Anthropic says Claude optimized more than 30 artificial intelligence models used in molecular biology in less than four weeks. The results reduce inference time by approximately 4x on average and make it possible to analyze much larger biomolecular systems with a single GPU-equipped machine.
The news matters because protein prediction and design often require large amounts of memory, computing time, and specialized expertise. What changes when those barriers come down? More researchers can experiment with molecules related to diseases, drugs, and complex cellular processes.
Claude optimizes molecular biology models
Biomolecular structure models try to predict the three-dimensional shape of a protein, a molecular complex, or a combination of proteins and other molecules based on their components. In protein design, the goal is reversed: to create a molecule with a specific function.
These tools are used in pharmaceutical research, treatment development, and disease studies. However, many open-source models require expensive infrastructure to run quickly.
Anthropic worked with Claude to optimize more than 30 models dedicated to tasks such as:
- Biomolecular structure prediction.
- Protein design.
- Protein language models.
- Genomic analysis.
According to the company, the optimizations sped up these tasks by approximately 4x with minimal loss of accuracy. When identical results to the originals were required, the improvement was close to 2x.
The difference is not only about making a model respond faster. It is also about reducing the resources needed so that more laboratories can use it.
The role of FlashPairformer
A significant part of the work focused on two operations used by models such as AlphaFold3, OpenFold3, and Boltz-2: triangle attention and triangle multiplication.
These operations help represent the geometry of a biomolecular system, but they are very expensive. Their time and memory requirements grow cubically with the size of the system. In simple terms, doubling the number of elements can multiply the computational cost by eight.
To address this problem, Anthropic and Claude developed FlashPairformer, a set of custom kernels for GPUs. A kernel is a piece of low-level software that allows operations to run more efficiently on specialized hardware.
Anthropic says FlashPairformer outperformed industry reference implementations, with improvements of between 2.7x and 2.9x in triangle attention, and between 1.7x and 3.2x in triangle multiplication, depending on the model configuration.
In addition to these reusable kernels, Claude introduced optimizations specific to each model. These included temporarily storing repeated calculations and removing code branches that always produced the same result.
The company verified that the accelerated versions preserved performance on final tasks, such as structure prediction. This detail is essential: a faster response is not useful if it loses biologically relevant accuracy.
Larger biomolecular systems
Anthropic also asked Claude to reduce memory usage. The result was a mode called Big, designed to work with systems containing more than 10,000 tokens on a single machine with an NVIDIA GPU.
In this context, a token does not necessarily represent a word. It may correspond to amino acids, nucleotides, atoms from small molecules, or ions. That is why the number of tokens serves as an approximate measure of the size and complexity of the molecular system.
With Big mode, Anthropic reports successful predictions for structures such as:
- The human mitochondrial complex I.
- The TRiC chaperone complex.
- A proteasome.
- A bacterial ribosome.
Some of these structures exceed 10,000 tokens and, according to the company, closely match their experimentally determined structures. As a reference, a 40S ribosome correctly predicted by AlphaFold3 had 7,663 tokens.
Claude also generated predictions of viral capsids and protein compartments ranging from more than 31,000 to more than 70,000 tokens using a machine with eight NVIDIA B300 GPUs. These structures were not predicted correctly, which Anthropic partly attributes to their size far exceeding the context on which the models were trained.
Even so, the experiment shows that the memory limit for attempting these calculations was significantly reduced. The next challenge will be improving accuracy, not just making inference possible.
Protein design with less computing power
Anthropic compared this work with an earlier protein design experiment. At that time, Claude could use up to $10,000 in Modal infrastructure per target, equivalent to around 2,500 hours on an NVIDIA H100 GPU.
In the new test, a single model was given access to an NVIDIA H200 GPU for 24 hours. It used a prompt of approximately 1,100 words, preinstalled tools, and had no subagents or human guidance during the design process.
The evaluation included three Claude models and 16 targets. Anthropic used ipSAE, a computational metric associated with the probability that a design will bind to its target in the laboratory.
The reported results indicate that designs with medium and high scores reached values similar to those of previous campaigns, but using approximately 100 times fewer GPU hours. Including the cost of tokens, the company estimates that around $150 in GPU and model usage was enough to match the computational performance of its previous campaigns.
This does not mean that an AI-designed protein automatically becomes a drug. The designs must be produced, tested, and validated in real experiments. Simulation can reduce the search space, but it does not replace experimental biology.
Competition with Adaptyv Bio
To test these capabilities outside its own experiments, Anthropic and Adaptyv Bio announced a protein design competition. The program will address five advanced problems, including:
- Reactivity across different species.
- pH sensitivity.
- Peptide specificity for MHC complexes.
- Designing proteins that target G protein-coupled receptors, known as GPCRs.
The organizations plan to experimentally validate more than 5,000 designs submitted by the community. The announced support includes up to $1 million in Claude credits, $250,000 in Modal computing credits, and DNA provided by Twist Bioscience.
Adaptyv Bio will conduct part of the validation in the laboratory. This means the competition will not be limited to comparing scores on a screen, but will attempt to determine which designs work under experimental conditions.
Anthropic also opened its Life Sciences Verification Program in public beta, aimed at organizations working in biological research that need to verify AI capabilities in scientific contexts.
What it means for research
The most relevant point in this announcement is not that Claude has replaced engineers or biologists. It is that a general-purpose model was able to collaborate on a highly specialized part of scientific software development: optimizing inference and GPU kernels.
The human team supervising the work had experience in biomolecular modeling, but no previous experience in inference optimization or kernel engineering. Anthropic says Claude accelerated more than 30 models in less than four weeks, a task that would normally require weeks of work from specialized teams for each model.
The optimized code will be published as open source, according to Anthropic. This could benefit academic laboratories, biotechnology startups, and developers that do not have large infrastructure budgets.
The most concrete promise of scientific AI is not that it will magically make discoveries. It is that it can turn slow, expensive, and difficult processes into tools that more people can try.
Accuracy, reproducibility, and experimental validation will remain essential. But if the optimizations hold up outside Anthropic's environments, the impact could be felt in a very practical part of research: more computational experiments, at lower cost and on more complex biological systems.
Original source
https://www.anthropic.com/research/claude-uplifts-biomolecular-modeling