GSOC’25 Project: Improving New Drug Modality Support

[Sonalilipsa17375]

Hello DeepChem Community ,

I’m Sonali Lipsa Patra, a 3rd yr. UG student in the EE department at NIT Rourkela, Odisha, India. I am excited to share that I will be working on the project Improving New Drug Modality Support for Antibody Drug Conjugates (ADCs) this summer as part of GSoC’25.

Throughout the project, I will be sharing weekly updates on my progress here. I am looking forward to engaging with the community and welcome feedback and suggestions from all.

Looking forward to learning, contributing, and having great discussions. Stay tuned for more updates.

Regards,
Sonali Lipsa Patra

[NTHamda]

Congratulations, wishing you a rewarding and impactful GSoC’25 journey ahead. Looking forward to seeing your project!

[Sonalilipsa17375]

Week-1 Updates :

• Explored the ADC Payload Bioactivity Dataset from ChEMBL to analyze the distribution of IC50 values for DNA topoisomerase I, a key target for ADC payloads. Link to Dataset
• Studied the DeepChem documentation on adding new datasets to MoleculeNet. Link
• Created a draft PR covering the introduction to ADCs and the overview of basic sciences. Draft PR
• Addressed previous review comments on ADC-related literature in the tutorial and will continue upgrading the section in phases.

Link to Slides Deck

[Sonalilipsa17375]

Week-2 Updates :

• Continued exploring the ChEMBL dataset for DNA topoisomerase I inhibitors, key for ADC payload identification. [Link to ChemBL dataset]
• Studied this amazing paper on ADC payload activity prediction. [Link]
• Uploaded the ChEMBL dataset to a personal AWS account, and preparing for Molecule Net integration.
• Visualized IC50 distribution with a log-scale histogram to assess compound potency. [Link to Notebook].
• Physicochemical Property Analysis: Analyzed IC50 value distribution. Planned visualizations for Molecular Weight and AlogP to study conjugation and delivery properties.
• Refined the ADC tutorial draft PR, addressing feedback on literature and incorporating script insights. [Link to Draft PR].

Link to Slides deck : https://docs.google.com/presentation/d/1GRsTgPoTBwrouery58OWFU7mrRUsvGaLdysTS_lebsA/edit?usp=sharing

[Sonalilipsa17375]

Week-3 Updates:

• Added the Approved ADCs dataset into the assets, (as the approved ADCs data is low, 22 approved ADCs, so added under assets).

• Created both 2D and 3D renderings of an example ADC (Brentuximab vedotin) payload to illustrate structural features.

• Standardized IC₅₀ values and generated a binary active/inactive label based on potency thresholds.

• Implemented code to visualize key payload properties - IC₅₀ (nM), and AlogP against the DNA Topoisomerase‑1 .

• [ Link to PR ] - https://github.com/deepchem/deepchem/pull/4441

• [Link to slides deck] :
  https://docs.google.com/presentation/d/1GRsTgPoTBwrouery58OWFU7mrRUsvGaLdysTS_lebsA/edit?usp=sharing

[Sonalilipsa17375]

WEEK 4 - updates -

• The Antibody–Drug Conjugates tutorial (#4441) is now nearly complete and ready for viewing.

• This week, I explored molecular dynamics for the first time by successfully setting up and running a LAMMPS simulation of a single water molecule, focusing on its structural conformations.

• I’m currently reviewing the ADCNet paper and working on implement in the second ADCs tutorial. Gone through the ADCNet paper, which proposes a deep learning framework for predicting ADC activity. It integrates ESM-2 for protein sequences and FG-BERT for small-molecule representations, using inputs like antigen/antibody sequences, SMILES strings, and drug-antibody ratio (DAR) to achieve accurate predictions. You can find the paper here: [link to paper: https://arxiv.org/pdf/2401.09176 ].

• I worked on incorporating feedback discussed during the OH session and will soon begin working on Tutorial 2, which focuses on the computational aspects of ADCs. This will include prototyping the ADCNet within the tutorial framework.

• [ Link to PR ] - https://github.com/deepchem/deepchem/pull/4441

• [Link to slides deck] :
  https://docs.google.com/presentation/d/1GRsTgPoTBwrouery58OWFU7mrRUsvGaLdysTS_lebsA/edit?usp=sharing

[Sonalilipsa17375]

Week-5 Updates-

• Implemented ADCNet prototype, using ESM-2 to generate embedding from protein sequences (antibody, and antigen) , and chemberta to generate embeddings from smiles string of payload and linker molecules.

1. Notebook 1-https://colab.research.google.com/drive/1lLpzgsIHNSDqU7_H2D1jk3lQFMsq9Ro7#scrollTo=FP3FvHthd_eS
2. Notebook- 2-https://colab.research.google.com/drive/1uPhG4HrpPx1rywcCIJKh5lrrTJfqqfgV#scrollTo=iCFGRzCzml9r

• Fixed review works on starter tutorial of ADC and is ready for final reviews. [PR Link - https://github.com/deepchem/deepchem/pull/4441] #4441
• Studied ADCNet architecture and went through the codebase. I runned ADCNet in local for better understanding. [Link- https://arxiv.org/abs/2401.09176 ]
• [Link to slides deck] :
  https://docs.google.com/presentation/d/1GRsTgPoTBwrouery58OWFU7mrRUsvGaLdysTS_lebsA/edit?usp=sharing

[Sonalilipsa17375]

Week-6 Updates-

• Implemented ADCNet tutorial: Added a new tutorial on Antibody-Drug Conjugates (ADCs) showcasing the ADCNet model for biological drug modeling using DeepChem.[PR: https://github.com/deepchem/deepchem/pull/4478]

• Merged starter tutorial on introduction to ADCs, and can be found at- https://github.com/deepchem/deepchem/blob/master/examples/tutorials/Introduction_to_Antibody_Drug_Conjugates.ipynb

• Worked on Code formatting CI fix, which was failing for unspecified types which mypy was unable to understand, in some files. [PR link: https://github.com/deepchem/deepchem/pull/4480]

• [Link to slides deck] :
  https://docs.google.com/presentation/d/1GRsTgPoTBwrouery58OWFU7mrRUsvGaLdysTS_lebsA/edit?usp=sharing

[Sonalilipsa17375]

Week -7 Updates:

1. Implemented ADCNet tutorial in PyTorch; currently working on adapting it to DeepChem style.
   PR: #4478
2. Studied DeepChem splitters, dataset, and Torch wrappers prior to implementation.
3. Minor fix made to the starter ADC tutorial; updated Colab link.
   PR: #4486

Slides Deck: https://docs.google.com/presentation/d/1GRsTgPoTBwrouery58OWFU7mrRUsvGaLdysTS_lebsA/edit?usp=sharing

[Sonalilipsa17375]

Week-8

• To make the tutorial deepchem specific, replaced the chemberta used earlier from hf to chemberta from deepchem.
• Utilized Roberta featurizer for tokenization replacing auto tokenizer, used before.
• Faced errors while utilizing deepchem’s chemberta had discussions with my mentors on the same.
• Tutorial is ready for review and can be accessed at:
• Link : https://github.com/deepchem/deepchem/pull/4478
  Slides Deck: https://docs.google.com/presentation/d/1GRsTgPoTBwrouery58OWFU7mrRUsvGaLdysTS_lebsA/edit?usp=sharing

[Sonalilipsa17375]

Week 9 Updates:

• My tutorial on ADCNet had received a LGTM before, but during rebasing, I encountered some recurring merge conflicts. To address this, I created a fresh pull request (PR) with all the previous work copied into the new PR, including the rebasing. Its ready to be merged.
  Here is the link: https://github.com/deepchem/deepchem/pull/4501.

• I have been busy throughout the week with academics and working on my final report for my GSoC project with my mentor, Ankita.
  Link: https://docs. google.com/document/d/1DuZ7EhPIadMwFKqvOy_h6Ojk9a5FShQJD10BpPytDms/edit?usp=sharing

Slides Deck https://docs.google.com/presentation/d/1GRsTgPoTBwrouery58OWFU7mrRUsvGaLdysTS_lebsA/edit?usp=sharing

As we approach the end of the summer and the timeline, I feel grateful to be a part of DeepChem and immensely thankful to the community for all their support and guidance throughout the course of this project.

I feel overwhelmed to put on my final report here:

GSOC FINAL REPORT