The target at the center of the study is the mechanistic target of rapamycin, or mTOR, a serine/threonine kinase that functions as one of the cell’s most important signaling hubs. mTOR integrates signals about nutrient availability, growth factors, energy status, and stress, and it operates through two distinct multiprotein assemblies, mTORC1 and mTORC2. mTORC1 drives anabolic metabolism, protein and lipid synthesis, and the suppression of autophagy, while mTORC2 promotes cell survival, cytoskeletal organization, and Akt-dependent signaling. When this pathway is persistently or inappropriately activated, the delicate balance between growth, metabolism, and cellular quality control breaks down, a scenario closely linked to cancer as well as metabolic and neurodegenerative disorders. Existing drugs such as rapamycin and its analogs modulate the pathway, but their clinical usefulness is limited by incomplete pathway inhibition, feedback activation, and off-target effects, which is precisely why the search for next-generation mTOR inhibitors with better specificity remains so intense.

Instead of screening existing chemical libraries, the Indian team turned the problem around and asked a machine to invent molecules from scratch. The researchers used the DeepScreening web server, which employs a recurrent neural network, a class of deep learning model particularly adept at learning sequential patterns. Molecules are represented as SMILES strings, compact text encodings of chemical structures, and the network learns the statistical grammar of these strings well enough to generate chemically plausible new ones. The crucial step was transfer learning: the team took a pre-trained model and fine-tuned it on CHEMBL2842, a curated subset of the ChEMBL database containing 5,178 experimentally validated mTOR inhibitors with reported IC50 values. Because this training set spans diverse scaffolds and activity ranges, the fine-tuned generator absorbed the structural features associated with mTOR inhibition and produced a library of 200 candidate molecules, each scored by the model for target relevance.

With 200 synthetic candidates in hand, the researchers moved to structure-based validation. They retrieved the crystal structure of the human mTOR kinase domain bound to Torin2 from the Protein Data Bank, entry 4JSX, determined by X-ray diffraction at 3.50 angstroms. After stripping the co-crystallized ligand and preparing the receptor with polar hydrogens and Kollman charges, they first redocked Torin2 into the catalytic pocket to prove the protocol worked. The redocked inhibitor reproduced its crystallographic pose with a root mean square deviation of just 0.99 angstroms and a binding energy of minus 10.3 kilocalories per mole, anchored by a hydrogen bond to Val2240. Only then did they dock the generated molecules, using AutoDock Vina with an exhaustiveness setting of 8 and a grid centered on the Torin2-defined binding site.

Four compounds emerged from the docking campaign with scores equal to or better than the reference drug. S000031 led the pack at minus 12.2 kilocalories per mole, followed by S000046 at minus 10.9, while S000090 and S000043 matched Torin2 at minus 10.3. The interaction analysis revealed why S000031 stood out. It formed multiple hydrogen bonds with Val2240, Thr2245, and Glu2190, and it wrapped itself in an extensive hydrophobic network involving residues such as Cys2243, Lys2187, Met2345, Ile2356, Leu2185, Ala2248, Trp2239, and Gly2238. Notably, the key residues engaged by Torin2, particularly Val2240, Glu2190, Met2345, and Trp2239, were conserved across all four designed compounds, suggesting a shared binding mode. The authors emphasize that superior affinity came not from the sheer number of hydrogen bonds but from an optimal balance between polar anchoring and hydrophobic packing, a mechanistic insight that could guide future lead optimization.

Binding energy alone means little if a molecule cannot survive in a human body, so the team ran the four hits through a battery of pharmacological filters. Lipinski’s Rule of Five analysis using DruLiTo showed that S000046 and S000043 satisfied every criterion with zero violations, while S000031 and S000090 each showed a single violation, still within the accepted threshold for orally active drugs. Toxicity predictions from ProTox-3.0 were striking: all four designed compounds were predicted inactive for hepatotoxicity, cardiotoxicity, carcinogenicity, mutagenicity, and cytotoxicity, whereas the reference compound Torin2 showed predicted activity for hepatotoxicity, carcinogenicity, and mutagenicity. The authors are careful to note that these predictions come from a single in silico platform and must be confirmed experimentally before any claims about safety can be made.

Structural novelty was verified through Tanimoto similarity analysis using ECFP4 molecular fingerprints. All four leads scored far below the novelty threshold of 0.7 relative to Torin2, with S000046 at 0.0948 and S000031 at 0.096, meaning the neural network had not merely produced close analogs of known inhibitors but had explored genuinely new chemical space. ADMET profiling with SwissADME added further encouragement: S000031, S000046, and S000090 showed high predicted gastrointestinal absorption comparable to Torin2, none of the compounds were predicted to cross the blood-brain barrier, and none triggered PAINS alerts, the red flags that often disqualify screening hits. Synthetic accessibility scores ranged from 3.32 to 5.95, indicating moderate but feasible synthetic routes. One caveat surfaced in metabolism: S000090 and S000043 showed predicted inhibition of CYP3A4 and multiple CYP isoforms respectively, which may require optimization, while S000031 and S000046 showed minimal cytochrome P450 inhibition, suggesting favorable metabolic stability.

The final layer of analysis came from density functional theory, the quantum mechanical method that computes electronic structure from first principles. Using the ORCA software package at the B3LYP/6-311G(d,p) level of theory, the team optimized the geometries of the leads and Torin2 and examined their frontier molecular orbitals. Torin2 showed a HOMO-LUMO energy gap of 3.601 electron volts, while the designed compounds displayed much smaller gaps, with S000031 at just 0.44 electron volts. Derived descriptors told a consistent story: S000031 exhibited the lowest global hardness at 0.22 electron volts, the highest softness at 4.545 inverse electron volts, and a strikingly high electrophilicity index of 96.496 electron volts, indicating strong potential for interaction with nucleophilic sites in the protein. The authors stress that these quantum descriptors should be read as indicators of electronic character and chemical reactivity rather than direct evidence of biological potency.

The most promising candidate to emerge from the entire workflow is S000031, which combines the strongest predicted binding affinity, a chemically novel scaffold, clean toxicity predictions, favorable absorption, minimal CYP liability, and exceptional electronic reactivity. The study’s genuine contribution lies less in any single algorithm than in the integration: a target-focused generative model, validated docking, drug-likeness and ADMET screening, toxicity prediction, novelty assessment, and quantum chemical characterization assembled into one reproducible pipeline that could in principle be redirected at any kinase or disease target. The researchers themselves are candid about the limitations. Docking scores are approximate estimates prone to false positives, the DeepScreening server’s internal architecture could not be independently tuned or benchmarked, and no experimental assays have yet been performed on the new molecules. The immediate next steps they propose include molecular dynamics simulations, MM-PBSA and MM-GBSA binding free-energy calculations, and ultimately laboratory testing of the top-ranked compounds. Until those experiments are done, the four designed inhibitors remain computational hypotheses, but they are unusually well-documented ones, and they demonstrate that a neural network fine-tuned on validated data can propose molecules that a rigorous computational gauntlet judges to be at least as promising as the best known drug for one of cancer biology’s most important targets.

Subject of Research: Computational de novo design of mTOR kinase inhibitors using recurrent neural networks, molecular docking, ADMET prediction, and density functional theory

Article Title: Computational de novo discovery of novel mTOR inhibitors using recurrent neural networks and integrated in silico approaches including docking and DFT

Article References: Bisht, A., Kumar, S., & Chandra, S. (2026). Computational de novo discovery of novel mTOR inhibitors using recurrent neural networks and integrated in silico approaches including docking and DFT. Discover Chemistry, 3(1), Article 529. https://doi.org/10.1007/s44371-026-00994-x

Image Credits: AI Generated

DOI: 10.1007/s44371-026-00994-x

Keywords: mTOR, de novo drug design, recurrent neural network, molecular docking, AutoDock Vina, ADMET, density functional theory, cancer, kinase inhibitors, transfer learning, ChEMBL, drug discovery

The biological target at the heart of the study is a pair of enzymes known as cyclin-dependent kinases 4 and 6, or CDK4 and CDK6. These kinases act as gatekeepers of the cell division cycle, driving the transition from the G1 phase, when cells grow and prepare, into S phase, when DNA is replicated. When the CDK4/6 machinery is dysregulated, cells can escape normal growth controls and proliferate uncontrollably, a hallmark of many malignancies. The importance of these enzymes is underscored by the clinical success of approved CDK4/6 inhibitors such as palbociclib, ribociclib and abemaciclib, which have transformed the treatment landscape for hormone receptor-positive breast cancer and are being explored in other tumor types, including lung cancer.

Yet the existing drugs are not the end of the story. Resistance mechanisms emerge, side effects limit some patients, and clinicians and researchers continue to call for chemically diverse scaffolds, meaning molecules with fundamentally different structural backbones, that can potently inhibit both CDK4 and CDK6 at the same time. Finding such dual inhibitors by traditional high-throughput screening is expensive and slow, which is precisely why the research team turned to computer-aided drug discovery. Their strategy was to build a computational funnel: a wide opening that could swallow an entire chemical library and progressively narrower stages that would discard weak candidates until only the most promising molecules remained.

The first stage of the funnel was ligand-based machine learning. Rather than requiring knowledge of the target protein’s three-dimensional structure, ligand-based models learn from the known activities of compounds that have already been tested. The researchers assembled curated datasets from the ChEMBL database, a public repository of bioactivity data, gathering 265 compounds with measured activity against CDK4 and 402 against CDK6. Each molecule was converted into a numerical representation called an extended-connectivity fingerprint, specifically the ECFP4 format, which encodes the local atomic environments of a molecule in a way that machine learning algorithms can digest. Fingerprint-based descriptors of this kind have become a workhorse of computational chemistry because they capture subtle structural features that often correlate with biological activity.

With the data prepared, the team benchmarked multiple machine learning algorithms to see which could best predict inhibitory potency. The winner was a Bayesian Ridge regressor, a statistical learning method that combines ridge regression with Bayesian regularization, making it robust against overfitting on modestly sized datasets. Trained on the ECFP4 fingerprints, the model achieved cross-validated coefficients of determination, R-squared values, of 0.731 plus or minus 0.022 for CDK4 and 0.721 plus or minus 0.070 for CDK6. In practical terms, this means the model could explain roughly seventy percent of the variance in compound potency from molecular structure alone, a level of predictive accuracy strong enough to make meaningful prioritization decisions. The consistency of performance across both targets was crucial, because the goal was a dual inhibitor, and a model that excelled at one kinase but stumbled on the other would have undermined the entire approach.

The trained machine learning filter was then deployed against a library of 22,823 compounds. The Bayesian Ridge model scored every molecule for its likely activity against CDK4 and CDK6, allowing the researchers to discard the vast majority of the library computationally before any expensive calculation was run. This is where the economics of the approach become compelling: instead of assaying or docking tens of thousands of compounds, the team could concentrate structural modeling resources on the small fraction that the statistical model deemed most promising. The survivors of the machine learning filter then entered the second stage of the funnel, molecular docking against both CDK4 and CDK6.

Docking is a structure-based technique that uses three-dimensional models of the target proteins, typically derived from experimentally determined structures in the Protein Data Bank, to predict how a small molecule physically fits into the binding pocket of the enzyme. The docking calculations, performed at standard precision, evaluated the geometric complementarity and predicted binding energy of each candidate within the ATP-binding sites of both kinases. By requiring candidates to dock well against both CDK4 and CDK6, the researchers enforced dual-target affinity at the structural level, complementing the statistical predictions of the machine learning stage. This dual-target docking refinement whittled the prioritized list down to three candidate hits selected for biochemical evaluation in the laboratory.

The experimental results validated the computational strategy emphatically. Of the three candidates tested, HY-18,623 emerged as a potent dual inhibitor, with a half-maximal inhibitory concentration, IC50, of 3.5 nanomolar against CDK4 and 17.4 nanomolar against CDK6. Values in the low nanomolar range represent the kind of potency typically associated with advanced lead compounds, and achieving it against both kinases simultaneously with a single molecule is exactly what the field has been seeking. The result suggests that the computational funnel did not merely enrich for plausible binders but genuinely identified a molecule with drug-development potential.

To understand why HY-18,623 binds so tightly, the researchers turned to the third pillar of their workflow: molecular dynamics simulations. Docking provides a static snapshot of a predicted binding pose, but proteins and ligands in solution are constantly in motion, and a pose that looks perfect in a single frame may fall apart within nanoseconds. The team ran 200-nanosecond molecular dynamics simulations of the CDK4 and CDK6 complexes with HY-18,623, tracking the stability of the bound state over time using metrics such as root mean square deviation of the atomic positions. The simulations revealed that the compound maintains persistent hydrogen bonds with the so-called hinge residues of both kinases, Val96 in CDK4 and Val101 in CDK6. Hinge-region hydrogen bonding is a canonical anchor of kinase inhibitor binding, and its persistence throughout the simulations indicates a stable, high-affinity binding mode. Binding free energy analyses further quantified the favorable thermodynamics of the interaction, corroborating the experimental potency measurements.

Beyond the specific discovery of HY-18,623, the study offers a template for how modern drug discovery campaigns can be organized. The sequential integration of ligand-based machine learning, structure-based docking and dynamics-based validation creates a hierarchy of evidence, in which each stage interrogates the candidates from a different angle: statistical structure-activity relationships, physical fit within the binding site, and dynamic stability of the complex. The researchers have also made their data publicly available through a GitHub repository, supporting the open-science ethos that is increasingly shaping kinase drug discovery. For a field in which approved CDK4/6 inhibitors have already changed the course of breast cancer treatment, yet resistance and toxicity continue to motivate the search for alternatives, the arrival of a computationally discovered, nanomolar dual inhibitor with a structurally characterized binding mode is a development worth watching. HY-18,623 now stands as a promising lead candidate for further therapeutic development in oncology, and the workflow that produced it as a potentially reusable engine for discovering the next generation of kinase inhibitors.

Subject of Research: Computational virtual screening combining machine learning, docking and molecular dynamics to discover dual CDK4/6 inhibitors for cancer therapy

Article Title: Harnessing machine learning, docking and molecular dynamics for the virtual screening of compounds as CDK4/6 dual inhibitors

Article References: Wang, Y., Fang, L., Liu, Q., Zhang, Z., Xu, C., & Jiang, Y. (2026). Harnessing machine learning, docking and molecular dynamics for the virtual screening of compounds as CDK4/6 dual inhibitors. Molecular Diversity. https://doi.org/10.1007/s11030-026-11716-x

Image Credits: AI Generated

DOI: 10.1007/s11030-026-11716-x

Keywords: CDK4, CDK6, virtual screening, machine learning, molecular docking, molecular dynamics, drug discovery, kinase inhibitors, cancer, ChEMBL, ECFP4 fingerprints, HY-18,623

The mathematical backbone of the work is elegantly simple. Many celebrated indices in chemical graph theory, including the Zagreb indices, the Randić index, Sombor-type measures, the Albertson irregularity index, arithmetic-geometric and geometric-arithmetic indices, the Forgotten index and inverse Nirmala indices, can all be written as a sum over the edges of a molecular graph of a symmetric function of the degrees of the two atoms joined by each bond. The authors multiply each of these contributions by a normalisation factor derived from the general sum-connectivity index introduced by Zhou and Trinajstić: the weight one over the square root of the sum of the two endpoint degrees. Because this factor decreases as the local connectivity increases, it moderates the dominance of edges attached to high-degree atoms, a known source of instability when classical descriptors are applied across chemically diverse datasets.

The framework is not merely a numerical tweak. The authors prove a set of structural results that give the new family a rigorous footing. For any molecular graph whose minimum and maximum vertex degrees are delta and capital delta respectively, each sum-connectivity descriptor is provably sandwiched between its classical counterpart scaled by one over the square root of twice capital delta and the same counterpart scaled by one over the square root of twice delta. On regular graphs, where every atom has the same degree, the relationship collapses to exact proportionality with a constant of one over the square root of two r. The paper further derives explicit min-max bounds for representative descriptors in terms of the number of edges and the degree extremes, develops complement-graph identities showing how the reweighting behaves under the transformation that maps each degree to n minus one minus the original degree, and introduces a ‘hyper sum-connectivity’ index obtained when the construction is applied to the classical sum-connectivity index itself.

To test whether this theory translates into predictive power, the team assembled a carefully curated antibacterial dataset from the ChEMBL bioactivity database, restricted to minimum inhibitory concentration measurements against Escherichia coli, an organism that is both a common member of the human microbiome and a leading cause of urinary tract, bloodstream and intra-abdominal infections increasingly complicated by multidrug resistance. After standardising all MIC values to micromolar units, aggregating duplicates by median and expressing potency as the negative base-ten logarithm, the final dataset comprised 6,657 unique compounds, each characterised by eight physicochemical properties, two blocks of fourteen topological indices, the classical family and the new sum-connectivity family, and a block of ninety RDKit descriptors capturing size, branching, ring content and related structural features.

The first empirical test was a deliberately controlled one: a head-to-head comparison of the two index families alone across eight physicochemical endpoints, with everything else held constant. Under a demanding five-times-repeated five-fold cross-validation protocol, the sum-connectivity family produced higher mean coefficients of determination for all eight targets, lifting the average R-squared from 0.6951 for the classical indices to 0.7015 for the reweighted ones. The gains were most visible for topological polar surface area, lipophilicity and the spacial score, while molecular weight barely moved. The authors are careful to characterise this as a modest but systematic improvement, a claim backed by the theoretical guarantees rather than by an isolated lucky result.

A second, broader QSPR experiment concatenated everything: RDKit descriptors, both index families and the physicochemical properties, into a single ‘All Combined’ feature set. This configuration delivered the highest scores of the study, with R-squared values ranging from 0.9349 for NP-likeness to 0.9997 for molecular weight. But the team resists the temptation to oversell these numbers. Correlation analyses reveal that several endpoints are almost directly encoded by closely related features already present in the pool: molecular weight correlates with the RDKit exact molecular weight at a Spearman coefficient of 0.99999, and topological polar surface area tracks the nitrogen-oxygen count at 0.9304. The near-unity scores are therefore best read as upper-bound benchmarks within the available descriptor space, not as proof of independent mechanistic insight. A sensitivity rerun with principal component compression lowered scores across the board, showing that blind dimensionality reduction is no substitute for careful proxy-descriptor removal.

The harder test was antibacterial activity itself. Predicting the pMIC endpoint from structure alone is intrinsically difficult because potency reflects a tangle of scaffold class, charge distribution, lipophilicity, hydrogen-bonding profile and, often, a specific mechanism of action that no purely topological number can see. The best-performing QSAR model was an Extra Trees regressor trained on RDKit descriptors, reaching an R-squared of 0.5959 plus or minus 0.0246 under repeated cross-validation. Critically, the team then repeated the evaluation using a far stricter Murcko scaffold-based validation, which keeps close structural analogues out of the training folds and probes whether models generalise to genuinely new chemotypes. Performance dropped, as expected, to an R-squared of 0.4799 plus or minus 0.0252, but the ranking of feature families remained stable, indicating reproducible rather than chance-driven signal.

Within that stricter landscape, the new descriptors tell an honest story. On their own, the standalone sum-connectivity indices edged out the classical family for pMIC prediction, 0.2887 versus 0.2767 under random cross-validation and 0.2025 versus 0.2019 under scaffold validation, but both were clearly weaker than the rich RDKit representation, and the sum-connectivity advantage effectively disappeared once RDKit features were present. Diagnostics on the best model reinforced the picture of genuine but moderate predictability: a Y-randomization test with fifty permutations produced a mean R-squared of minus 0.0356, separated from the real model by a Z-score of 91.8, while permutation importance highlighted heterocycle counts, aliphatic ring content, charge-sensitive surface-area terms and electronic-state measures, a chemically plausible fingerprint of the determinants of antibacterial potency. A learning curve still rising with training-set size suggests further curated data would continue to help.

The study’s conclusions are notably measured for a field where bold claims are common. The authors state plainly that their descriptors are most strongly supported as systematically improved graph-topological descriptors in matched comparisons, with added value for the heterogeneous pMIC endpoint that is complementary and modest. They also acknowledge that every classical index is almost collinear with its sum-connectivity counterpart, with Spearman correlations between 0.9967 and 0.9986, and that a PCA rerun reduced the strongest benchmark scores while leaving constitution-driven targets comparatively easy to encode, so targeted removal of proxy descriptors and regularised feature selection remain the more promising next steps.

For the wider cheminformatics community, the work offers a template as much as a tool. It shows how a single, mathematically transparent transformation can generate a coherent descriptor family with provable bounds, how controlled matched benchmarks should be separated from headline-grabbing combined-model scores, and how scaffold-based validation should be the default standard for activity modelling on real drug-discovery data. The mathematical side also leaves a clear opening: extending the reweighting to the general chi-alpha family and studying how the exponent on the degree sum affects extremal behaviour and descriptor discrimination is flagged as an open direction. In an era when machine learning models are only as trustworthy as the features fed into them, a rigorous attempt to make one small corner of the descriptor space measurably better is a quietly consequential contribution.

Subject of Research: Sum-connectivity molecular descriptors for QSPR and antibacterial QSAR modelling

Article Title: Advancing QSPR with sum-connectivity descriptors: physicochemical and antibacterial modelling via molecular graph connectivity

Article References: Altairi, A., Alsharafi, M., & Alhaj, Z. (2026). Advancing QSPR with sum-connectivity descriptors: physicochemical and antibacterial modelling via molecular graph connectivity. Journal of Saudi Chemical Society, 30(4), Article 50. https://doi.org/10.1007/s44442-026-00105-6

Image Credits: AI Generated

DOI: 10.1007/s44442-026-00105-6

Keywords: QSPR, QSAR, topological indices, sum-connectivity descriptors, molecular graphs, antibacterial activity, Escherichia coli, ChEMBL, machine learning, MIC prediction, chemical graph theory, RDKit

In a study published in Molecular Diversity, researchers led by Huang Zeng and Bo Qiu of Jiaying University describe a systematic computational framework that pairs five types of molecular fingerprints with both classical machine learning and deep learning algorithms. Their goal was twofold: to rigorously compare how different combinations of chemical representations and algorithms perform at predicting TOP1 inhibition, and to convert the best performers into practical tools that any medicinal chemist can use. The work culminated not just in models but in validated hit compounds, two of which showed meaningful inhibition of the enzyme in laboratory tests.

The foundation of the study is data. The team mined the ChEMBL database for compounds with experimentally measured TOP1 inhibitory activity, assembling a dataset suitable for training classification models. Each molecule was then encoded using five distinct fingerprint families: AtomPairs, which describe pairs of atoms and their topological separation; MACCS keys, a set of 166 structurally predefined substructure features; Morgan fingerprints, circular fingerprints that capture local atomic environments much like extended-connectivity fingerprints; PharmacoPFP, which encodes pharmacophore features rather than pure connectivity; and RDKit descriptors, a collection of calculated physicochemical and topological properties. Because each fingerprint represents a different view of what a molecule is, the choice of representation can matter as much as the choice of algorithm.

On top of these representations, the researchers built a total of forty models spanning classical methods such as support vector machines, random forests, and k-nearest neighbors, alongside deep neural networks. The head-to-head comparison produced a clear result: combinations built on Morgan fingerprints dominated the leaderboard. Under random data splitting, the four top performers were SVM paired with Morgan fingerprints, random forest with Morgan, a deep neural network using MACCS keys, and k-nearest neighbors with Morgan. All four achieved area under the receiver operating characteristic curve values, or ROC-AUC, between 0.93 and 0.94, a level of discrimination that places them firmly in the range useful for prioritizing compounds in virtual screening campaigns.

High scores alone can deceive, particularly in cheminformatics where models sometimes memorize dataset quirks rather than learning genuine structure-activity relationships. To guard against this, the team applied Y-scrambling, a validation technique in which the activity labels are randomly shuffled and models are retrained. If the original models had been learning noise, their scrambled counterparts would perform equally well. They did not, supporting the conclusion that the top models had captured real, non-random relationships between molecular structure and TOP1 inhibition. The authors are also candid about a second, more sobering test: when compounds were split by molecular scaffold rather than randomly, AUC values dropped to between 0.67 and 0.82. This scaffold-split performance reveals that the models are less reliable when asked to extrapolate to chemically novel frameworks, a well-known limitation of fingerprint-based approaches that the researchers explicitly quantify rather than hide.

Interpretability was another pillar of the work. Black-box predictions carry little weight in drug design unless chemists can see what drives them. Using SHAP analysis, a game-theoretic method that assigns each molecular feature a contribution to individual predictions, the team identified the key structural elements that the models associated with TOP1 inhibition. This step transforms the models from opaque ranking machines into sources of chemical hypotheses, allowing medicinal chemists to see which substructures and descriptor values push a compound toward or away from the predicted active class, and potentially to use those insights when designing next-generation analogs.

What distinguishes this study from many computational papers is its insistence on usability. The researchers deployed their best models as a freely accessible web application, reachable at drugpred.top:5000, where users can input a molecule as a SMILES string, the standard text notation for chemical structures, or simply draw the structure in a browser and receive a prediction of TOP1 inhibitory activity. For laboratories that cannot or prefer not to rely on an internet connection, standalone desktop executables are available for download on GitHub at github.com/zenghuang8006/TOP1-inhibitor-prediction. The authors note that earlier AI-driven studies of TOP1 inhibitors lacked both systematic fingerprint and algorithm comparisons and accessible predictive tools, and this pipeline is designed to fill both gaps at once.

The real proof, however, came from the laboratory. The team screened a commercial library of 189,554 compounds from SPECS using their computational framework, then took the highest-ranking candidates into in vitro testing. Two compounds, designated AG60 and AI61, emerged as potential TOP1 inhibitor hits. At a concentration of 400 micromolar, AG60 inhibited the enzyme by 64 percent, while AI61 achieved 90 percent inhibition. Converting a computational prediction into experimentally confirmed enzyme inhibition, even at moderate potency, is a meaningful step; the compounds now represent starting points, or hits, that can be chemically optimized into more potent and selective leads through medicinal chemistry.

The biological context underscores why this target continues to attract attention. TOP1 is essential in proliferating cells, and its inhibition selectively harms cancer cells that cannot tolerate the resulting DNA damage. Approved TOP1 poisons remain mainstays of treatment for colorectal, ovarian, and other cancers, and newer agents such as antibody-drug conjugates carrying TOP1 inhibitor payloads have expanded the target’s therapeutic reach. But resistance, toxicity, and the pharmacological limitations of existing scaffolds keep the demand for chemical novelty high. The authors’ own prior work, including the computational identification of FTY720 and COH29 as TOP1 catalytic inhibitors, illustrates a sustained research program aimed at expanding the toolbox of TOP1-directed chemistry.

For the broader field, the study offers a template as much as a set of hits. It demonstrates that fingerprint choice can be systematically benchmarked, that interpretability tools like SHAP can be integrated into the screening workflow, and that honest reporting of scaffold-split performance should accompany headline accuracy figures. It also shows that the distance between a machine learning model and a wet-lab result can be shortened when prediction tools are built for practicing chemists rather than for other data scientists. The forty models, the validated hits AG60 and AI61, and the freely available web and desktop predictors together form a practical framework that other target-focused screening campaigns can adapt. As artificial intelligence continues to compress the early stages of drug discovery, studies that pair rigorous benchmarking with experimental follow-through, and that are transparent about both strengths and limitations, offer the clearest picture of what these methods can actually deliver for anticancer drug development.

Subject of Research: AI-driven screening of topoisomerase I inhibitors using machine learning, deep learning and molecular fingerprints

Article Title: Integrating machine learning and deep learning with multiple molecular fingerprints for topoisomerase I inhibitor screening and lead identification

Article References: Integrating machine learning and deep learning with multiple molecular fingerprints for topoisomerase I inhibitor screening and lead identification. (n.d.). https://doi.org/10.1007/s11030-026-11709-w

Image Credits: AI Generated

DOI: 10.1007/s11030-026-11709-w

Keywords: topoisomerase I, TOP1 inhibitors, machine learning, deep learning, molecular fingerprints, virtual screening, drug discovery, SHAP, ChEMBL, anticancer targets, SPECS compounds, QSAR

A new open-access study published in BMC Bioinformatics tackles this long-standing aggregation problem head-on. The research, authored by Mohammed Alsharafi, Azzam Altairi, Zaied Alhaj, and Yusuf Zeren, introduces and rigorously evaluates a fourteen-member family of product-connectivity descriptors, abbreviated PCI, designed to complement rather than replace the classical unweighted degree-based indices. The central idea is a product normalization applied to an existing degree kernel. Instead of summing edge contributions, the product formulation transforms how edge degrees combine, dampening the disproportionate influence of highly connected substructures and restoring sensitivity to the broader connectivity pattern of the molecule. The authors emphasize that the novelty lies not in proposing yet another fixed topological index, but in offering a general product-normalization scheme that can be grafted onto an existing degree kernel, making the approach modular and broadly applicable.

The theoretical half of the paper is devoted to establishing the mathematical credentials of the new family. The authors derive formal relations between each product-connectivity index and its corresponding unweighted classical counterpart, and they establish degree-extreme bounds—analytical limits that describe how the indices behave for graphs with extreme degree configurations. Such bounds matter in chemoinformatics because they guarantee that a descriptor is well-behaved across the full space of possible molecular graphs, not merely convenient on the handful of molecules in a training set. By proving these relationships, the team provides a principled foundation for the empirical work that follows, ensuring that the new descriptors are mathematically coherent and their behavior under degenerate conditions is understood in advance.

The empirical half of the study is unusually comprehensive by the standards of descriptor-evaluation papers. The authors assembled a curated dataset of 10,558 antiviral records drawn from ChEMBL, the public database of bioactivity data for drug-like molecules, covering 10,486 unique compounds. This scale is significant: many descriptor studies are validated on datasets of a few hundred molecules, which makes it difficult to distinguish genuine predictive value from statistical noise. The breadth of the ChEMBL antiviral collection also means the evaluation spans many viral targets and assay types rather than a single narrow endpoint, giving a more realistic picture of how the descriptors would perform in a working drug-discovery pipeline.

On the quantitative structure–property side, the study benchmarks the descriptors against twelve distinct QSPR endpoints, testing whether the product-connectivity family improves the prediction of physicochemical and related molecular properties. On the activity side, the authors tackle antiviral pEC50 prediction, where pEC50 is the negative base-10 logarithm of the half-maximal effective concentration, a standard measure of compound potency against viral targets. The modeling setup was deliberately controlled. For the QSPR tasks, the team used hold-out testing, reserving a portion of the data to evaluate models that had never seen it during training. For the QSAR task, they employed five-fold cross-validation, rotating the data so that every compound is tested exactly once. Both protocols are standard safeguards against overoptimistic performance estimates.

Equally important is the feature-block ablation design, which the authors describe as leakage-controlled. In descriptor studies, a common pitfall is information leakage, where features that implicitly encode the answer—such as experimental values or near-duplicate structures shared between training and test sets—inflate apparent accuracy. By organizing the feature space into blocks, including a block of RDKit descriptors, a block of physicochemical properties, and the new product-connectivity indices, and then systematically ablating, or removing, each block while controlling for leakage, the researchers could isolate the marginal contribution of the PCI family. This design directly addresses the question a skeptical reader would ask: do the new descriptors add anything beyond what established descriptor suites already capture?

The results, as reported in the study, show that the observed gains from the product-connectivity descriptors are dataset-dependent. The authors are explicit on this point, cautioning that the improvements should not be interpreted as universal superiority over richer descriptor spaces. This honesty is notable in a field where new descriptors are sometimes promoted with sweeping claims. What the study does establish is that the product-normalization strategy is a viable, theoretically grounded way to mitigate the dominance of highly connected substructures in additive degree-based indices, and that in appropriate datasets—particularly the antiviral QSAR setting that motivated the work—the family can measurably improve predictive performance within a fixed five-regressor modeling suite.

The choice to hold the regressor suite fixed across all experiments is itself methodologically meaningful. If the modeling algorithm changed between baseline and augmented feature sets, any performance difference could be attributed to the algorithm rather than the descriptors. By keeping five regressors constant and varying only the feature blocks, the authors ensure that observed differences trace back to the information content of the descriptors themselves. This controlled comparison, combined with the large curated dataset and the dual QSAR/QSPR evaluation, makes the study a useful template for how descriptor families should be validated going forward: theory first, then large-scale empirical benchmarking with leakage controls and ablations.

For the antiviral research community, the practical implications are straightforward. Antiviral drug discovery remains a pressing global need, and computational pre-screening that can reliably rank candidate compounds by predicted potency saves both time and laboratory resources. Degree-based topological indices are attractive precisely because they can be computed for millions of candidate molecules in seconds, and a normalization strategy that improves their signal quality without adding significant computational burden could sharpen the first filter through which candidate antivirals pass. The modular nature of the product-normalization approach means that practitioners can apply it to degree kernels they already use, rather than adopting an entirely new descriptor vocabulary.

The study also used only public chemical and bioactivity records and involved no human participants, human data, human tissue, animals, or animal tissue, and the article is published open access under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International license, permitting non-commercial sharing and distribution with appropriate credit. As the field of machine-learning-guided drug discovery continues to expand, work of this kind—careful, theoretically grounded, and empirically honest about its limits—plays a quiet but essential role. New descriptors rarely make headlines on their own, but the cumulative effect of better-constructed, better-validated molecular representations is a more reliable computational foundation for the search for next-generation antiviral therapies.

Subject of Research: Product-connectivity topological descriptors for improving antiviral QSAR and QSPR prediction

Article Title: A theoretical and empirical evaluation of product connectivity descriptors for improving antiviral QSAR and QSPR prediction

Article References: A theoretical and empirical evaluation of product connectivity descriptors for improving antiviral QSAR and QSPR prediction. (n.d.). https://doi.org/10.1186/s12859-026-06654-2

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06654-2

Keywords: topological indices, molecular descriptors, QSAR, QSPR, antiviral drug discovery, ChEMBL, chemoinformatics, machine learning, product connectivity, pEC50 prediction, BMC Bioinformatics, feature ablation

Small molecules face a formidable structural challenge, however. The PD-1/PD-L1 interface is a broad, shallow protein-protein contact surface, long considered a difficult target for drug-like compounds. Progress has come from chemotypes such as the biphenyl-based inhibitors developed by Bristol-Myers Squibb, including BMS-202, which bind PD-L1 and induce dimerization that blocks PD-1 recognition. Compounds such as INCB086550 and the oral modulator CA-170 have entered early clinical evaluation, proving that small-molecule checkpoint inhibition is feasible. Still, only a small fraction of explored scaffolds has advanced to the clinic, and screening the relevant chemical space experimentally is costly and slow. Computational triage has therefore become essential for deciding which candidates deserve synthesis and bioassay follow-up.

A team led by Monsin Sangsawat and Pornchai Rojsitthisak at Chulalongkorn University, with collaborators including Worathat Thitikornpong, Cong Feng, and Ming Chen, has now published an integrated solution in the journal Results in Chemistry. Their study reports an open-source, end-to-end cheminformatics framework for predicting the potency of small-molecule PD-L1 inhibitors, built entirely in the KNIME Analytics Platform. Every processing step, from data retrieval to model deployment, is encoded as an explicit node sequence, and the executable workflow file, curated datasets, and fixed split identifiers have been deposited in a public GitHub repository. The design directly confronts a persistent problem in the field: many published machine-learning models for PD-L1 are difficult to reproduce because training data, preprocessing details, or workflow code are inaccessible, and few are evaluated across independent data sources.

The foundation of the framework is a rigorously curated dataset. The researchers queried ChEMBL, targeting the human PD-L1 entry CHEMBL4523993, and retrieved more than 15,000 bioactivity records. To reduce label noise, they applied assay-level filters restricting the modeling set to homogeneous Alpha and HTRF assay formats, the two readouts with the most consistent annotation and the broadest activity coverage. After removing ambiguous qualifiers, duplicates, and non-numeric annotations, and after converting IC50 values to logarithmic pIC50 units, 2413 unique compounds remained. These were stratified into three potency tiers: active compounds with IC50 below 10 nanomolar, moderate inhibitors between 10 and 100 nanomolar, and weak or inactive compounds above 100 nanomolar. The curated set spans 953 unique Bemis-Murcko scaffolds, preserving the chemical diversity reported for PD-L1 inhibitors, from biphenyls and combretastatin analogs to cyclic peptides.

Molecular representation was treated as a central experimental variable. Each compound was encoded with 119 RDKit physicochemical descriptors, which were filtered for zero variance, pruned for high inter-correlation, and refined using recursive feature elimination with cross-validation, ultimately yielding 23 descriptors covering lipophilicity, ring content, van der Waals surface areas, and molecular quantum numbers. In parallel, two fingerprint families were generated: MACCS structural keys and Morgan circular fingerprints with a radius of two, together with a hybrid concatenation of the two. Four regression algorithms were benchmarked across the three representations, producing twelve configurations: Random Forest, a Keras deep neural network, Gradient Boosting, and XGBoost. Each configuration was trained under four fixed random seeds, and results were reported as means with standard deviations to capture stochastic sensitivity.

The Random Forest consistently delivered the most reliable balance between fit and generalization. Using the hybrid MACCS-Morgan representation, it achieved a coefficient of determination of 0.871 on the held-out test set and 0.863 on a strictly external validation set of 403 compounds that were never touched during model selection. Cross-validated Q2 values ranged from 0.888 to 0.902 across descriptor sets, and external Qext2 values reached as high as 0.919 with MACCS fingerprints alone. The deep neural network, although it attained higher training performance for some representations, showed a larger decline on held-out data, suggesting a greater tendency toward overfitting at the current sample size. Importantly, the team verified that no test or external compound shared an identical canonical SMILES with any training compound, and performance was unchanged after excluding high-similarity nearest neighbors, ruling out information leakage through duplicates.

Generalization was then stress-tested with a scaffold-disjoint partition, in which entire Bemis-Murcko scaffold families were assigned exclusively to training or held-out sets. Under this stricter regime, the hybrid Random Forest model achieved a mean test R2 of 0.858 and external validation R2 of 0.851, closely matching the random-split results. This indicates that the reported accuracy was not inflated by scaffold sharing and that the model genuinely extends to structural families absent from training. Prediction reliability was strongest in the moderate-to-potent activity range most relevant to virtual screening, while weak inhibitors with pIC50 below 7 showed larger errors and a modest tendency toward over-prediction, a pattern attributed to the relative scarcity of weak compounds in the curated data.

A distinctive strength of the study is its emphasis on knowing when to trust the model. Two applicability-domain diagnostics were implemented: a distance-based criterion derived from the mean and standard deviation of minimum training-set distances, and a leverage-based Williams plot with the classical threshold of three times the number of descriptors divided by the number of training compounds. Nearly all test and external-validation compounds fell inside the domain, and out-of-domain compounds in the independent BindingDB application set showed markedly larger errors than in-domain ones, with mean absolute errors of 1.884 versus 0.809 under the distance criterion. The researchers applied the model to 124 unique BindingDB compounds after removing 444 overlapping entries. Compounds sharing biphenyl scaffolds or cyclic peptide chemotypes with the training distribution showed small residuals, while structurally distinct pyrazolone derivatives and arylindanyloxypyridine frameworks showed substantially larger errors, pinpointing precisely where future data expansion would pay off.

The framework also translates statistical features into medicinal chemistry insight. SHAP analysis of the Random Forest identified the number of aromatic heterocycles, molar refractivity, and lipophilicity as the most influential descriptors, with threshold-like dependencies. The strongest potency-associated MACCS key simply records the presence of chlorine, but the corresponding Morgan features resolve it more sharply as a chlorinated aryl ring system and a chloro-substituted biaryl, discriminating potency with correlations up to 0.75. Conversely, a potency-decreasing CH2-O linkage characteristic of dibenzyl ether motifs is captured by MACCS but not the leading Morgan features, explaining why the two fingerprint blocks carry non-redundant, oppositely signed information. Scaffold analysis showed simple dibenzyl-ether and stilbene-like cores averaging pIC50 values of 6.3 to 7.3, while phenylpyridine and phenylpyrazine cores bearing basic amines averaged 8.4 to 10.2. These trends map directly onto the predominantly hydrophobic PD-L1 binding cleft, defined by residues such as Tyr56, Met115, and Tyr123, where halogenated aromatic extension and nitrogen heterocycles favor productive binding, and outlier compounds flagged by the model were examined through GNINA docking at this cleft against the reference inhibitor BMS-202.

The authors are candid about limits. Assay-format annotations come from ChEMBL descriptions rather than independently verified conditions, which may bias the set toward chemotypes preferentially tested in Alpha and HTRF formats, and a comparison against the published PDL1-inhi.predictor showed that an external model captured the BindingDB chemical space somewhat better. Yet the study’s contribution extends beyond any single metric: it delivers a transparent, inspectable, and adaptable platform that couples potency prediction with uncertainty quantification, supports pre-docking prioritization, and can be extended with uncertainty-guided active learning. The team suggests that high-ranking compounds within the applicability domain, especially those representing novel chemotypes, could be selected for prospective testing in the same Alpha or HTRF assays, followed by model retraining. For a field where reproducibility is often the missing ingredient, an open workflow anyone can audit and rerun on an ordinary laptop may prove as valuable as the model itself.

Subject of Research: Machine learning-guided discovery and prioritization of small-molecule PD-L1 inhibitors using an open-source KNIME workflow with curated bioactivity data and applicability-domain assessment

Article Title: Machine-learning-guided discovery and prioritization of PD-L1 inhibitors: An open-source KNIME workflow with curated bioactivity data, molecular fingerprints, and applicability-domain assessment

Article References: Sangsawat, M., Watpuang, B., Nathapinthu, T., Srisiriroj, P., Nalinratana, N., Taechawattananant, P., Thitikornpong, W., Feng, C., Chen, M., & Rojsitthisak, P. (2026). Machine-learning-guided discovery and prioritization of PD-L1 inhibitors: An open-source KNIME workflow with curated bioactivity data, molecular fingerprints, and applicability-domain assessment. Results in Chemistry, 30, Article 103877. https://doi.org/10.1016/j.rechem.2026.103877

Image Credits: AI Generated

DOI: 10.1016/j.rechem.2026.103877

Keywords: PD-L1, PD-1/PD-L1 immune checkpoint, machine learning, KNIME workflow, Random Forest, molecular fingerprints, MACCS, Morgan fingerprints, applicability domain, ChEMBL, BindingDB, cancer immunotherapy