# AI-designed CAR-T binders showed cell activity; four of ten killed targets specifically

> A study screened 12,000 designed CD20 binders in CAR-T cells. It found 707 pooled growth hits; four of ten individually tested showed significant specific target-cell killing.

By BIG CHANGE Editorial

Published: 2026-10-10T11:09:00.637Z
Updated: 2026-10-10T11:09:00.637Z
Canonical: https://bigchange.ai/blog/ai-designed-car-t-binders-cd20-study

![Conceptual illustration of an engineered immune cell and a separate target cell suspended in a culture well; not to scale.](https://bigchange.ai/api/media/file/car-t-binder-study-culture-well-hero-v2.png)
AI-generated conceptual illustration by BIG CHANGE.

Researchers asked 28 teams to design 12,000 short proteins that might recognize CD20, a marker on the surface of certain cancer cells. Instead of judging the designs only by whether an isolated protein could bind CD20, the [Bits to Binders study](https://link.springer.com/article/10.1038/s44320-026-00246-1) put each 80-amino-acid sequence into a chimeric antigen receptor (CAR) and screened how human T cells behaved. The open-access paper appeared in *Molecular Systems Biology* on October 8.

The results became narrower at each step. A pooled cell-growth screen flagged 707 designs with statistically significant enrichment of at least twofold in the condition containing CD20-positive cells. The researchers then tested ten leading, non-redundant designs one at a time. Seven showed target-associated proliferation, expansion and cytokine production above the negative control. Four produced significantly greater *specific* killing of CD20-positive cells than that control. These are laboratory measurements from one T-cell donor isolate, not evidence of a treatment working in patients.

## The big change

- **What changed:** A competition evaluated thousands of computationally designed CD20 binders as working parts of CARs in primary human T cells, then checked ten leading designs in individual functional assays.
- **How:** Teams combined varied generative, sequence-design, structure-prediction and selection methods. The researchers inserted the submitted sequences into one CAR scaffold, screened T-cell growth in a pool and measured several functions of selected constructs separately.
- **Why it matters:** The experiment tests a harder question than isolated binding: whether a designed recognition domain can be expressed by a T cell and help it respond to a target. The drop from pooled hits to specific killing shows why each assay must be read on its own terms.
- **What remains uncertain:** The study used one CAR scaffold and one donor isolate. Its pooled comparison cannot by itself isolate CD20-specific effects, and none of the designs was tested in animals or patients.

## From a designed binder to a cell response

A CAR gives a T cell an engineered surface receptor. Its outward-facing domain recognizes a target; the rest of the receptor helps turn that recognition into a cell response. A protein that looks promising in a binding model can still fail when it must be made, displayed and used by a living T cell. That gap was the competition's test.

Participants from 42 countries submitted designs during a six-week stage. Their workflows were varied. Some used diffusion models such as RFdiffusion or Chroma; others used methods including BindCraft, ColabDesign and protein language models. Many also used ProteinMPNN to propose or diversify sequences, structure predictors to assess candidates, and other computational filters. Several started from CD20 in a known antibody-bound structure or fragments of CD20-specific antibodies. The submitted sequences therefore came from mixed human and computational workflows, rather than a single AI model designing every successful receptor unaided.

The researchers synthesized DNA for the submitted 80-amino-acid domains and placed them in the same second-generation CAR backbone. Each domain's sequence served as a barcode in a pool of engineered T cells. They compared the abundance of those barcodes after two weeks of culture with CD20-positive Raji cells against cultures without those target cells. Of the 12,000 designs, 98.3% passed the DNA supplier's quality check, but 56.8% passed the study's minimum read threshold after cell growth. A failed or missing sequence at this stage cannot simply be counted as a binder that did not recognize CD20; production, expression and cell fitness also affect the funnel.

The pooled screen found 707 sequences with significant enrichment of at least twofold. That is 5.9% of all submissions, and team-level hit rates ranged from 0.6% to 38.4%. The experiment used three technical replicates per condition from one donor. Its comparison was Raji cells versus no Raji cells, rather than CD20-positive versus otherwise matched CD20-negative Raji cells. The authors accordingly caution that the pooled result is more directly a Raji-associated growth signal, even though subsequent binding and individual assays support CD20 recognition for some candidates.

## Ten individual tests changed the reading

The ten leading distinct designs were rebuilt as individual CAR constructs. The researchers measured receptor expression, T-cell division and expansion, production of two signaling proteins, and killing of target cells. Seven exceeded the unedited T-cell negative control in the target-associated proliferation, expansion and cytokine assays. Three of those seven, however, did not show significant specific target-cell killing.

The killing assay compared CD20-positive Raji cells with CD20-negative K562 cells. Only four designs, numbered 1506, 2383, 3494 and 3718, had significantly greater specific lysis than the negative control. The paper also reports background killing of both cell lines by several designs. That matters because killing a CD20-positive cell is not, by itself, proof that the CAR discriminates well between target and non-target cells. Across the functional readouts, the designed binders generally fell short of the study's established antibody-derived CAR control, with a reported exception in one cytokine measure.

An isolated-protein binding assay added another caution. Only three of the ten showed appreciable binding to detergent-solubilized CD20 under those conditions. The authors say this assay may not represent a binder's behavior when displayed in a CAR on a T cell. Conversely, a pooled growth signal is not sufficient to establish precise binding or safe target selectivity. The assays answer related, narrower questions.

## What the benchmark can guide

The study also compared design choices and sequence features with experimental outcomes. It found broad differences among methods, but the teams mixed tools and submitted uneven numbers of related sequences. The authors warn that a method's apparent success cannot generally be read as a clean causal comparison. They report that common structure-based confidence scores predicted little about recovery and proliferation in this dataset.

Some simpler sequence checks looked more useful within the same dataset. In a retrospective analysis, filters for features including DNA composition, repeated sequence motifs and particular amino-acid patterns increased the share of submitted designs with target-associated proliferation. Retrospective filtering of this competition is a way to analyze its failures; it has not yet demonstrated the same improvement in a new prospective design round.

The immediate contribution is an open experimental benchmark that connects design choices to several cell-level outcomes. Its limits are substantial: one donor isolate, one CAR scaffold, a pooled assay with an imperfect specificity comparison, and background killing in individual tests. The authors explicitly say the competition was meant to evaluate design methods in a therapeutic context, not to produce a clinic-ready CAR-T therapy.

## Sources & further reading

- [Kosonocky and colleagues, “Validation and analysis of 12,000 AI-driven CAR-T designs in the Bits to Binders competition,” *Molecular Systems Biology*, October 8, 2026](https://link.springer.com/article/10.1038/s44320-026-00246-1). The primary open-access paper reports the competition methods, pooled and individual assays, computational analysis and limitations. It is the basis for the numerical results here.
- [Study data and design sequences on Zenodo](https://doi.org/10.5281/zenodo.20451404). The paper links this repository for the submitted sequences and related data; BIG CHANGE did not rerun its analysis.

## Sources

- [Validation and analysis of 12,000 AI-driven CAR-T designs in the Bits to Binders competition](https://link.springer.com/article/10.1038/s44320-026-00246-1) — Primary open-access experimental paper; supports design methods, pooled and individual assays, analysis and stated limitations.
- [Bits to Binders study data and submitted design sequences](https://doi.org/10.5281/zenodo.20451404) — Repository linked from the paper. BIG CHANGE did not independently analyze the files.