Skip to content

Latest commit

 

History

History
177 lines (130 loc) · 5.81 KB

README.md

File metadata and controls

177 lines (130 loc) · 5.81 KB
Raw

Humatch

Fast, gene-specific joint humanisation of antibody heavy and light chains.

                        @
     )  __QQ    -->    /||\
    (__(_)_">           /\

Install

Follow the steps below to download the code and the necessary packages. Requires Python 3.9.

# clone the repo
git clone https://github.com/oxpig/Humatch.git
cd Humatch/

# create your virtual env e.g.
python3 -m venv .humatch_venv
source .humatch_venv/bin/activate

# install
pip install .

ANARCI is required for aligning and padding sequences. We recommend installing ANARCI from github.com/oxpig/ANARCI.

If you are having issues installing, try upgrading pip: pip install --upgrade pip

Specific python versions can be used to initiate the environment using e.g. /usr/bin/python3.9 -m venv .humatch_venv

CNN weights and germline likeness lookup arrays are automatically downloaded from zenodo.org/records/13764770 when Humatch is first run.

$${\color{red}If \space you \space have \space issues \space with \space the \space auto \space downloads}$$ then the 3 weights (.h5) files and 24 germline likeness lookup arrays (.npy) files can be manually downloaded and saved in Humatch/Humatch/trained_models and Humatch/Humatch/germline_likeness_lookup_arrays respectively. Once these files are downloaded and saved in the right folders, please rerun the pip install . command to add these files to Humatch's package data.

Humanness classification

Humatch can be used to obtain heavy, light, and paired predictions for a given VH/VL pair. As Humatch was trained on complete VH/VL sequences, only complete sequences should be used as input. An example notebook is provided and predictions can also be obtained from the command line e.g.

Humatch-classify
    -H EVQLVESGGG...VSS
    -L DIVMTQGALP...EIK
    -s

Output:

hv:     hv3
lv:     kv2
CNN_H:  0.000
CNN_L:  0.000
CNN_P:  0.000

Predictions for all heavy and light V-genes are returned by ommitting the -s summary flag, otherwise only the top-scoring human v-genes are returned. Classification can be run for individual heavy/light chains - in this instance, only the heavy/light CNN score will be returned.

For high throughput screening, Humatch can be run on a csv file of antibody sequences e.g.

Humatch-classify
    -i data/example.csv
    --vh_col heavy
    --vl_col light

Output:

VH VL hv1 ... hv7 lv1 ... lv10 kv1 ... kv7 CNN_P
QVQ...VSS EIV...IK- 0.999 ... 0.000 0.000 ... 0.000 0.000 ... 0.000 0.998
... ... ... ... ... ... ... ... ... ... ... ...

The output csv will contain Humatch's predictions alongside the aligned, padded VH/VL sequences of the columns specified. Output paths can be specified with the -o argument.

Humanisation

Humatch is primarily designed to offer experimental-like humanisation in seconds. Like humanness classification, an example notebook is provided in addition to the command line interface e.g.

Humatch-humanise
    -H QVNLLQSGAA...VSA
    -L DTVLTQSPAL...EIK
    -v

Output:

Humanised sequences:
        EVKLVESGGG...VSS
        DTVLTQSPAL...EIK
        Edit:   24
        HV:     hv1
        LV:     kv3
        CNN_H:  0.958
        CNN_L:  0.999
        CNN_P:  0.981

Using the verbose -v flag will show you the default config parameters used by Humatch. Users may design their own config file and point to this instead using the --config argument if they wish to specify target genes or add/remove residues Humatch cannot mutate.

If humanising many sequences, Humatch can be run on a csv file of antibody sequences similarly to classification e.g.

Humatch-humanise
    -i data/example.csv
    --vh_col heavy
    --vl_col light

Output (the first example sequence is predicted to be human, so no edits are suggested):

Humatch_H Humatch_L Edit HV LV CNN_H CNN_L CNN_P
QVQ...VSS EIV...IK- 0 hv1 kv3 0.999 1.0 0.998
... ... ... ... ... ... ... ...

Sequence alignment

Users can run sequence alignment in isolation to determine how ANARCI has numbered a sequence. With this information, users can then specify IMGT positions to remain fixed during humanisation e.g. add CNN_fixed_imgt_positions_H: ["9 ", "81A", "120 "] to the config file (note - the spaces where insertion codes are not present are required)

Humatch-align
    -H QVQLVQSGAE...VSS
    -L EIVLTQSPVT...EIK

Output

1       Q       E
2       V       I
3       Q       V
3A      -       -
4       L       L
5       V       T
...     ...     ...
126     V       I
127     S       K
128     S       -

For small speed increases, users may also wish to pre-align many sequences to avoid repeating this step if wanting to trial many different humanisation configs e.g.

Humatch-align
    -i data/example.csv
    --vh_col heavy
    --vl_col light

This can be run with the --imgt_cols flag to return unique columns for each IMGT position, otherwise only two columns are returned - padded VH and VL. If csvs are pre-aligned (without the --imgt_cols flag), the alignment step can be avoided during classification and humanisation by including the --aligned flag.

Citation

@article{Chinery2024,
  title = {Humatch - fast, gene-specific joint humanisation of antibody heavy and light chains},
  author = {Lewis Chinery, Jeliazko R Jeliazkov, and Charlotte M Deane},
  journal = {bioRxiv},
  year = {2024},
  doi = {10.1101/2024.09.16.613210}
}