Machine learning and free energy methods discover new tuberculosis KasA inhibitors

In the race to develop new antibiotics against tuberculosis, one of humanity’s oldest and deadliest infectious diseases, a team of computational chemists has turned to an unconventional strategy: letting fragments of known drug molecules recombine into entirely new chemical entities, guided by machine learning and rigorous physics-based simulations. The result, published in the journal Molecular Diversity, is a shortlist of five synthetic molecules that appear to bind the tuberculosis protein KasA more tightly than the reference inhibitor used to solve the protein’s crystal structure, offering a promising starting point for the next generation of anti-TB drug candidates.

The study, led by Md Ataul Islam of SilicoScientia Private Limited in Bengaluru, India, together with collaborators at King Saud University in Saudi Arabia, University College London, and M.H. Saboo Siddik College of Engineering in Mumbai, focused on KasA, a beta-ketoacyl-acyl carrier protein synthase that sits at the heart of the bacterium’s cell wall manufacturing machinery. KasA is not an arbitrary target. Mycobacterium tuberculosis owes much of its remarkable resilience to an exceptionally waxy cell wall rich in mycolic acids, long-chain fatty acids that shield the pathogen from antibiotics and the host immune system alike. KasA catalyzes a key carbon-carbon bond-forming, Claisen-type condensation step in the FAS-II fatty acid synthesis pathway that builds these mycolic acids, working in concert with the FAS-I system. When KasA is depleted experimentally, the bacterium’s cell wall integrity collapses and the cells lyse, which is precisely why disrupting this enzyme has long been considered a compelling way to kill the bacillus. The enzyme is also the known target of thiolactomycin, a natural product inhibitor, and of a synthetic anti-tubercular scaffold whose cellular target was confirmed as KasA in earlier work, underscoring that the protein is pharmacologically tractable.

Rather than screening millions of commercially available compounds, the team embraced fragment-based drug design, an approach that starts from small, simple chemical fragments and grows or merges them into larger, drug-like molecules. Fragment-based methods have matured considerably since their early days and now underpin numerous clinical candidates, but they traditionally rely on biophysical techniques such as X-ray crystallography or nuclear magnetic resonance to detect weak fragment binding. In this study, the researchers took a fully computational route. Using a tool called MacFrag, which segments large molecules into smaller fragments with desirable properties consistent with the so-called “rule of three” for fragments, they decomposed known KasA inhibitors, including thiolactomycin-based chemotypes, into a library of chemical building blocks. These fragments then served as raw material for Fragmenstein, an open-source computational platform that merges and links fragments by respecting the binding poses they would occupy in the protein’s crystallographic binding site. Fragmenstein effectively stitches new molecules together such that each fragment-derived portion retains its predicted binding interactions, a strict conservation-of-binding philosophy that maximizes the odds that the resulting hybrid compounds will actually bind.

From this fragment-merging exercise, the team generated a library of 8,390 candidate molecules. That number was then cut down through a multi-tiered computational funnel designed to mimic the successive filters of a real drug discovery campaign. First came pharmacokinetic and toxicity filtering using ADMET-AI, a machine learning platform trained to predict how compounds behave in terms of absorption, distribution, metabolism, excretion and toxicity. Compounds with red flags for oral drug-likeness or safety were removed. Next, the surviving molecules were docked into the KasA active site, the elongated channel where the enzyme normally accommodates fatty acyl substrates. Docking, which computationally samples the orientations and conformations of a small molecule within a protein pocket and scores the resulting interactions, was performed with multiple engines including AutoDock Vina, allowing the researchers to prioritize compounds showing consistently favorable poses rather than single-program artifacts. Chemical space analysis using dimensionality-reduction techniques helped confirm that the selected compounds explored diverse regions of scaffold space rather than clustering around a single chemotype.

The rigorous funnel ultimately delivered five standouts, named KasA_FB1 through KasA_FB5. In molecular docking, their predicted binding energies were −7.80, −8.30, −9.00, −7.80 and −9.00 kcal/mol respectively, all outperforming the co-crystallized reference ligand thiolactomycin, which scored −7.20 kcal/mol against the same target. In drug discovery terms, a full kilocalorie-per-mole advantage in docking scores is meaningful, since each kilocalorie corresponds to roughly an order of magnitude change in binding affinity. But docking alone is notorious for producing false positives, because it typically evaluates a rigid or only partially flexible protein snapshot and relies on approximate scoring functions. The team therefore subjected all five molecules to molecular dynamics simulations, in which the protein-ligand complexes are embedded in explicit water and simulated over time using the GROMACS engine with the CHARMM36 protein force field and SwissParam-derived parameters for the small molecules. These simulations reveal whether a ligand remains stably seated in the binding pocket under realistic thermal motion or whether it drifts, flips, or loses its key interactions within nanoseconds.

The dynamics results were encouraging. All five candidates maintained dynamic stability comparable to or better than thiolactomycin, preserving the hydrogen bonds and hydrophobic contacts that anchor them in the KasA channel. The researchers then quantified binding strength more precisely using two complementary free energy methods. The first, MM-GBSA, combines molecular mechanics energies with implicit-solvent estimates to compute binding free energies from simulation snapshots, a widely used middle-ground approach that balances accuracy and computational cost. The second, free energy perturbation, or FEP, is considered one of the most rigorous techniques in computational chemistry. FEP calculates the free energy change of chemically transforming one ligand into another while bound to the protein, using a thermodynamic cycle and a series of intermediate “alchemical” states. Because it samples the explicit solvent environment and accounts for entropic effects far more faithfully than docking or end-point methods, FEP is often referred to as the gold standard for comparing the affinities of related compounds. Both approaches confirmed that the new molecules out-bind the reference ligand, and per-residue energy decomposition pinpointed which amino acids contribute most to the favorable binding, information that medicinal chemists can exploit when optimizing the compounds further.

The significance of this work extends beyond the five specific molecules. Tuberculosis remains the world’s leading infectious killer in recent global tallies, and the rise of multidrug-resistant, extensively drug-resistant, and even totally drug-resistant strains has made the discovery of agents with novel mechanisms an urgent priority. Existing front-line therapies rely heavily on disrupting mycolic acid biosynthesis, as isoniazid and ethionamide do, and resistance to these drugs frequently maps onto the very pathway that KasA controls. A new chemotype that inhibits KasA through a distinct binding mode could therefore sidestep existing resistance mechanisms. Fragment-based approaches have been applied against tuberculosis targets before, but the pipeline demonstrated here, combining automated fragment generation, fragment merging with binding-pose conservation, machine learning-based ADMET prediction, ensemble docking, and alchemical free energy calculations, represents a template that could be rapidly redeployed against other essential mycobacterial enzymes.

The authors have also made an unusually complete set of data publicly available, hosting the MacFrag fragment library, all 8,390 generated molecules with their ADMET predictions, docking outputs, both rounds of molecular dynamics trajectories, per-residue energy decompositions, hydrogen bond occupancy analyses, chemical space plots, and the FEP energies on a public repository. This transparency allows other groups to independently verify the predictions, repurpose the fragment library for related targets, or extend the optimization campaign without repeating the initial computational heavy lifting. The work was funded by the Ongoing Research Funding program at King Saud University, Riyadh.

The researchers are careful to note that these are computational predictions, and the five candidate molecules now require experimental validation, beginning with enzymatic inhibition assays against purified KasA and extending to minimum inhibitory concentration testing against live M. tuberculosis cultures, followed by structural confirmation of the binding mode. Many computationally promising compounds fail at this stage, often due to factors that simulations capture imperfectly, such as cell wall penetration in mycobacteria, which presents one of the most formidable permeability barriers in microbiology. Nevertheless, the demonstrated superiority of these fragment-derived molecules over the existing reference ligand, across orthogonal computational methods ranging from docking to alchemical free energy calculations, gives the candidates a stronger than usual prior. If experimental testing bears out the predictions, the study will stand as a compelling example of how machine learning, fragment-based design, and physics-based simulation can compress the earliest and most expensive phases of antibiotic discovery into a purely computational campaign, delivering chemists a prioritized, synthesizable shortlist in a fraction of the traditional timeline.