytic transient directing group strategy with benzylamine substrates gave ortho-sulfanylation, and so directly set up amine sulfides for oxidative cyclisation to form S=N bonds. Unusually sulfenamides were used as the source of sulfanyl radicals. https://lnkd.in/ea3A8__d

A bit quiet recently, but happy to share an update on recently published #compchem work from our team. Although it was not planned, I am very happy to see these papers appear back-to-back in Digital Discovery #RSCDigital. First up is work from Valdas Vitartas, Hanwen Zhang, and Veronika Juraskova on developing a data-efficient strategy for training reactive machine-learned interatomic potentials, now implemented in our code mlp-train. The approach combines automated active learning with well-tempered metadynamics to iteratively and selectively explore chemically relevant regions of configuration space. We illustrate its use across a range of reactions, including the calculation of free energy surfaces in explicit solvent. https://lnkd.in/e-ZAu6D6 Next is Sara Tanovic/ Ewa Wieczorek's work on #retrosynthesis prediction. Through different experiments, they explored how dataset size and diversity, measured through reaction templates, affect the performance of single-step retrosynthesis models. The results show that template diversity is key to a model’s accuracy, and although models can generalise to unseen products, even template-free models cannot generalise to new chemical disconnections. These findings emphasise a fundamental class-imbalance problem in retrosynthesis prediction and highlight the need for datasets that are both balanced and diverse. https://lnkd.in/exnhueFz Finally, in Chloe’s work on reaction rate prediction, we introduce a BERT-based model to predict experimental log k values for SN2 reactions. We compare it with existing models and evaluate both accuracy and interpretability, identifying the chemical features that drive the model’s predictive performance. https://lnkd.in/eUH6dCYx A recurring theme in these papers is the use /development of quite different ML frameworks to help address current challenges in chemistry, including solvent effects, kinetic predictions, and how best use available data. We appreciate the constructive feedback from reviewers on both the manuscripts and the associated code

🧬 Multi-instance machine learning (MIL) in chemoinformatics and bioinformatics🧪 I recently read a new paper, "Chemical complexity challenge: Is multi-instance machine learning a solution?" by Zankov et al. (doi 10.1002/wcms.1698), after seeing a post about it from Pat Walters. See below for my summary! 📚 The study explores the application of MIL algorithms for handling complex chemical and biological data. MIL is a learning framework that considers objects as sets of multiple alternative instances, called a "bag". The goal, in contrast to single-instance learning (SIL), is to predict a label for the bag rather than for a single instance. For example, a single chemical structure can be represented by an ensemble of multiple conformations to predict bioactivity. This leads to a better representation of the dynamic nature of chemicals in equilibrium. 💡 MIL algorithms can be categorised as instance-based or bag-based. Instance-based algorithms consider instances in a bag as a separate objects. Predictions are generated for each instance in the bag, and a predefined rule (an aggregation function) is applied to aggregate the instance-level predictions into a bag-level prediction. Bag-based algorithms consider the whole bag as a training object, leading to a single prediction for the bag as a whole, rather than for individual instances. 🔬The paper showcases various applications of MIL in chemoinformatics and bioinformatics. For example, MIL has been used to model bioactivity with conformation ensembles, improving prediction accuracy over SIL from 71% to 91% on a benchmarking dataset. It has also been applied to predict protein-protein interactions, taking into account different protein isoforms synthesised from the same gene arising from alternative splicing. 🎯 One interesting aspect of MIL is its ability to identify key instances, such as specific molecular forms or fragments, that are responsible for a particular property or function of a molecule (key instance detection or KID). This opens up new possibilities for understanding molecular mechanisms and designing targeted interventions, for example by identifying key domains in proteins responsible for biological functions. 📊 However, the authors note the need for benchmark datasets to validate MIL models, especially in the chemistry and biology domains. Current benchmark datasets are limited in size and scope, and KID in particular is challenging to evaluate. They anticipate the development of new datasets to stimulate progress in this field. 🌟 Overall, this paper sheds light on the potential of MIL in chemoinformatics and bioinformatics, offering new insights and avenues for research. Exciting times ahead for data-driven approaches in understanding complex chemical and biological systems! #MachineLearning #Chemoinformatics #Bioinformatics #MultiInstanceLearning #DataScience #ArtificialIntelligence Image: Graphical abstract from paper

Two recent studies have highlighted how induced proximity can be used bring proteins together to rewire biology and generate downstream effects that have therapeutic value. In two related studies, the Gray team developed bifunctional compounds linking the BRD4 binder JQ1 to the PARP1 ligand olaparib (DD-CIP1/2 compounds), while the Erb group pursued a similar strategy using rucaparib (PCIP-1 compound). An interesting observation in both studies is that the compounds function via an event driven mechanism (still to be delineated), as indicated by a disconnect between the target engagement and anti-proliferative potency. This anti-proliferative activity has been shown to translate to in vivo anti-tumour activity for DD-CIP2. Notably, mechanistically this differs conceptually from RIPTACs, where potency arises from the formation of a cooperative and stable ternary complex that sequesters an essential cellular protein. Interestingly, the data supports that PCIP-1 functions via PARP2, rather than PARP1, possibly due to preferential formation of a productive ternary complex. Both studies make for excellent reading, and exemplify event driven pharmacology achieved by induced proximity between two proteins, providing further evidence that this modality has the potential to leverage mechanisms that are difficult to predict and challenging to measure but may carry significant therapeutic value. #inducedproximity #innovation Figure adapted from papers. DD-CIP2 paper: https://lnkd.in/eSS9F7wK PCIP-1 pre-print: https://lnkd.in/e6bFnnUD RIPTAC paper: https://lnkd.in/eAf5zBTd  

📗 I came across The Logic of Chemical Optimization by Kombo & LaMarche (2025) recently, and it has some really interesting ideas on how we think about optimization in medicinal chemistry. They introduce something called retro-optimization analysis—basically working backward from a development candidate (DC) to earlier leads and hits to look at how the molecule evolved and whether a more efficient path could have been taken. Some of the ideas that stood out: 📚 Optimizons – substructures in a molecule that were the main focus of optimization—kind of like SAR building blocks. The paper tracks how these change over time to understand the trajectory of design decisions. 📚 Optimization signatures – a way to capture which parts of the molecule were modified (e.g., 1-0-1), helping to visualize how structure–activity relationships were explored across analogues. 📚 Network metrics (like betweenness centrality and edge count) are used to identify key compounds in the optimization network. Some “theoretical leads” (TLs) had stronger network properties and were discovered earlier than the actual leads (ALs), suggesting alternate paths that might have accelerated the process. A few broader takeaways: 💡 Optimization tends to follow an “outside-in” trajectory—starting with peripheral changes and moving toward the core. 💡 Molecular complexity, synthetic accessibility, and conformational properties (like radius of gyration) were tracked over time, giving a clearer picture of how compounds evolve not just in potency, but in developability. 💡 The idea of using retrospective data to inform prospective design really stood out—there’s potential to use these methods in real time during lead optimization to prioritize more efficient paths. If you're working in drug discovery/compound design, it's a solid read with ideas that could be useful both for looking back and planning ahead. 📖 Read the full paper here: https://lnkd.in/guKJyMrk #MedicinalChemistry #LeadOptimization #DrugDiscovery #Cheminformatics #StructureActivity

I am pleased to share that our latest work has just been published in JACS Au. In this study, we report a catalyst- and additive-free olefin 1,2-halo-hydroxylamination enabled by anomeric amides, providing direct access to multifunctional hydroxylamines (MFHAs) from both activated and unactivated alkenes. The transformation proceeds under mild conditions with excellent atom economy and broad functional-group tolerance. Mechanistic experiments, DFT calculations, and molecular dynamics simulations reveal a rare polar-to-radical crossover mechanism, in which initial heterolytic activation triggers a homolytic cascade leading to electrophilic nitrogen transfer. This mechanistic paradigm offers a fundamentally different entry point to alkene difunctionalization. The resulting MFHAs contain both electrophilic N–O and C–X bonds and serve as versatile intermediates for the synthesis of structurally complex amines and N-heterocycles, including aziridines, pyrrolidines, and tetrahydroquinolines. I would like to thank Dr. Young-Do Kwon (Young Do Kwon) and Dr. Daniel Joaquin (Daniel Joaquin) from the Kürti Group at Rice University (Rice Chemistry) for their outstanding experimental work, as well as Professor Daniel H. Ess (Daniel Ess) and his team (Michael Davenport and Jeff Olsen) at Brigham Young University (Brigham Young University) for the insightful computational studies that were essential for elucidating the reaction mechanism. Special thanks go to Professor Muhammed Yousufuddin (Muhammed Yousufuddin) and Melanie Carpio at University of North Texas at Dallas (University of North Texas at Dallas) for obtaining crystal structures for several key compounds. The article is fully open access: https://lnkd.in/gEGbedJt

SMILES isn't just a string. It's a molecular programming language. And like any language, fluency separates beginners from experts. After 19 years of cheminformatics, here's what I wish someone told me earlier: 𝗧𝗵𝗲 𝗛𝗶𝗱𝗱𝗲𝗻 𝗣𝗼𝘄𝗲𝗿 𝗼𝗳 𝗖𝗮𝗻𝗼𝗻𝗶𝗰𝗮𝗹 𝗦𝗠𝗜𝗟𝗘𝗦 Same molecule, infinite representations: CCO | OCC | C(O)C All ethanol. But only one canonical form. Why it matters: → Database deduplication fails without it → ML models see these as different molecules → Virtual screening misses hits due to representation mismatch One-line fix: Chem.MolToSmiles(mol, canonical=True) 𝗦𝗠𝗔𝗥𝗧𝗦: 𝗪𝗵𝗲𝗿𝗲 𝗥𝗲𝗮𝗹 𝗣𝗼𝘄𝗲𝗿 𝗟𝗶𝘃𝗲𝘀 SMILES describes molecules. SMARTS queries them. Patterns I use weekly: → [#7;H2] - Primary amines (metabolic hotspot) → CX3[OX1-] - Carboxylate anion → [#6;X4][F,Cl,Br,I] - Alkyl halide (reactive!) 𝗣𝗿𝗼 𝗠𝗼𝘃𝗲𝘀 𝗠𝗼𝘀𝘁 𝗠𝗶𝘀𝘀: 1️⃣ 𝗦𝘁𝗲𝗿𝗲𝗼𝗰𝗵𝗲𝗺𝗶𝘀𝘁𝗿𝘆 𝗲𝗻𝘂𝗺𝗲𝗿𝗮𝘁𝗶𝗼𝗻 : One SMILES → 2ⁿ stereoisomers. Your "hit" might be the wrong enantiomer. 2️⃣ 𝗧𝗮𝘂𝘁𝗼𝗺𝗲𝗿 𝗮𝘄𝗮𝗿𝗲𝗻𝗲𝘀𝘀 : Keto-enol forms are different SMILES but same compound. 3️⃣ 𝗣𝗔𝗜𝗡𝗦 𝗳𝗶𝗹𝘁𝗲𝗿𝗶𝗻𝗴 : 480 SMARTS patterns that flag promiscuous compounds. Use FilterCatalog before you waste months on false positives. 4️⃣ 𝗦𝗰𝗮𝗳𝗳𝗼𝗹𝗱 𝗱𝗲𝗰𝗼𝗺𝗽𝗼𝘀𝗶𝘁𝗶𝗼𝗻 : Murcko scaffolds reveal the core. Essential for series analysis and IP landscaping. 𝗧𝗵𝗲 𝗿𝗲𝗮𝗹 𝗲𝘅𝗽𝗲𝗿𝘁𝗶𝘀𝗲? Knowing WHEN to use SMILES vs InChI vs SELFIES vs molecular graphs. Each representation has failure modes. The best computational chemists understand the trade-offs. What's your most useful SMARTS pattern? I'll share my full SMARTS collection in future posts. #Cheminformatics #DrugDiscovery #Python #RDKit #SMILES #ComputationalChemistry #AI #MedicinalChemistry #AIDD #Molcura College of Pharmacy DMU Ajman University American University of Sharjah American University of Ras Al Khaimah Gulf Medical University