ESMFold, AlphaFold 2, and ColabFold are often lined up as three folders and ranked by accuracy. That framing misses the point. They answer different questions, and the deepest difference is not speed, it is what each one reads before it predicts. AlphaFold 2 and ColabFold read an evolutionary alignment. ESMFold reads only your sequence. Once you know which of those inputs your problem has, the choice is usually obvious.
This post walks through the real difference, then a decision rule that holds up in practice.
The real difference is the input, not the speed
AlphaFold 2 (Jumper et al., Nature 2021) learns from a multiple-sequence alignment (MSA). It gathers hundreds of evolutionary relatives of your sequence and reads the pattern of conservation and covariation between columns to infer which residues sit close in space. The alignment is where most of the signal lives. ColabFold (Mirdita et al., Nat Methods 2022) is the same AlphaFold 2 weights with a faster MSA frontend, MMseqs2 instead of jackhmmer plus HHblits, so it builds that alignment in seconds rather than the better part of an hour.
ESMFold (Lin et al., Science 2023) takes a different route. It is built on ESM-2, a protein language model trained on hundreds of millions of sequences, with a folding head on top. The evolutionary signal that AlphaFold reads from an alignment at prediction time, ESMFold has already absorbed into the weights of the language model during training. So it folds from a single sequence, no alignment step at all. That is why it is fast, and it is also why it behaves differently on the sequences where an alignment would have mattered.
So the three are not a simple accuracy ladder. AlphaFold and ColabFold are the MSA route to a fold. ESMFold is the single-sequence route.
How we actually use them: triage, then decide
The mistake is treating these as rivals for a single job. In practice they are a pipeline.
It comes down to throughput. Cofold every candidate with its target and you spend most of your compute confirming that obviously bad designs are bad. ESMFold cuts those cheaply on the single chain first, so the expensive complex prediction runs only where the answer changes a decision.
When single-sequence is a feature, not a compromise
The instinct is that no alignment always means worse. That is true for a natural protein with a deep MSA, and false for a designed one.
A de novo designed binder has no evolutionary relatives by construction. There is no real alignment to build. Forcing one, by aligning against distant natural homologs that share nothing meaningful with the design, does not add signal; it adds noise, and it can actively pull the prediction off. On a designed or orphan sequence, folding from the single sequence is the honest input, not a shortcut. This is exactly the regime ESMFold was built for, and it is why single-sequence folding has become a standard self-consistency check in design pipelines.
The rule of thumb is short. Designed, orphan, or highly diverged sequence with no meaningful alignment: single-sequence ESMFold is a legitimate call. Natural protein with a deep, high-quality MSA: use the alignment, AlphaFold 2 or ColabFold will be more accurate, and skipping the MSA there leaves accuracy on the table for no reason.
The trap that costs wet-lab money
Here is the point that matters most, and the one most often missed. A confident fold is not a binding prediction.
A high pLDDT from ESMFold tells you the model is confident about the structure of that single chain. It says nothing about whether your designed binder engages its target. Reading a clean monomer fold as evidence of binding is the most expensive mistake in this workflow, because it sends designs to the bench that were never going to work, and wet-lab time is the scarce resource, not GPU time.
Binding is a separate question that a monomer folder cannot answer. It needs a prediction of the complex, a cofold of binder and target together, with interface scoring, and past that, an actual assay. ESMFold is an excellent way to check that a designed sequence folds into the shape you intended. It is not, and was never meant to be, a filter for whether that shape binds anything.
Quick reference
| ESMFold | ColabFold | AlphaFold 2 | |
|---|---|---|---|
| Input | single sequence, no MSA | AF2 weights, fast MMseqs2 MSA | AF2 weights, full MSA |
| Reads evolution from | the language-model weights | the alignment | the alignment |
| Speed | fastest, seconds per model | fast | slowest |
| Best for | large batches, designed and orphan sequences | the set that matters, with reasonable homology | hardest, highest-stakes natural proteins |
| Watch out | pLDDT is not a binding signal; weaker where an MSA would have helped | MSA depth varies by database | wall-clock and cost per prediction |
Practical recommendations
A workflow that holds up across most design and analysis tasks:
- Triage large designed pools with ESMFold. Single-sequence, seconds per model, and it is the correct input for sequences with no natural homologs. Cut the pool here.
- Take survivors to a cofold with the target. The binding question needs the complex, not a monomer fold. Spend the expensive complex prediction where it changes a decision, with a full MSA on the natural target chain.
- Do not read pLDDT as binding. A confident monomer fold means the shape is plausible, nothing more. Take binding to a cofold and then to the bench.
- Match the input to the sequence. Designed or orphan, single-sequence is honest. Natural with a deep MSA, use the alignment.
You can run ESMFold on tools.ranomics.com with no install, and switch to the AlphaFold 2 tool when you want the full-MSA fold on the same target. The two share the same hosted environment, so triaging with one and confirming with the other is a single workflow.
Summary
ESMFold, AlphaFold 2, and ColabFold are not an accuracy ladder. AlphaFold and ColabFold read an evolutionary alignment; ESMFold folds from the single sequence because it learned that signal into a language model. Use ESMFold to triage large pools and to fold designed and orphan sequences that have no meaningful MSA. Take the survivors to a cofold of the complex, the prediction that actually speaks to binding, where a full MSA on the target earns its wall-clock. And whichever one you run, remember that a confident fold answers a question about shape, not about binding.
Related Ranomics services
- ESMFold technology page: how the ESM-2 single-sequence folder sits in our triage pipeline.
- ColabFold vs AlphaFold 2: the companion piece on the two MSA routes to a fold.
- Binder Pilot: take your top folded designs to a ranked, wet-lab-validated hit list, where binding is actually measured.