Non-opioid pain, discovery stage
Selective blockers for the channels
that carry pain
One in five adults suffers from chronic pain. Twenty years after recognizing that problem, opioids are still our primary means of addressing it. Some real progress, though, in the form of January 2025's approval of Suzetrigine, the first drug in a new non-opioid class in around 20 years. Appropriately, that event sent a rush of funding to companies working on sodium channels, but unfortunately, has herded a great deal of development activity to a single target. There's so much more to be done in pain.
Background
Why these drugs fail on selectivity
Selectivity is critical for pain targets because the channels that transmit pain have very closely related channels that perform different functions. Blocking those channels will lead to side effects that derail drug development efforts. This was roughly the fate of many Nav1.7 programs and has taught the industry an expensive lesson about selectivity being much more than an afterthought in pain programs.
Selectivity at pain targets is a physics problem because these closely related proteins can only be resolved in terms of their motions. The motion that distinguishes one channel from the next is the mechanism of selectivity for a given molecule so simulating the target in a state that the drug must recognize for function is absolutely critical. Unfortunately these simulations are the most computationally challenging in drug discovery, so are commonly truncated leading to an inability to understand selectivity. Our group has become very confident in our ability to perform these simulations to completion. We've tuned our computing to match the standards set by the best labs in this field and have numbers to prove it.
Where the money went
One example is the problem of matching targets to disease. Most failed clinical programs come from molecules applied to the wrong type of pain.
Venom peptides
Nature has solved pain signaling with remarkable efficacy. Ziconotide is made from cone snail peptides and exhibits exquisite specificity and selectivity, rivaling what you would see in any drug.
Three kinds of pain
Nociceptive, neuropathic and nociplastic pains each require different approaches to treatment. Reproducing that level of efficacy body-wide without a pump connected to the spine would be a remarkable feat of medicinal chemistry.
Measuring mental pain
Even modeling mental pain is an opportunity as the field is crying out for an acceptable way to measure it.
Programme 1 · CaV2.2 · Discovery
A cone snail peptide inside a human calcium channel.
Ziconotide (sold as Prialt) is a peptide analgesic drug which is extremely potent and targeted for pain but can only be delivered using a pump into the spinal fluid of patients. As such, it's only available to the tiniest fraction of patients for severe, intractable pain that does not respond to other interventions. At Anvelda we're redesigning the chemistry to maintain this incredible specificity whilst removing the need for the pump.
In PDB 7MIX all dots (amino acids) have been experimentally positioned. In total there are 1327 amino acids in this structure (for the channel), and the ziconotide drug has 25 amino acids. The peptide drug sits above the membrane blocking the pore.
| Programme | Target | Approach | Stage |
|---|---|---|---|
| Peptide redesign | CaV2.2 | Shrink and stabilise a venom blocker, keep the selectivity | Discovery |
| Selectivity modelling | Nav1.8, Nav1.7 | Score candidates against every channel state, not only the target | Discovery |
| Measurement | Mental pain | Build a measure the field agrees on, from published brain data | Research |
Each programme has a stage (column) which shows progress on each of these, as well as our next planned experiment, as we're often reminded, it's not a programme unless it has a specific next measurement!
Benchmark
The measurement
The same simulation on the same computer, before and after our engine tunes it to the chip.
We can now simulate 14.3x more time per day (on the same computer!). This isn't by performing less physics, it's entirely down to optimisations we make to the simulations in our engine to squeeze everything out of the chip. Being able to simulate 14.3x faster means we can simulate all of the states of the channels (not just the one in the structure above), which is incredibly useful to help us understand selectivity.
Figure 02 · throughput on the same machine, before and after tuning.
Figure 03 · a side by side comparison of identical runs, started from the same minimised structure and with the same random seed, so the proteins move identically.
Anvelda started life from two parts; training and running large ML models for world-leading AI research teams, and trading systems in finance that involve optimising everything down to the last nanosecond on specific pieces of hardware. Applying these learnings to HPC simulations of molecules was applying the same optimisations on bare metal to a new problem!
Reproducibility
The signed record
One last point, this figure comes with a fingerprint that contains all the details of how to recreate the figure, atom placement, force field, settings etc, so that people can check we've not fiddled, or recreate the figure themselves so they can run their own comparison studies.
| Structure | PDB 7MIX, membrane oriented copy from OPM |
|---|---|
| Complex | CaV2.2 alpha1 pore with ziconotide, 25 residues |
| Protein | 21,932 atoms, 1,351 residues |
| Solvent | 105,727 waters, 288 Na, 294 Cl at 0.15 M |
| Force field | amber14 with TIP3P water, particle mesh Ewald at 1.0 nm |
| Integrator | Langevin middle, 300 K, 4 fs step |
| Software | OpenMM 8.6.1 on Apple M5 Max |
It's also worth mentioning that as part of the setup procedure we energy minimise the system to ensure we have reasonable physics to start our timings from (i.e. nothing moving violently!). In this case the energy goes from a silly looking 9,684,191 to -5,604,909 kilo Joules a mole (!!), the point is the energy goes down!
Programme 3 · Mental pain · Research
Comparing the published pain maps
The statement that there is a common circuit in the brain for physical pain and for pain of rejection is becoming cast in stone simply by repeated mention in the scientific literature. However the published maps of activity tell a slightly different story, which may have quite significant implications for developing a drug against the second kind. We downloaded them from the published references over the course of an afternoon.
Figure 04 · every published pain map compared with every other inside the cingulate. Physical pain maps agree with each other. The rejection map agrees with none of them.
| Maps compared | Whole brain | In the cingulate |
|---|---|---|
| Physical pain against social rejection | 0.098 | 0.134 |
| Two physical pain maps against each other | 0.229 | 0.493 |
| Heat pain against fibromyalgia pain | 0.047 | 0.513 |
| Social rejection against fibromyalgia pain | minus 0.033 | minus 0.006 |
The maps for physical pain are strongly similar to each other (the degree of spatial similarity is about 1 for identical shape, and 0 for uncorrelated; the numbers for the first two pairs are about 0.49 and about 0.51), but the third one is very different to all the others. This implies that the pain of rejection is very different to the others, so any drug developed to target one form of pain is likely to have little effect on the others. More work needs to be done, of course, but we think it important to point out this discrepancy now.
Collaborate
Work with us
If you have a pain lab, or a tissue bank, or a project that needs large scale computations done properly, get in touch. We've got the computation and the physics, let's sign results together.
Contact usLaela Zorana, woman founder in production high performance computing. This work is carried out by CodeZorana LLC, Burr Ridge, Illinois, trading as Computelab.co, using Anvelda, the bare-metal engine.
Benchmark run 19 October 2026