Transform PAE matrix to LIS scores. Args: pae_matrix: Input PAE matrix pae_cutoff: PAE threshold for transformation Ret

Implementation of TriangleMultiplication used in AF2 and AF-M<2.3.

dynamically update template features

Generate redesigned sequences using AbMPNN for a given PDB structure. Args: trajectory_pdb: Path to the input PDB file. targ

Add features to distinguish between chains. Args: all_chain_features: A dictionary which maps chain_id to a dictionary of features for ea

Add beta sheet secondary structure loss function for CDR regions. Implements a sophisticated loss function that promotes beta sheet formation

Add beta strand secondary structure loss function. Implements a loss function that promotes beta strand secondary structure in specified

add missing CB atoms based on N,CA,C

Add helical secondary structure loss function. Implements a loss function that promotes alpha-helical secondary structure in specified re

Add interface predicted Template Modeling score loss function. Implements an interface pTM loss to optimize the predicted confidence of t

Add radius of gyration loss function to control protein compactness. Implements a radius of gyration constraint to encourage compact protein

Add N- and C-terminus distance constraint loss function. Implements a loss function that constrains the distance between the N- and C-ter

Perform average product correct, used for contact prediction.

Applies dropout to a tensor.

Multiply the inverse of a rotation matrix by a vector.

Applies inverse Rot3Array to point.

compute autoregressive mask, given order of positions

Aggregate all metrics into comprehensive evaluation dict (floats rounded to 4 decimals). Returns: Dict[str, Any]: Confidence, interf

N-chain pDockQ: - Builds a full distance matrix across all chains - Uses only cross-chain contacts (< contact_cutoff Å) - Avera

Calculate structural clash score for protein structure validation. Analyzes a protein structure to identify and count atomic clashes based on

Calculate shape complementarity between loop residues on chain A and all residues on chain B, with automatic loop detection. Args:

Check if a design seed has already been processed in this run. Searches through all structure directories (trajectories, redesign_can

Evaluate termination conditions for the design optimization loop. Checks multiple termination criteria including maximum accepted des

Clear GPU and system memory caches for optimal performance. Performs comprehensive memory cleanup including Python garbage collection, Py

Compute CDR residue positions from framework and CDR length specifications. Calculates the absolute residue positions for all CDRs based on f

Compute CDR contacts with target hotspot residues (5.3Å threshold, ≥3 contacts required). Args: external_relaxed_pdb: Relaxed comple

Compute docking quality metrics. pDockQ2 and LIS now come exclusively from the upstream ``ipsae`` tool (single scalar each), avoiding th

Compute MLM pseudolikelihood by masking each residue position once. For each residue, mask it and score log p(residue | all other po

Convert Chai-1 output CIF files to PDB format. Chai-1 outputs structures in mmCIF format, but many downstream tools expect PDB format. T

Reshapes and modifies monomer features for multimer models.

Create complete directory structure for a new Germinal design run. Initializes the full directory hierarchy required for a Germinal d

Expand extra_msa into 1hot and concat with other extra msa features. We do this as late as possible as the one_hot extra msa can be very large.

Expand extra_msa into 1hot and concat with other extra msa features. We do this as late as possible as the one_hot extra msa can be very large. Ar

Create input JSON data for AlphaFold3 inference. Args: binder_seq (str): Amino acid sequence of the binder protein. target_s

Create input JSON data for Protenix inference. Protenix uses a different JSON schema from AF3: sequences are specified as 'proteinChain'

Create combined PDB structure from separate binder and target files. Combines a binder PDB and target PDB into a single complex structure,

crop features to specified [pos]itions

three stage design (logits→soft→hard)

optimize argmax(logits)

Evaluate metrics against quality filters (operators: <, <=, >, >=, ==, =). Args: filter_set: {metric_name: {"value": threshold, "ope

given msa (N,L,20) expand to (1+N*copies,L*copies,22) if block_diag else (N,L*copies,22)

input: 3 coords (a,b,c), (L)ength, (A)ngle, and (D)ihedral output: 4th coord

Extract predicted structure and confidence scores from Protenix output. Converts the best/worst CIF output to PDB format and extracts confid

Extract predicted structure and confidence scores from AF3 output. Processes AF3 results by converting the output CIF file to PDB format,

train model

Args: seq_logits (torch.Tensor): Input sequence (seq_length, 20). Returns: ce_loss (nll) log_lik

Construct Rigid3Array from homogeneous 4x4 array.

Generate predicted affines for a single chain. Jumper et al. (2021) Suppl. Alg. 20 "StructureModule" This is the main part of the structure modu

Generate predicted Rigid's for a single chain. This is the main part of the iterative fold head - it iteratively applies folding to produce a set

Generate Multiple Sequence Alignments (MSAs) for protein chains. Creates MSAs for each protein chain in the input data using the specified

Generate a unique hash identifier for temporary directories. Creates a collision-resistant hash using timestamp, process ID, and random numb

Conduct binder hallucination with ColabDesign AF2 model. Args: design_name: Unique name for this design trajectory run_setti

Calculate antibody sequence pseudo-log-likelihood using AbLang (MLM scoring). Each residue is masked once and scored against the full bidire

Generate AbMPNN redesigned sequences for a given trajectory. Args: trajectory_pdb_af: Path to the trajectory PDB file from AF2/Colab

get inputs from alphafold model

Get the number of residues in a specific chain Parameters: - pose: PyRosetta pose - chain_id: Chain identifier (default 'A') Re

Returns atom indices needed to compute chi angles for all residue types. Returns: A tensor of shape [residue_types=21, chis=4, atoms=4]. The re

Returns atom indices needed to compute chi angles for all residue types. Returns: A tensor of shape [residue_types=21, chis=4, atoms=4]. The re

Identify mutations outside CDR regions (format: 'A123B'). Args: trajectory_sequence: Final designed sequence framework_seque

Unmasked CE gradient (fast, one forward pass). NOTE: NOT Salazar-style masked PLL. The embedding hook supplies the actual (soft) toke

Masked PLL gradient (Salazar-style), chunked to bound peak GPU memory. Processes AA positions in chunks of chunk_size. Each chunk runs a forw

Calculate antibody sequence log-likelihood using IgLM language model. Attribution: Shuai, R. W., Ruffolo, J. A., & Gray, J. J. (2023). IgLM:

Get Initializer for weights and scale to multiply activations by.