For D. E. Shaw Research and computational drug-discovery / scientific ML applications:
- Put substantive research, including the Hasan research described in the project brief, in Research Experience, above Projects when it is your strongest evidence.
- In Projects, put Solubility Bench first: it directly shows molecular representations, numerical modeling, scientific Python, controlled experimentation, and honest uncertainty analysis.
- Put DistillTwin next to demonstrate depth in chemical engineering and process modeling.
- Put Titrate next if space remains and it adds distinct substance. If it has stronger real-world use, experimental validation, or technical depth than the other projects, adjust the ordering accordingly.
For process-control or traditional chemical-engineering applications, put DistillTwin first and use Solubility Bench to demonstrate scientific computing breadth. This recommendation uses the descriptions supplied in the conversation; the other repositories were not reviewed.
Solubility Bench — Molecular ML & Uncertainty Quantification
Python · PyTorch · RDKit · scikit-learn
- Built a bond-aware molecular graph network and RDKit descriptor/fingerprint baselines for aqueous solubility prediction across 1,117 unique ESOL structures.
- Benchmarked eight model variants across three seeds and random/scaffold splits; isolated training, model selection, uncertainty calibration and test evaluation, with controlled descriptor, bond-feature and 3D ablations.
- Implemented three-member neural ensembles and split-conformal intervals; evaluated RMSE, interval coverage and width, with reproducible configurations, CLI workflows and automated invariance/leakage tests.
For a one-page resume, use the first bullet and a shortened combination of the other two. The measured-performance option below can replace one methods bullet after reading the result interpretation. Use only claims you can explain, reproduce and defend.
Measured-performance alternative:
- Achieved 0.593 ± 0.039 logS RMSE on three random ESOL splits with a three-member GNN/descriptor ensemble, a 15.9% reduction versus the descriptor random forest; evaluated scaffold generalization at 0.764 ± 0.063 RMSE.
State that these are custom four-way splits. Avoid claiming an official MoleculeNet score, state-of-the-art performance, or guaranteed 90% scaffold coverage.
Why solubility? It is a chemically meaningful property whose observed value reflects competing phase and solvent interactions. It is relevant to early molecular screening, while being a tractable regression task for a small CPU project. It does not establish efficacy or binding affinity.
What makes this chemistry-informed? The representation exposes bonding, chemical neighborhoods and global descriptors related to size, polarity and hydrogen bonding. The optional shape experiment adds invariant conformer geometry. There is no thermodynamic residual in the loss, so “physics-informed neural network” would overstate the implementation.
Why compare a GNN against a random forest? Molecular descriptors already encode strong domain information, and ESOL is small. A more complex representation must earn its additional complexity through controlled comparisons. A baseline outperforming a GNN is a valid scientific finding.
Why four partitions? Reusing early-stopping validation data to calibrate uncertainty would mix model selection with calibration. The separate calibration set isolates that step. The cost is fewer training examples and noisier evaluation.
Why scaffold splits? Random splits can place related chemical series in both training and test. A scaffold split asks a different generalization question. It is still imperfect: acyclic molecules collapse into a large group, side-chain similarity can cross scaffold boundaries, and largest-first assignment has its own distribution biases.
What do the intervals mean? They are prediction intervals for a future measured logS under the relevant assumptions, not confidence intervals for model weights or proof that a new molecule is safe. Exchangeability is crucial. Under scaffold shift, show empirical coverage and width instead of promising nominal coverage.
What does ensemble spread measure? Variation between fitted predictors with different initialization and optimization paths. It does not separately identify aleatoric and epistemic uncertainty. The conformal step calibrates residual intervals; uncertainty-floor and scaled-score choices are explicitly reported.
Why does single-conformer 3D sometimes fail to help? A gas-phase force-field geometry misses solvation, conformational populations and crystal packing. Shape can be noisy or redundant with existing descriptors. Nonconvergence is handled as missingness, not silently accepted.
What would you improve with more time? Start with an external labeled dataset and an overlap audit. Test whether the observed ranking survives different data, rather than repeatedly optimizing against the same small test partitions. For physical modeling depth, add a separate QM9 experiment or molecular simulation task with appropriate symmetries and physically defined targets.
Before interviewing, reproduce the smoke run and one full model, then:
- Trace a SMILES string through canonicalization, featurization, graph batching and a prediction.
- Recompute an RMSE from
predictions.csvand a conformal quantile by hand on a small example. - Explain why turning off bond features retains adjacency and atom chemistry.
- Change a validation-only hyperparameter, save a new run directory and explain the comparison's limits.
- Show an error case and a conformer failure rather than only favorable predictions.
- Explain what you wrote, what you learned, and any tooling assistance used when asked. Do not claim experimental or research experience beyond the work actually done.
- External evaluation with canonical identity/scaffold overlap audits and fixed train-derived calibration.
- More scaffold partitions and group-aware uncertainty on reported benchmark differences.
- Validation-budget comparisons for kernel/fingerprint methods and the graph network.
- Multi-conformer shape aggregation with explicit force-field and sampling limits.
- A separate physically defined QM9 task; 3D-invariant/equivariant architectures with geometric tests and a clearly specified quantum target.