RapidDock

Screening one million compounds against the roughly 20,000 proteins of the human proteome is 2 × 10^10 docking runs. At the 35 seconds per complex that DiffDock-L takes on a single GPU, a 512-GPU cluster would need about twenty years; with a docking-while-folding method like AlphaFold 3 the figure is closer to two hundred. The obstacle to knowing what else a drug candidate binds — the off-target interactions that surface as toxicity late in trials — is arithmetic, not accuracy.

RapidDock, released as a preprint in October 2024 by Ingenix.ai with collaborators at the University of Warsaw, Warsaw University of Technology, IDEAS NCBR, IMPAN and Cardinal Stefan Wyszyński University, attacks that arithmetic by changing the shape of the computation rather than tuning it. There is no diffusion trajectory to integrate and no structure module to iterate: one transformer encoder reads the protein and the ligand together and emits, in a single forward pass, the complete matrix of pairwise distances in the docked complex — ligand atom to ligand atom, and ligand atom to amino acid. Cartesian coordinates are recovered from that matrix afterwards by L-BFGS optimization, at a cost comparable to the forward pass itself.

Docking is blind: no binding pocket is supplied, and the protein may be an unbound or computationally folded structure, so the method applies to targets with no ligand-bound crystal structure. Average inference time is 0.04 seconds on one GPU.

#Key Features

• Single-pass inference: The whole prediction is one non-autoregressive transformer evaluation — 0.03 seconds per complex on PoseBusters and 0.05 on DockGen — rather than a sampled trajectory that must be repeated and ranked.
• Distances as attention bias: Protein and ligand geometry enter as learnable embeddings of discretized inter-atomic distances added to the attention matrices, which makes the model invariant to translation and rotation without an equivariant architecture.
• Folding pre-training: The network is first trained to predict amino-acid distances on ~440k AlphaFold-2 SwissProt structures with 97% of input distances masked; removing that stage costs 5.7 points of PoseBusters accuracy.
• Symmetry-invariant loss: For molecules such as benzene, where one pose maps to several equivalent atom orderings, the loss is evaluated over all permutations and only the best-matching one is backpropagated.
• Robust to unbound inputs: Given AlphaFold-2 apostructures instead of the bound proteins, success rates fall only to 51.7% on PoseBusters and 42.3% on DockGen.

#Technical Details

The backbone is a 16-layer Mistral-style transformer with 4 attention heads and a hidden size of 512 — about 60 million parameters, chosen deep-and-narrow over shallower alternatives of equal size — run with a full rather than causal attention mask. Inputs are 56 atom tokens and 26 amino-acid tokens; atoms additionally carry embedded Gasteiger partial charges, and residues carry per-position embeddings from the 650M-parameter ESM-2. The protein's structure is its Cα distance matrix; the ligand's is a rigid-part distance matrix averaged over 96 generated conformers, with entries whose standard deviation exceeds 0.3 Å marked unknown to separate rigid fragments from rotatable ones. Training ran about three days on eight A100 GPUs: 48 hours of folding pre-training, then 16 hours of docking fine-tuning on ~30k complexes from PDBBind and BindingMOAD, with apostructures substituted for about 30% of training proteins. On the held-out benchmarks RapidDock reaches 52.1% of poses under 2 Å RMSD on PoseBusters (median 1.90 Å) and 44.0% on DockGen (median 2.83 Å), against 40.8% and 22.6% for DiffDock-L; the ESM-2 embeddings are the single largest contributor, worth 9.5 points on PoseBusters.

#Applications

The intended use is screening at a scale that per-complex runtimes have previously ruled out: the authors dock twelve well-studied drugs and toxins against 19,222 human proteins from the Human Protein Atlas, averaging 74 seconds per ligand on eight A100s. That makes exhaustive off-target and polypharmacology profiling a routine step rather than a project, and it suits hit triage in virtual screening campaigns where a fast, pocket-free first pass can narrow a library before slower and more accurate methods are applied.

#Impact

RapidDock is presented as the first fully transformer-based model for blind molecular docking, and it is a useful demonstration that a distance-prediction formulation with learned distance biases can match generative docking accuracy at a fraction of the cost. Its limits are real: AlphaFold 3 remains substantially more accurate on PoseBusters at 76.9%, the model emits no confidence score to say which of its poses to trust, and it requires a protein structure as input. The work was submitted to ICLR 2025 and remains an unrefereed preprint; no code or weights have been published. Readers should also note an unrelated 2024 model named RAPiDock, a diffusion generative method for protein–peptide docking from a different group — the two share no authors, architecture, or artifacts.