HelixFold-S1: MSA-free & Contact-Guided Models

• The paper introduces HelixFold-Single (2022), an MSA-free protein structure predictor that utilizes a large protein language model to replace traditional evolutionary inputs.
• The 2025 model adapts a HelixFold3 variant by integrating explicit intermolecular contact prediction with guided sampling for biomolecular complex structure prediction.
• Both models, built on the HelixFold ecosystem, offer significant speed improvements over MSA-based methods while addressing distinct structural prediction challenges.

HelixFold-S1 is a name used for two distinct models within the HelixFold line. In the 2022 literature and PaddleHelix ecosystem, “HelixFold-S1” commonly denotes HelixFold-Single, an end-to-end, MSA-free protein structure predictor that replaces AlphaFold2’s MSA stack with a large protein LLM while retaining most of AlphaFold2’s geometric machinery (Fang et al., 2022). In the 2025 literature, HelixFold-S1 denotes a sampling-optimized variant of HelixFold3 for biomolecular complex structure prediction, introducing explicit intermolecular contact prediction and contact-guided test-time sampling within an AlphaFold3-like architecture (Liu et al., 12 Jul 2025). The shared label therefore refers either to a single-sequence monomer-folding model or to a contact-guided complex-sampling model, depending on context.

1. Nomenclature and lineage

The ambiguity of the term reflects the evolution of the HelixFold family. “HelixFold” originally referred to an efficient PaddlePaddle implementation of AlphaFold2, with systems-level optimization for training and inference speed and memory reduction, while maintaining accuracy on CASP14 and CAMEO comparable to AlphaFold2 (Wang et al., 2022). On top of that base, HelixFold-Single was introduced as a single-sequence predictor and was often abbreviated as “HelixFold-S1” in code and deployments (Fang et al., 2022). Later, a different model, also named HelixFold-S1, was introduced as a sampling-optimized HelixFold3 variant for complex prediction (Liu et al., 12 Jul 2025).

Usage of “HelixFold-S1”	Base model	Primary task
HelixFold-Single	HelixFold / AlphaFold2-style geometry	Single-sequence protein structure prediction
HelixFold-S1	HelixFold3 / AlphaFold3-like architecture	Biomolecular complex structure prediction via contact-guided sampling

The 2022 and 2025 usages are therefore not minor revisions of a single model family member. They occupy different positions in the HelixFold trajectory: the earlier system addresses the removal of MSA dependence for protein folding, whereas the later system addresses inference-time exploration of docking modes for complexes when co-evolutionary signals are weak or absent (Fang et al., 2022, Liu et al., 12 Jul 2025).

A plausible implication is that references to “HelixFold-S1” in software, deployments, or secondary discussions require disambiguation by year, task, or architectural context. Without that disambiguation, the label conflates two substantially different modeling objectives.

2. HelixFold-Single as the 2022 meaning of HelixFold-S1

HelixFold-Single is an end-to-end, MSA-free protein structure predictor designed to infer 3D structures directly from a single amino-acid sequence, without computing MSAs or using templates (Fang et al., 2022). Its stated motivation is that MSA search is time-consuming, often taking dozens of minutes per protein, and that many downstream workloads require large numbers of predictions, making MSA-based inference a bottleneck (Fang et al., 2022).

The model substitutes AlphaFold2’s explicit evolutionary input pathway with a large protein LLM (PLM). The PLM is trained on primary sequences by masked language modeling and is used as an alternative mechanism for learning co-evolutionary information (Fang et al., 2022). The resulting architecture keeps the essential geometric learning stack of AlphaFold2—specifically a modified Evoformer and Structure Module—but removes the MSA stack and template search at inference (Fang et al., 2022).

The pipeline has three major components. First, the PLM Base takes the primary sequence $\mathbf{x} = (x_1,\dots,x_L)$ and produces per-residue embeddings and per-layer multi-head attention maps (Fang et al., 2022). Second, an Adaptor linearly projects those outputs into the dimensions required by the folding trunk, producing an initial single representation and an initial pair representation: $$\begin{aligned} \tilde{\mathbf{x}}^{(0)} &= \text{Linear}\bigl(\mathbf{x}^{(n_{\text{PLM}})}\bigr), \ \tilde{\mathbf{z}}^{(0)} &= \text{Linear}\Bigl([\mathbf{z}^{(1)}, \mathbf{z}^{(2)}, \dots, \mathbf{z}^{(n_{\text{PLM}})}]\Bigr). \end{aligned}$$ Third, Geometric Modeling applies a revised Evoformer and a Structure Module, with recycling, to predict atomic coordinates (Fang et al., 2022).

Within the PLM, the paper uses a DeBERTa-based Transformer with disentangled attention and switches from Post-Norm to Pre-Norm for stability (Fang et al., 2022). For the main model, the PLM-1B configuration has approximately 1.09B parameters, $n_{\text{PLM}} = 20$, hidden size $d_{\text{PLM}} = 2048$, intermediate size 8192, and $h_{\text{PLM}} = 16$ attention heads (Fang et al., 2022). The Evoformer has approximately 87M parameters, and the Structure Module approximately 1.7M parameters, for a total of about 1.18B parameters plus adaptor layers (Fang et al., 2022).

A central architectural detail is that the pair representation is derived from attention maps across all PLM layers, while the single representation comes from the final-layer residue embeddings (Fang et al., 2022). This makes the model an explicit attempt to relocate MSA-like statistical structure into the PLM, then reuse AlphaFold2-level geometric reasoning downstream.

3. Architecture, training regime, and empirical properties of HelixFold-Single

HelixFold-Single’s revised Evoformer closely follows AlphaFold2 but removes column-wise gated self-attention, since there is no MSA dimension to attend over in the MSA-free setting (Fang et al., 2022). The input dimensions after the adaptor are $$\tilde{\mathbf{x}}^{(0)} \in \mathbb{R}^{L \times 512}$$ and $$\tilde{\mathbf{z}}^{(0)} \in \mathbb{R}^{L \times L \times 64}$$ (Fang et al., 2022). The Structure Module follows AlphaFold2 and uses Invariant Point Attention (IPA) and SE(3)-equivariant operations to infer residue frames, torsions, and atomic coordinates (Fang et al., 2022).

Training proceeds in two stages (Fang et al., 2022). In Stage 1, the PLM is pre-trained on UniRef30 (2021-03), comprising about 260–300 million sequences, using masked language modeling with 15% residue masking (Fang et al., 2022). In Stage 2, the PLM Base and Geometric Modeling are jointly optimized with structural supervision on about 1.1M structures, combining filtered RCSB PDB structures with two distillation datasets derived from AlphaFold2 and the AlphaFold Protein Structure Database (Fang et al., 2022). The structure prediction objective follows AlphaFold2, including FAPE and auxiliary losses such as distogram, angle/torsion, violation, and pLDDT-related terms, although the paper does not rederive all formulas (Fang et al., 2022).

The main reported runtime result is that inference avoids external database search entirely. On a single NVIDIA A100 (40G), median inference times for HelixFold-Single are reported as 1.5 s for proteins of length 1–100 and 100–200, 2.1 s for 200–400, 6.2 s for 400–800, and 37.5 s for $>800$, compared with AlphaFold2 total pipeline times of 766.1 s, 795.8 s, 908.3 s, 1125.2 s, and 1611.2 s, respectively (Fang et al., 2022). The paper states that for sequences shorter than 100 residues, HelixFold-Single is roughly 1000× faster than the full AlphaFold2 pipeline, while for $>800$ residues it remains approximately 40–50× faster (Fang et al., 2022).

On CASP14, HelixFold-Single significantly outperforms the MSA-free baselines discussed in the paper but still lags MSA-based AlphaFold2 overall, especially on challenging FM targets (Fang et al., 2022). On CAMEO, its average accuracy is described as comparable to AlphaFold2 with MSA input and better than RoseTTAFold with MSA input (Fang et al., 2022). When stratified by MSA depth, the reported behavior is that HelixFold-Single improves as MSA depth increases, and for targets with MSA depth $>1000$ its TM-scores are comparable to MSA-based methods (Fang et al., 2022).

The paper also reports negative correlations between PLM perplexity and both MSA depth and downstream TM-score (Fang et al., 2022). This suggests that PLM modeling quality functions as a proxy for how much family-level information the model has internalized. The authors further show that scaling the PLM from 100M to 1B parameters lowers perplexity and improves long-range contact prediction measured by top $$\begin{aligned} \tilde{\mathbf{x}}^{(0)} &= \text{Linear}\bigl(\mathbf{x}^{(n_{\text{PLM}})}\bigr), \ \tilde{\mathbf{z}}^{(0)} &= \text{Linear}\Bigl([\mathbf{z}^{(1)}, \mathbf{z}^{(2)}, \dots, \mathbf{z}^{(n_{\text{PLM}})}]\Bigr). \end{aligned}$$0 long-range contact precision $$\begin{aligned} \tilde{\mathbf{x}}^{(0)} &= \text{Linear}\bigl(\mathbf{x}^{(n_{\text{PLM}})}\bigr), \ \tilde{\mathbf{z}}^{(0)} &= \text{Linear}\Bigl([\mathbf{z}^{(1)}, \mathbf{z}^{(2)}, \dots, \mathbf{z}^{(n_{\text{PLM}})}]\Bigr). \end{aligned}$$1 (Fang et al., 2022).

4. HelixFold-S1 as the 2025 contact-guided sampling model

The 2025 HelixFold-S1 addresses a different problem: biomolecular complex structure prediction across interaction types including protein-antibody, protein-protein, protein-ligand, protein-RNA, and protein-DNA (Liu et al., 12 Jul 2025). Its stated motivation is that co-evolutionary signals are often weak or absent for these tasks—especially for antibody–antigen, protein–ligand, and many nucleic-acid complexes—and that pure large-scale sampling can yield diminishing returns because many extra samples remain minor variants of the same binding mode (Liu et al., 12 Jul 2025).

HelixFold-S1 is defined as a sampling-optimized variant of HelixFold3 that explicitly predicts intermolecular contacts and uses them to guide test-time sampling (Liu et al., 12 Jul 2025). Conceptually, it converts “aimless” sampling into targeted, contact-driven exploration of docking modes while remaining within an AlphaFold3-like architecture (Liu et al., 12 Jul 2025).

The model introduces two contact-centric additions (Liu et al., 12 Jul 2025). The Contact Prediction Module takes pair representations from HF3’s Pairformer and outputs contact probabilities between tokens from different chains. The contact definition is binary: tokens $$\begin{aligned} \tilde{\mathbf{x}}^{(0)} &= \text{Linear}\bigl(\mathbf{x}^{(n_{\text{PLM}})}\bigr), \ \tilde{\mathbf{z}}^{(0)} &= \text{Linear}\Bigl([\mathbf{z}^{(1)}, \mathbf{z}^{(2)}, \dots, \mathbf{z}^{(n_{\text{PLM}})}]\Bigr). \end{aligned}$$2 and $$\begin{aligned} \tilde{\mathbf{x}}^{(0)} &= \text{Linear}\bigl(\mathbf{x}^{(n_{\text{PLM}})}\bigr), \ \tilde{\mathbf{z}}^{(0)} &= \text{Linear}\Bigl([\mathbf{z}^{(1)}, \mathbf{z}^{(2)}, \dots, \mathbf{z}^{(n_{\text{PLM}})}]\Bigr). \end{aligned}$$3 are in contact if any atom pair across the two tokens is within 5 Å (Liu et al., 12 Jul 2025). The Contact Conditioning Module accepts a binary matrix $$\begin{aligned} \tilde{\mathbf{x}}^{(0)} &= \text{Linear}\bigl(\mathbf{x}^{(n_{\text{PLM}})}\bigr), \ \tilde{\mathbf{z}}^{(0)} &= \text{Linear}\Bigl([\mathbf{z}^{(1)}, \mathbf{z}^{(2)}, \dots, \mathbf{z}^{(n_{\text{PLM}})}]\Bigr). \end{aligned}$$4, with $$\begin{aligned} \tilde{\mathbf{x}}^{(0)} &= \text{Linear}\bigl(\mathbf{x}^{(n_{\text{PLM}})}\bigr), \ \tilde{\mathbf{z}}^{(0)} &= \text{Linear}\Bigl([\mathbf{z}^{(1)}, \mathbf{z}^{(2)}, \dots, \mathbf{z}^{(n_{\text{PLM}})}]\Bigr). \end{aligned}$$5, projects it into pair-feature space, and fuses it into the pair representation at the Input Embedder stage so that the diffusion model is conditioned on specified contacts (Liu et al., 12 Jul 2025).

The 2025 paper describes a three-stage inference pipeline (Liu et al., 12 Jul 2025). In the first stage, the model is run in contact-prediction mode five times by default and the resulting contact probability matrices are averaged, with intra-chain entries masked out (Liu et al., 12 Jul 2025). In the second stage, inter-chain token pairs are sorted by predicted contact probability and iteratively selected as constraints, using redundant contact pruning so that contacts already realized in previous predictions are not repeatedly used for new samples (Liu et al., 12 Jul 2025). When the maximum remaining contact probability falls below

$$\begin{aligned} \tilde{\mathbf{x}}^{(0)} &= \text{Linear}\bigl(\mathbf{x}^{(n_{\text{PLM}})}\bigr), \ \tilde{\mathbf{z}}^{(0)} &= \text{Linear}\Bigl([\mathbf{z}^{(1)}, \mathbf{z}^{(2)}, \dots, \mathbf{z}^{(n_{\text{PLM}})}]\Bigr). \end{aligned}$$6

where $$\begin{aligned} \tilde{\mathbf{x}}^{(0)} &= \text{Linear}\bigl(\mathbf{x}^{(n_{\text{PLM}})}\bigr), \ \tilde{\mathbf{z}}^{(0)} &= \text{Linear}\Bigl([\mathbf{z}^{(1)}, \mathbf{z}^{(2)}, \dots, \mathbf{z}^{(n_{\text{PLM}})}]\Bigr). \end{aligned}$$7, the system switches from exploring new contacts to enriched sampling from previously realized contact sets until the sample budget is exhausted (Liu et al., 12 Jul 2025).

In the third stage, the selected contact or contact subset is converted into a binary contact matrix and passed through the Contact Conditioning Module to generate a full 3D structure (Liu et al., 12 Jul 2025). Reported experiments use 10 recycles and 200 diffusion steps per sample (Liu et al., 12 Jul 2025). Candidate structures are ranked by

$$\begin{aligned} \tilde{\mathbf{x}}^{(0)} &= \text{Linear}\bigl(\mathbf{x}^{(n_{\text{PLM}})}\bigr), \ \tilde{\mathbf{z}}^{(0)} &= \text{Linear}\Bigl([\mathbf{z}^{(1)}, \mathbf{z}^{(2)}, \dots, \mathbf{z}^{(n_{\text{PLM}})}]\Bigr). \end{aligned}$$8

and the highest-confidence structure is returned as the final prediction (Liu et al., 12 Jul 2025).

The paper emphasizes the exploration–exploitation structure of the sampler: early high-probability contact proposals broaden the search space, while later enriched sampling refines already promising interface patterns (Liu et al., 12 Jul 2025). A plausible implication is that the contact map acts as an intermediate, lower-dimensional interface representation that makes diversity controllable rather than incidental.

5. Training, evaluation, and observed behavior of the 2025 HelixFold-S1

The 2025 HelixFold-S1 retains HelixFold3’s multi-term loss and adds an inter-chain binary contact loss (Liu et al., 12 Jul 2025). The total loss is given as

$$\begin{aligned} \tilde{\mathbf{x}}^{(0)} &= \text{Linear}\bigl(\mathbf{x}^{(n_{\text{PLM}})}\bigr), \ \tilde{\mathbf{z}}^{(0)} &= \text{Linear}\Bigl([\mathbf{z}^{(1)}, \mathbf{z}^{(2)}, \dots, \mathbf{z}^{(n_{\text{PLM}})}]\Bigr). \end{aligned}$$9

with $n_{\text{PLM}} = 20$0, $n_{\text{PLM}} = 20$1, $n_{\text{PLM}} = 20$2, and $n_{\text{PLM}} = 20$3 only for contact-prediction samples and 0 for contact-conditioned structure-prediction samples (Liu et al., 12 Jul 2025). The contact loss is binary cross-entropy over inter-chain token pairs (Liu et al., 12 Jul 2025).

Training is organized into three fine-tuning stages (Liu et al., 12 Jul 2025). Stage 1 trains only contact-conditioned structure prediction, with 70% probability of using 1–10 true inter-chain contacts as constraints and 30% probability of using none; the crop size is 384, the training length 10,000 steps, and the optimizer is Adam with $n_{\text{PLM}} = 20$4, $n_{\text{PLM}} = 20$5, $n_{\text{PLM}} = 20$6, learning rate $n_{\text{PLM}} = 20$7, and batch size 128 (Liu et al., 12 Jul 2025). Stage 2 jointly trains contact prediction and contact-conditioned structure prediction for 20,000 steps at crop size 384, and Stage 3 repeats that setup with crop size 640 for 3,000 steps (Liu et al., 12 Jul 2025). The new contact modules are randomly initialized, while all other parameters are copied from HF3 (Liu et al., 12 Jul 2025).

Evaluation sets are drawn from PDB and SAbDab in 2022–2024, filtered to be non-redundant and test-only relative to HF3’s training set (Liu et al., 12 Jul 2025). The benchmark includes 221 protein–antibody complexes from SAbDab, 199 protein–protein, 194 protein–ligand, 94 protein–RNA, and 252 protein–DNA complexes (Liu et al., 12 Jul 2025). Reported metrics are DockQ for protein–protein and antibody–antigen, interface lDDT (iLDDT) for protein–nucleic-acid complexes, and symmetry-corrected ligand heavy-atom RMSD after pocket alignment for protein–ligand complexes, with the main ligand metric being the fraction of predictions with ligand RMSD $n_{\text{PLM}} = 20$8 Å (Liu et al., 12 Jul 2025).

The baseline set includes unguided HF3, HF3 w/ AFsample, Chai-1, Boltz-2, and AFsample itself for sampling-strategy comparison (Liu et al., 12 Jul 2025). Across complex types, both HF3 and HF-S1 improve as the number of samples increases, but under oracle selection the performance gap between HF-S1 and HF3 widens with more sampling, indicating that HF-S1 discovers better structures within its ensemble because its search is more diverse and better targeted (Liu et al., 12 Jul 2025). The paper highlights especially large gains on protein–antibody, protein–ligand, and moderately difficult nucleic-acid complexes (Liu et al., 12 Jul 2025).

For a 2024-only protein–antibody subset of 95 complexes using 80 samples, with success defined as DockQ $n_{\text{PLM}} = 20$9, the paper reports that HF3 and Boltz-2 are strong baselines but show limited improvement with additional samples, Chai-1 shows virtually no gain from additional sampling, HF3 w/ AFsample improves with more samples, and HF-S1 shows substantially faster and larger performance improvements than all baselines (Liu et al., 12 Jul 2025). In contact-focused analyses, the predicted contact matrix of HF-S1 has substantially higher AUPRC than a posterior contact matrix derived from HF3 samples, protein–antibody complexes have the lowest AUPRC, and protein–ligand complexes the highest (Liu et al., 12 Jul 2025). The contact satisfaction rate is reported as $d_{\text{PLM}} = 2048$0 across complex types, indicating that conditioning usually yields structures satisfying the imposed contacts (Liu et al., 12 Jul 2025).

The paper also analyzes target-level contact probability $d_{\text{PLM}} = 2048$1 as an empirical difficulty indicator (Liu et al., 12 Jul 2025). Lower $d_{\text{PLM}} = 2048$2 is associated with lower single-sample accuracy in DockQ, iLDDT, and ligand RMSD, whereas higher $d_{\text{PLM}} = 2048$3 is associated with stronger single-sample predictions (Liu et al., 12 Jul 2025). Intermediate $d_{\text{PLM}} = 2048$4 values show the largest gains under oracle selection, suggesting that these are the cases where contact-guided sampling is most beneficial (Liu et al., 12 Jul 2025). The authors stop short of implementing adaptive sampling allocation, but the reported analysis supports that interpretation (Liu et al., 12 Jul 2025).

6. Relation to the broader HelixFold ecosystem and interpretive cautions

The broader HelixFold program begins with HelixFold, a PaddlePaddle implementation of AlphaFold2 optimized by operator fusion, tensor fusion, hybrid parallelism, Recompute, BFloat16, and in-place memory read/write (Wang et al., 2022). HelixFold reports end-to-end training in 7.5 days on 128 A100 GPUs and 5.3 days with hybrid parallelism, compared with about 11 days for the original AlphaFold2 and OpenFold, while maintaining near-parity on CASP14 and CAMEO (Wang et al., 2022). That work does not define HelixFold-S1 explicitly, but it provides the implementation substrate on which later variants were built (Wang et al., 2022).

Against that backdrop, the two meanings of HelixFold-S1 exemplify two different research directions. The 2022 HelixFold-Single asks whether the MSA stack can be replaced by a large PLM while retaining AlphaFold2’s geometric trunk (Fang et al., 2022). The 2025 HelixFold-S1 asks whether complex prediction can be improved at inference time by replacing indiscriminate sampling with contact-guided exploration and contact-conditioned generation (Liu et al., 12 Jul 2025). The former is primarily an input-representation substitution for monomer folding; the latter is primarily an inference-time search strategy for complexes.

Several common misconceptions follow from the name collision. One is to treat HelixFold-S1 as a single, continuously upgraded model. The source material does not support that interpretation: the 2022 and 2025 systems differ in task, architecture, and training setup (Fang et al., 2022, Liu et al., 12 Jul 2025). Another is to assume that the 2025 S1 is simply a continuation of the “Single” naming convention. The paper explicitly presents it instead as a sampling-optimized variant of HelixFold3, and the “S1” designation is not expanded in the text; the authors only infer from context that it denotes a sampling-oriented variant (Liu et al., 12 Jul 2025).

A plausible implication is that “HelixFold-S1” should be read as a context-dependent identifier rather than a stable taxonomic label. In bibliographic usage, the safest interpretation is therefore to pair the name with its arXiv record, publication year, and task domain: HelixFold-Single for MSA-free protein structure prediction (Fang et al., 2022), or HelixFold-S1 for contact-guided biomolecular interaction structure prediction (Liu et al., 12 Jul 2025).