Pythia Labs AI Antibodies Bind 15 of 22 Targets

A computer model designed antibodies from scratch against 22 disease-linked proteins, and lab screening found working binders for 15 of them after testing no more than 48 designs per target. That is the headline result from Pythia Labs, a Los Angeles and San Diego biotech company that announced its launch on 27 October.

The company made the announcement on Business Wire “alongside a preprint in bioRxiv” describing Pythia-1, the AI model behind its antibody designs. The real test of the claim is in that preprint, which has not been peer reviewed and stops at lab and cell experiments. None of the antibodies has been tried in an animal or a person.

In brief

Pythia Labs says its Pythia-1 model designed full-length human antibodies for 22 targets, with lab hits on 15, including all seven hard-to-reach cell receptors it tried.
Some designs blocked or switched on those receptors in cell tests, which the company calls a first for antibodies designed from scratch.
The work is an unreviewed company preprint with no animal or human data, and no funding has been disclosed.

Seven receptor targets, all with hits

The preprint, written entirely by Pythia Labs staff, split its targets into two groups. Fifteen were soluble proteins or the outer parts of receptors. Seven were G protein-coupled receptors, or GPCRs, a family of proteins that sit in the cell membrane and pass signals inward. The paper says conventional antibody discovery has struggled with such membrane proteins, partly because so little of the receptor pokes out of the cell.

According to the paper, first-round screening produced 31 hits against 8 of the 15 soluble targets and 18 hits across all seven GPCRs. Twelve of the soluble-target hits bound in the single- to double-digit nanomolar range, a measure of how tightly an antibody grips, where lower is tighter. The computing took about a day per target, the authors say, and initial lab screening took less than three weeks.

The antibodies were made in full-length IgG format, the standard Y-shaped antibody. The authors note that earlier working AI designs against GPCRs mostly used nanobodies, smaller single-domain antibodies that lack the part of an IgG that keeps it in the bloodstream for longer. The model was also steered towards human antibody frameworks and away from sequence features that cause manufacturing problems.

Designs that block a receptor, and one that switches it on

The more striking claim is about what the antibodies did, not just what they stuck to. Against interleukin-11, a protein with a possible role in lung and heart fibrosis, the best design, IL11_04, bound more tightly than a reference antibody from a patent (2.5 nanomolar against 9.4) and blocked signalling in a cell assay at lower doses than that control.

Among the receptors, a design called CCR8_38 bound CCR8, which is found on immune cells inside tumours, at 3 nanomolar, and five designs blocked the receptor’s signalling in a cell test. A design called MRGPRX2_15 did the opposite, activating MRGPRX2, a receptor on mast cells involved in allergic reactions, with a strength the authors call comparable to one of its natural triggers. “Generating antibodies that activate these receptors has been a longstanding goal in antibody discovery,” the company’s release says.

The authors write: “To our knowledge, this study provides the first demonstration of both GPCR agonism and antagonism by de novo designed full-length IgG antibodies.” That is the company’s own assessment, and it has not been checked by outside reviewers. Pythia’s release adds that the antibodies’ physical properties “compared favorably with clinical-stage therapeutic antibodies”.

A lab result built for checking, not yet a medicine

The most useful thing Pythia has done may be the least flashy. On its news page, the company offers the full sequences of every lead antibody in the paper “to support independent validation and further research”. That matters because the field is crowded with claims made by companies about their own models. When Anthropic said Claude designed protein binders against scorpion neurotoxins, it also released its designs. Open models such as RFD4-Proteina let anyone test the tools. Published sequences let rival labs make the antibodies and see whether the numbers hold.

The limits are just as clear. Every result comes from purified proteins or engineered cells, and the paper itself says that “in vivo assessment of pharmacokinetics, efficacy, and safety will be required” before any candidate can be judged as a drug. Seven of the 22 targets produced no hits at all. The preprint discloses that “this work was conducted by employees of Pythia Labs”, so no outside group has yet repeated it.

There is also a difference in emphasis between the two documents. The release speaks of designing for “hard-to-drug targets”, while the preprint singles out the seven GPCRs as the particularly challenging class. It also says the targets were picked partly because earlier computational antibody-design studies had already tested them. The release says “around 48” candidates were screened per target, while the preprint says at most 48. Neither changes the core result, but both are worth noting before the hype outruns the data.

Where Pythia sits in the AI drug-design race

Pythia joins a busy field. On Monday TSN covered AbOptimizer, an AI tool that made an existing antibody grip IL-6 13 times more tightly, which tunes antibodies rather than inventing them. Big drugmakers are also putting money behind AI-led development, as in Bristol Myers Squibb’s lead role in Pathos AI’s $237m Series E.

Pythia has not said who funds it or how much it has raised; neither the release nor its website names investors. Its team page lists 17 scientists and engineers, including scientific founder Joshua Laniado, and says the group has a collective track record of 12 or more biologics advanced to the clinic. The next sign of whether Pythia-1 delivers will be animal data on any of these leads, or an outside lab reporting results with the published sequences.