How It Works
Why cyclic peptides
Most disease-relevant protein interactions don't fit the two tools most drug discovery programs reach for first. Small molecules are cheap to make and easy to dose, but they struggle against the large, relatively flat interfaces typical of protein-protein interactions. Antibodies can engage those interfaces well, but they're large, expensive to produce, generally can't cross a cell membrane, and carry a much longer development timeline.
Cyclic peptides sit in the gap between them: large enough to cover a real protein-protein interface, small enough to keep some of the developability advantages of a small molecule. Cyclization - closing the chain into a ring, most commonly through a disulfide bridge - locks the peptide into a defined shape instead of leaving it to flop through an ensemble of conformations, which improves both binding affinity and resistance to enzymatic degradation.
Why this is a hard computational problem
The shape of the ring and the sequence that occupies it are coupled - you can't solve one without the other - and the space of valid, cyclizable options is enormous. Even after you have a candidate, predicting whether it will actually bind (not just whether it looks structurally plausible) is its own open problem. Treating these as one blurred-together guess is a common source of false positives in computational design.
Our approach
Rather than exploring blind, every design is anchored to a real, already-established protein interaction - a target, and a fragment already known to engage it - so the search starts from a physically grounded starting point instead of a guess. From there:
- You choose the angle of attack. Design a binder against the target's surface, against the interacting fragment, or both at once - your call, and you only pay compute for the side(s) you select.
- Every candidate is independently validated, not self-graded - structural plausibility and predicted binding strength are checked by a separate step before anything is ranked, and low-confidence signals are surfaced rather than hidden.
- The ranking is honest about its own limits. When a batch's predicted binding strengths are too close together to be statistically meaningful, the system says so and falls back to structural confidence instead of presenting a false ordering.
- Favorites can be refined further, on demand - physically relaxed into a cleaner geometry, or converted into a more protease-stable bicyclic form - only for the candidates you choose, never applied automatically.
Built-in developability screening.
Before a raw candidate sequence is even sent to the expensive fold-and-score validation step, it's screened for practical manufacturability - net charge, hydrophobicity, hydrophobic moment, and aggregation propensity - the same kind of checks a chemist runs before ordering a peptide. Without this step, a design search can gravitate toward sequences that look great on paper but are hydrophobic and aggregation-prone enough that they're difficult or impossible to synthesize and purify at bench scale. That failure mode is filtered out before a candidate ever reaches your shortlist - what you see ranked isn't just structurally plausible, it's buildable.
This general approach is the same one used for classic, well-studied drug-discovery targets Example: MDM2 / p53 is one well-known example of the kind of protein interaction it's built to attack - but the method itself is target-agnostic.
What you get, and why it's built to be trusted
The deliverable is a ranked shortlist, not a dump of raw model output: for each candidate, a 3D structure, a sequence, and the confidence and affinity signal behind its rank - plus an honest flag whenever the affinity numbers aren't reliable enough to lean on. No single step grades its own homework: generation, validation, and scoring are separate checks, and a candidate has to clear all of them to reach your shortlist. You decide which side of the interaction to attack, and nothing about that choice - including what it costs - is hidden from you.
Your data does not become part of training data for our models or any third party services. All your data is encrypted and only accessible by the logged in user (you) with your password.
More to come
This page is the first entry in what we intend to build out as an ongoing knowledge base - future additions will go deeper on how to choose good hotspot residues, how to read and interpret confidence and affinity metrics, and case studies across different target classes. If there's a topic you'd want covered first, that's worth flagging back to us directly.
More coming soon. In the meantime, see the User Guide or the pricing details.