AI is no longer just finding antibodies, it is starting to decide how they act

With Pythia-1, researchers design via AI human antibodies capable of activating or blocking cellular receptors known to be difficult to target.

Blocking a cellular receptor or, conversely, triggering its activation: these two operations can produce radically different biological effects. Until now, designing fully human antibodies from scratch capable of precisely exerting these functions on certain membrane proteins represented a considerable challenge. Pythia Labs claims to have reached a milestone with Pythia-1, a generative model that simultaneously designs the atomic structure and sequence of antibodies. The experimental results are encouraging: several AI-imagined molecules successfully altered receptor activity in cells after a single design campaign.

In a study published on bioRxiv, researchers present antibodies capable of inhibiting or activating G protein-coupled receptors, known as GPCRs. This family of proteins is ubiquitous in our body, notably involved in cellular signal transmission and targeted by many drugs. Particularly Challenging Therapeutic Targets

Therapeutic antibodies have profoundly transformed the treatment of many diseases. Their ability to recognize specific molecular structures allows them to target biological mechanisms with high specificity.

Yet their discovery remains difficult when target proteins are partially embedded in cell membranes.

GPCRs illustrate this problem. These receptors cross the cell membrane multiple times and change conformation when receiving a signal. Their externally accessible areas are sometimes small, while their binding sites can be located in cavities that are difficult to reach with a conventional antibody.

Traditional methods, based notably on animal immunization or screening of molecular libraries, can identify antibodies capable of binding to a protein. However, this binding does not guarantee that the molecule will produce the desired therapeutic effect.

An antibody can recognize its target perfectly without preventing its function. Conversely, it can sometimes alter its activity in an unintended manner.

The ambition of Pythia-1 consists precisely of integrating this functional dimension right from the design phase. Designing the Molecule Based on the Desired Effect

Developed by teams at Pythia Labs, Pythia-1 is a generative model based on flow matching, capable of working at the atomic scale.

It jointly uses amino acid sequences, three-dimensional protein geometry, and constraints defined by researchers.

The process begins with the structure of the biological target. Scientists select a specific region to target, called an epitope, as well as a particular conformation of the receptor. This choice can guide the design toward a molecule intended to prevent receptor activation or, on the contrary, to promote it.

The model then proposes different antibody structures and their molecular sequences. Candidates are filtered according to their structural plausibility, predicted affinity, and several properties necessary for potential pharmaceutical development.

The approach also stands out due to the format of the molecules produced. Pythia-1 designs complete human immunoglobulin G, or IgG, including the heavy and light chains that characterize these antibodies. This architecture is more complex to design than that of nanobodies, smaller antibody fragments already used in several AI-assisted design works.

To limit difficulties encountered later in the laboratory, researchers also impose constraints on the sequences, notably to avoid certain characteristics likely to impair the stability or manufacturing of the molecules. A Single Design Campaign, 22 Biological Targets

The study covers 22 therapeutic targets: fifteen soluble proteins or extracellular domains of receptors, and seven GPCRs.

For each, researchers generated numerous computational proposals before selecting a maximum of 48 antibodies intended for experimental trials.

This detail is important: the 48 candidates correspond to the molecules selected for testing, not the total number of structures computationally explored.

The design and selection process could be completed in a single day. Initial biological assays, following DNA synthesis and antibody production, were carried out in less than three weeks.

The results are not uniform across all targets. Among the fifteen soluble proteins or extracellular domains, eight yielded antibodies showing detectable initial binding, representing 53% of the studied targets.

In contrast, all seven GPCRs tested produced at least one candidate with measurable binding.

In total, the team identified 31 positive candidates among the soluble targets and 18 among the GPCRs. Some molecules exhibited nanomolar affinity, meaning they could interact with their target at very low concentrations.

These results are interesting, but the true difficulty was not limited to achieving binding. It was necessary to demonstrate an effect on cellular function. Antibodies Capable of Blocking Receptors

Researchers notably worked on CCR8, a receptor associated with certain immune cells present in tumors, and CXCR4, involved in several immune cell trafficking and regulation mechanisms.

For CCR8, several antibodies generated by Pythia-1 demonstrated an ability to inhibit receptor activation by its natural ligand. Five molecules showed measurable antagonist activity, including two with inhibitory potency expressed by IC50 values below 10 nanomolar.

On CXCR4, the CXCR423 antibody achieved an IC50 of 5.5 nanomolar in a cell-based assay. Under the same experimental conditions, the reference antibody REGN7663 exhibited an IC50 of 33 nanomolar.

The difference suggests a more potent inhibition in this specific assay, without allowing for a conclusion of general therapeutic superiority.

The proposed structural models also show that certain antibody loops could penetrate deep into GPCR binding cavities, where their natural ligands usually act.

This observation helps explain how relatively large molecules could reach regions long considered difficult to access. The Most Surprising Experiment: An Antibody That Activates Its Target

One of the