Day 1: Baseline Architecture & Dataset Prep

10 Sep 2026

• Initialized scaffold-disjoint split on BindingDB 2024 benchmark dataset.
• Extracting raw sequences for target proteins and calculating ESM2 (650M parameter model) residue-level embeddings.
• Goal: Ensure zero overlap between training compound scaffolds and test compound scaffolds.

import torch
from transformers import AutoTokenizer, EsmModel

tokenizer = AutoTokenizer.from_pretrained("facebook/esm2_t33_650M_UR50D")
model = EsmModel.from_pretrained("facebook/esm2_t33_650M_UR50D").cuda()

# Generate mean-pooled sequence representation
def extract_embedding(seq):
    inputs = tokenizer(seq, return_tensors="pt").to("cuda")
    with torch.no_grad():
        outputs = model(**inputs)
    return outputs.last_hidden_state.mean(dim=1)

Observation: Initial preprocessing shows 14.2% sequence similarity drop between train and test sets when strict scaffold grouping is enforced.

Day 3: Initial Benchmarking & Leakage Checks

12 Sep 2026

• Benchmarking 1D CNN baseline against 3D-Graph Transformer.
• Metric tracked: Concordance Index (CI) and Root Mean Squared Error (RMSE) on test set.

Preliminary Results

• Math verification for loss penalty weighting: $\mathcal{L} = \mathcal{L}_{\text{MSE}} + \lambda \|\mathbf{W}_h\|_2^2$

Day 6: Checkpoint Reached! 🎯

15 Sep 2026

• Reached target performance threshold: RMSE $< 0.55$ under zero-shot target scaffold splits.
• All evaluation scripts verified with zero data leakage.
• Checkpoint Action: Compiled findings, loss curves, and architectural diagrams into a full research article post!