GSoC 2025 - Continuous Interaction Fingerprints

[VAKozyrev]

Hello everyone,

My name is Vladimir. I'm a PhD student in computational chemistry. I'm interested in the Continuous (i.e., non-binary) interaction fingerprints (IFPs) project and wanted to leave some of my thoughts here.

Also, hello to the mentors @cbouy and @talagayev. My apologies for submitting the proposal on the final day; I appreciate your understanding and hope it still receives full consideration.

I think my main questions are whether you want to focus on fine-tuning the parameters for the thresholds used to define "ideal" and "minimum" distances, or just on implementing the possibility of providing these parameters? Also, I see how it should be more or less straightforward to implement distance cutoffs, while angles and combinations of angle-distance cutoffs might be more challenging.

In my pre-proposal, I focused more on the technical side of the question—implementing the functionality—rather than on setting the exact values for the thresholds.

Firstly, I would start by implementing two new parameters for the Interaction class and its subclasses like Hydrophobic, HBAcceptor, etc. They should accept both minimal_distance and ideal_distance parameters for their initialization. These are thresholds corresponding to 1 and 0 in the fingerprint.

Then, these parameters must be provided to the Fingerprint class as part of the parameters dictionary during its initialization. This is then used in Fingerprint._set_interactions to set the list of considered interactions and their parameters.

Next, if the ideal and minimum distances are provided and a continuous fingerprint should be generated, the current detect methods:
https://github.com/chemosim-lab/ProLIF/blob/c2daf5738ae528c067d95ed401fe8c6c83d74a98/prolif/interactions/base.py#L140-L147
should be modified to output metadata of the interaction (including the real distance between interacting groups) if the distance is below minimum_distance:

if dist <= self.minimum_distance:
    yield self.metadata(...)

This way, while constructing a fingerprint, we will retrieve the interaction metadata if the distance is below minimum_distance. The construction of the fingerprint happens in the Fingerprint.bitvector method:
https://github.com/chemosim-lab/ProLIF/blob/c2daf5738ae528c067d95ed401fe8c6c83d74a98/prolif/fingerprint.py#L282-L287

Here, each interaction from the list of interactions considered in the fingerprint is called for two residues, which calls the Interaction.__call__ method:
https://github.com/chemosim-lab/ProLIF/blob/c2daf5738ae528c067d95ed401fe8c6c83d74a98/prolif/interactions/base.py#L50-L52
This method returns either True or int_metadata, which is the interaction metadata, containing the real distance between interacting groups. In our case, we need to make sure that the metadata flag is set to True.

Returning to the Fingerprint.bitvector method, we can modify it by adding something like:

bitvector = [
    interaction(res1, res2) for interaction in self.interactions.values()
]
if self.continuous:
    return np.array(map(activation_function, bitvector), dtype=np.float32)

Here, activation_function is a function that needs to be implemented to map the interaction metadata from bitvector to a continuous value between 0 and 1. The activation_function will take the interaction metadata and output 1 if the distance is below ideal_distance, or apply a sigmoid function if it is between ideal_distance and minimum_distance.

As a starting point, the sigmoid function can be defined as:

def sigmoid(x, m, k): 
    return 1 / (1 + np.exp(-k * (x - m)))

Here, m can be set to (ideal_distance + minimum_distance) / 2 so that the function returns 0.5 halfway between the two cutoffs. The parameter k controls the steepness of the curve and can be fine-tuned.

I'd really appreciate any feedback, and everyone is very welcome to join the discussion!

[cbouy]

Hi @VAKozyrev,

Sorry for the (very) late reply here, this discussion completely fell off my radar 😅 You should have received some feedback in the pre-proposal acceptance email, hopefully it clarified some things!
Feel free to submit your proposal on the GSoC portal with a link to a Google docs in the PDF so that any mentor can provide feedback (hopefully in a timely manner)!