TxBench-AB tested AI agents on therapeutic antibody decisions: the best score was 53,0%
TxBench-AB tested AI agents on therapeutic antibody decisions: the best score was 53,0%
On 2 September, LatchBio published TxBench-Antibody Discovery, a set of 100 tasks in which an agent receives experimental data before making the next decision in therapeutic antibody development. According to the authors, Opus 5 running in Claude Code passed 53,0% of the scored attempts.
A team developing a therapeutic antibody must integrate results from different experiments. Binding to a biological target measures one property of the molecule, activity in cells measures another, and manufacturability measures a third. The choice of the next experiment depends on which of these measurements can answer the relevant question about the candidate.
LatchBio previously applied this framework to small molecules in the TxBench-PP benchmark. The new benchmark extends the evaluation to antibodies. The same numerical value can have different meanings in different measurement systems, so results from one experiment must be compared with results from another.
In TxBench-AB, the agent receives a dataset before making one of these decisions: select a candidate, modify its design, repeat the assay, or terminate the program. The automated evaluator accepts a decision only when it includes the measurements and data that support it. The authors describe the benchmark’s aim as follows:
“The goal is to determine what the available data support and which research action should follow, rather than merely reproducing an experimental result.”
One of the open tasks concerns an antibody intended to cross the blood-brain barrier, the layer of cells between the blood and brain tissue. The agent receives the antibody concentration in the whole brain, separate measurements from capillaries and tissue, and microscopy data. The decision requires these results to be compared so that antibody remaining in blood vessels can be distinguished from antibody that has reached brain tissue.
These 100 tasks cover ten stages of development, from target selection and the search for a binding molecule to testing cellular activity and assessing candidate risks. The authors ran three attempts for each combination of model and execution environment. Opus 5 in Claude Code passed 152 of 287 scored attempts. In 28 of 92 tasks with three scored attempts, this configuration failed all three times.
The overall score combines different scientific tasks. Opus 5 performed better on questions about targets and molecular activity in cells, while Grok 4.6 performed better on experimental design, binding kinetics, and candidate risk assessment. Each model’s strengths varied with the type of decision and the data it had to interpret.
The benchmark authors classified 82,3% of the 3 517 scored failed attempts as errors in interpreting data during drug development. Failures often began with selection of the wrong metric, transfer of a result between different biological conditions, or failure to check experimental quality before drawing a conclusion. Additional tokens and tool calls sometimes merely extended a calculation based on an invalid comparison.