Use ESMFold2 Online

ESMFold2: High-Affinity Structure Prediction and De Novo Protein Design

ESMFold2 is a state-of-the-art computational engine designed to predict protein structure and power the de novo design of entirely new, functional proteins. Built on the fourth generation of Evolutionary Scale Modeling (ESM), this world model family moves beyond simple static snapshots to reason about protein sequences at the true scale of evolution.

Released as a completely open-source tool under the MIT license, ESMFold2 enables researchers to bypass traditional, resource-intensive laboratory screening. By compressing years of experimental work into days, it allows therapeutic candidates to be fully optimized and tested computationally before they ever reach the bench.

Key Innovations: Recurrent Architecture & MSA-Free Folding

ESMFold2 introduces an architectural paradigm shift that optimizes both structural accuracy and computational throughput.

• Looped Transformer Architecture: Rather than passing sequence representations through a network a single time, ESMFold2 loops them through the same parameters repeatedly. The model learns to use each pass to iteratively refine its 3D structural representation based on what it computed in the previous loop.

• Inference-Time Scaling: Because its recurrent depth uses fixed parameters rather than an overfitted pool of explicit layers, users can allocate more computational budget to harder targets at inference time by simply running more loops—increasing accuracy without retraining the base model.

• Alignment-Free Generation: Operating directly from the learned biological representations of the ESMC language model, ESMFold2 captures complex evolutionary patterns encoded during pretraining. It does not require explicit Multiple Sequence Alignments (MSAs), though optional MSA inputs are supported to further boost accuracy.

• Two-Stage Design Algorithm: Seamlessly generates binders through a simple, parallelizable search procedure:

  1. Candidate Generation: Searches the model's latent representation space to generate tens of thousands of de novo candidates.

  2. Scoring & Ranking: Evaluates and filters candidates for predicted structural stability and high-affinity binding based entirely on internal confidence scores.

Performance Benchmarks

ESMFold2 and its computationally lighter counterpart, ESMFold2-Fast, offer a premier combination of speed and accuracy across macromolecular complex benchmarks.

Scientific Breakthroughs in Therapeutic Binder Design

Computational Scaling vs. Lab Outcomes

To evaluate the impact of inference compute scales, researchers generated candidate pools using varying computational budgets and screened the top 84 designs. Higher computational allocation directly translated into superior experimental success in the lab:

• Minibinders: De novo compact scaffolds with no predetermined structures saw success rates rise from 54% to 70%.

• Single-Chain Antibodies (scFvs): The highly demanding immunoglobulin framework—requiring custom design of hypervariable loops—saw lab success rates nearly double, moving from 12% to 21%.

Nanomolar Affirmation Across Disease Targets

Without any specialized task-specific training or fine-tuning, ESMFold2 successfully designed, generated, and lab-validated high-affinity binders against five clinically relevant cancer and immunoregulatory targets:

• EGFR & PDGFRβ: Critical receptor tyrosine kinases implicated in tumor growth.

• CTLA-4 & CD45: Major checkpoint regulators of immune cell signaling.

• PD-L1 Checkpoint Inhibition: An ESMFold2-designed de novo scFv bound to PD-L1 with an ultra-precise measured affinity of 4.3 nM. In cell-based biophysical assays, this computational design successfully relieved PD-L1-mediated T-cell suppression with a functional potency comparable to approved blockbusting cancer therapies.

De Novo Innovation vs. Database Retrieval

Biophysical and cell-based characterization confirmed that the high-affinity binders designed by ESMFold2 show minimal sequence similarity to any existing sequences in public biological databases. This confirms that the model is actively generating novel biochemical solutions based on an internalized understanding of biological rules, rather than simply retrieving or copying pre-existing historical answers.

ESMFold2 on Tamarind Bio: Run Millions of Virtual Experiments

By partnering with Tamarind Bio, ESMFold2 is democratized for therapeutic discovery pipelines, eliminating the overhead of GPU acceleration and customized inference orchestration.

By incorporating highly optimized, context-parallel kernels developed in collaboration with NVIDIA, the Tamarind Bio platform allows researchers to scale prediction to all of protein biology. Running millions of virtual design checkpoints is compressed into a matter of hours.

How to Use ESMFold2 on Tamarind Bio

1. Access the Engine: Log in to tamarind.bio and navigate to the open-source ESMFold2 workspace.

2. Define Your Design Objective: Upload your target molecular receptor structure (such as an immune checkpoint or oncogenic receptor).

3. Select Your Format: Choose between designing de novo minibinders or modeling hypervariable loop contacts on structural scFv frameworks.

4. Allocate Inference Scale: Set your computational budget (number of loops and sampling seeds) to scale inference resolution based on target difficulty.

5. Execute Parallel Scans: Run the generative pipeline to perform autonomous candidate generation, scoring, and confidence ranking.

6. Export High-Affinity Leads: Download the resulting high-confidence PDB structures and affinity predictions to immediately transition into wet-lab expression and synthesis.

Source