How Molecule Uses LiteFold to Design Non-Canonical Peptide Binders for Undrugged GPCR Targets
Molecule ·
Molecule's discovery team designs peptide binders against difficult, often poorly-characterized targets — the kind of GPCR class A receptors where the binding mechanism itself is still an open scientific question, not just an engineering one. That combination of exploratory biology and heavy computational lift is where the team turned to LiteFold.
Background
Molecule's team works at the harder edge of peptide therapeutics: designing binders for receptors where the literature is thin and the mechanism of action isn't fully resolved. That work depends on two things happening well at once — fast, thorough literature synthesis to ground every design decision, and molecular design tooling precise enough to model non-standard chemistry like non-canonical amino acids (ncAAs).
Most of Molecule's binders for other programs were built outside LiteFold, using tools like RFdiffusion3 and BindCraft that existed before LiteFold could support them. LiteFold's role sharpened on the harder, more novel slice of the pipeline: peptides built from non-canonical amino acids against two class A GPCR targets, KISS1R and OX2R.
Challenge
Designing against KISS1R and OX2R meant working in territory where standard peptide-receptor binding assumptions don't hold. The team needed to run literature review across many candidates and hypotheses in parallel without losing context between them, and it needed peptide design tooling capable of simulating non-canonical amino acid chemistry — capability that, at the time, didn't exist in an off-the-shelf platform.
The harder failure mode wasn't a tooling gap, though — it was scientific over-investment. On one program, deep reasoning about analog design (ncAA substitutions, cyclization strategies) continued for a long time before it became clear the underlying mechanism and target ensemble were not well enough resolved for the peptide to be improvable at all. The interaction looked more like adjacent-receptor signaling than a standard peptide-receptor complex — a case where more reasoning wasn't going to fix an unresolved biology problem.
Solution
Molecule ran its workflow through LiteFold's multi-agent orchestration and Rosalind, LiteFold's AI co-scientist, using two capabilities built specifically for this kind of work:
Literature review as a multi-agent workflow. Rosalind orchestrated multiple queries across multiple candidate molecules simultaneously — organizing results, data, candidates, and hypotheses without losing context between them, something Molecule's team called out as the most useful task LiteFold performed for them.
MD simulation of non-canonical amino acids. LiteFold's peptide design tooling was extended, specifically for Molecule's needs, to support molecular dynamics simulation of ncAAs — the modeling capability the two GPCR programs depended on.
Using this combination, Molecule designed ncAA peptides for KISS1R and OX2R that are now staged for wet-lab validation directly in a functional, in-vivo cell assay — a step up from a simple binding assay that will also surface agonism, proteolytic degradation, half-life, and off-target selectivity. Eurofins and O2H Discovery are the CRO partners running that validation, targeted for the end of Q3.
Results
Molecule's team compared Rosalind directly against other AI co-scientist tools they'd used, including Claude Science and K-dense:
"Rosalind seems to reason better than other AI co-scientists and always performed deeper analysis, retrieving more findings and generating more hypotheses. It performed similarly to K-dense in literature review and analysis, but wins in agentic orchestration of bioinformatic tools."
That same depth of reasoning surfaced a real limitation. Rosalind pursued a full analog-design workthrough for a peptide before Molecule's team established that the target mechanism wasn't scientifically resolved enough to justify it:
"One can argue this was also my error."
Rather than treat that as a dead end, Molecule closed the gap on their side — adding an internal "druggability" check to their own workflow to filter out targets where the mechanism isn't resolved enough to reason about further, avoiding wasted compute on both sides.
On day-to-day usefulness, the split between human and agentic use was clear:
"For human use it's the ease of visualizing designs and working together with Rosalind like a scientific co-pilot, able to jump from idea to execution. For an agent, retrieval of information, sequences, production of high quality scientific papers."
What's next
Molecule's next ask points at a gap in the field rather than in LiteFold specifically — an agonist prediction tool, which needs training data from functional assays to build:
"I have already started defining the problem and working on the solution, and once we get data on Q3 this could be accelerated."
With the KISS1R and OX2R validation results expected by the end of Q3, Molecule and LiteFold are positioned to turn that functional assay data into exactly the training set an agonist prediction tool would need.