[Communications Biology | (2024) 7:1051]

Structural shifts in a tiny channel protein help microbes pump out life-saving antibiotics. Antibiotic resistance is often described as a looming catastrophe, but the battle isn’t fought only in hospital wards. It also plays out at the atomic scale, deep inside bacterial membranes. There, a protein called TolC acts as a molecular escape hatch, helping E. coli and other Gram-negative bacteria expel dangerous compounds, including the β-lactam antibiotics that form the backbone of modern medicine.

TolC is no ordinary protein. It is a barrel-shaped channel that spans the bacterium’s outer membrane, linking up with partner proteins inside the cell to form powerful efflux pumps. These pumps can spit out a dizzying array of antibiotics, rendering many treatments ineffective. But TolC is also a double-edged sword: while it keeps poisons out, it can also serve as an entry point for viruses and bacterial toxins. Evolution, the researchers note, has tuned TolC into a delicate balancing act.

To probe how TolC opens and closes, we turned to molecular dynamics simulations, atom-by-atom movies of the protein in action. We discovered that even in its “resting” state, TolC is primed with collective motions that naturally favor pumping molecules outward. Then, using steered molecular dynamics, we pulled three different antibiotics, oxacillin, piperacillin, and carbenicillin, through TolC’s channel, watching where the drugs stuck and where they slipped through. Oxacillin, it turned out, slid out easily. Carbenicillin, by contrast, tangled in a web of hydrogen bonds, explaining why bacteria struggle to pump it out efficiently.

Next, we performed a feat of experimental saturation mutagenesis, mutating every single residue of TolC and testing how each mutant fared under antibiotic stress. Most mutations left TolC’s function intact. But a handful, especially near the channel’s entrance and exit, crippled efflux and made bacteria dramatically more vulnerable.

One mutation, D371G, was particularly revealing. It disrupted a set of salt bridges that normally choreograph the channel’s opening. Rather than freeing the passage, the mutation caused drugs to linger, effectively jamming the pump. We describe this as an allosteric occlusion effect, a local tweak that ripples through the protein’s long body and alters function at a distance.

Taken together, the results paint TolC as an evolutionarily optimized machine: tuned for efficiency, but vulnerable at specific hinges and gates. The work also points to a strategy for antibiotic design: target the weak spots at TolC’s periplasmic entryway, where drugs first encounter the pump. By crafting molecules that bind or stall here, scientists may be able to sidestep efflux and restore antibiotic potency.

As we may put it, understanding TolC’s shifting architecture isn’t just a matter of molecular curiosity. It could be the key to slowing the relentless arms race between bacteria and modern medicine, one salt bridge, one channel opening at a time.

How a “multiply perturbed response” maps hidden control knobs for biosensors and beyond

Proteins don’t just exist; they perform. They flex, breathe, and communicate, flipping between shapes that switch biological circuits on and off. Understanding where to push a protein has long been a guessing game. In biosensor design, researchers splice fluorescent reporters into sensing proteins, hoping that a ligand-binding event will ripple through the structure and light up the signal. Trial and error can work; but it’s a painfully slow way to listen to molecular whispers.

Now a physics-guided framework called Multiply Perturbed Response (MPR) offers a shortcut. Instead of probing one amino acid at a time, it asks which small combination of residues, when pushed together, most efficiently drives a protein from its inactive “apo” shape to its active “holo” form. Built on linear-response theory, MPR extends classical Perturbation Response Scanning, mapping not just single-site sensitivity but the cooperative choreography of multiple sites that is the essence of allostery.

When applied across model proteins that serve as the sensory cores of fluorescent biosensors, MPR consistently rediscovered experimentally validated “insertion neighborhoods.” They are the hot spots where mechanical signals travel cleanly from binding pocket to reporter domain. These allosteric hubs act like transmission gears, coordinating local ligand binding with global conformational motion. The analysis also revealed which regions to avoid: rigid scaffolds that mute signals or structural hinges whose disruption could cause misfolding.

The implications reach far beyond biosensing. Many enzymes and receptors are naturally allosteric, integrating cellular signals or resisting drugs by rerouting communication through alternative pathways. MPR provides a way to see those paths quantitatively. The same physics that guides fluorescent-sensor optimization could, in principle, expose the hidden relay networks that let pathogens evolve drug resistance or that govern how therapeutic molecules bias a receptor’s motion. In silico, the identified residues can even serve as collective variables that are explicit levers for steering molecular simulations or designing molecules that modulate motion rather than block binding.

Linear models can’t capture every nuance of a protein’s restless landscape. Yet MPR delivers something long missing from that landscape: a simple, interpretable atlas of the few levers most likely to move the machine; and, perhaps, to re-tune life’s most adaptable engines.

[Mathematical Biosciences | (2024) 372: 109191]

When bacteria evolve to shrug off antibiotics, doctors usually switch drugs and hope for the best. Our framework suggests something more audacious: plan a sequence of drugs that steers evolution backward, nudging microbes toward a genotype that’s once again vulnerable.

The idea starts with lab measurements of how bacterial genotypes grow under different antibiotics. From those data, we construct a genotype graph where each drug induces its own set of likely evolutionary moves. Stack those moves over time and you get a map for plotting multi-step treatment plans. Two simple rules translate growth rates into movement on the graph. In the Correlated Probability Model (CPM), the bigger a neighbor’s growth advantage, the more likely evolution flows that way. In the Equal Probability Model (EPM), all fitter neighbors are treated the same. Either way, each antibiotic corresponds to a transition matrix, the probability of hopping from one genotype to another in a single step.

With these maps in hand, we tackle two versions of the problem: static schedules commit to a fixed drug order in advance. This is tricky, NP-hard in general, so we reformulate the search as a mixed-integer linear program to find high-probability routes efficiently. Dynamic policies choose the next drug based on where evolution actually lands, a GPS that can reroute in real time. Here, dynamic programming delivers the optimal choice at each step.

Real data are noisy, so we build uncertainty into the planning. Instead of locking into a single, possibly wrong transition matrix per drug, we average over many plausible ones estimated via sample-average approximation. That yields robust schedules without blowing up the computation. What happens when you run the numbers? Across dozens of scenarios, fifteen starting genotypes, up to sixteen steps, CPM and EPM, longer plans help, and dynamic policies help even more. Static schedules steadily lift the chance of returning to the wild type, reaching roughly greater than 0.70 with CPM and greater than 0.50 with EPM by sixteen steps. Dynamic policies, however, can push success above 0.99, achieved by eleven steps under CPM and fourteen under EPM, while remaining fast to compute.

Why it matters: collateral sensitivity, the tendency for resistance to one drug to increase sensitivity to another, won’t save us automatically. But if we can chart where it exists and sequence drugs to exploit it, treatment becomes navigation, not brute force. The time machine formalizes that navigation and bakes in measurement uncertainty, offering clinicians and lab scientists a reproducible way to test smarter drug orders.

Caveats remain. The approach relies on having genotype–drug growth data, and real infections involve immune systems, pharmacokinetics, and spatial structure that lab plates can’t capture. Still, the framework is modular and the code is public, inviting new datasets, transition rules, and constraints. The big idea endures: don’t just hit bacteria harder, steer them, step by calculated step, back to where our medicines work.