Machine Learning in Biology

One-hot encoding, k-mers, embeddings: how you represent a sequence determines what a model can and can't learn from it. This is the most consequential modelling decision you'll make.

Genomic data is not i.i.d. Sequence similarity between train and test sets inflates performance metrics. This paper is required reading before you report any benchmark number.

The progression from DeepBind to Enformer to modern attention models. Understanding this history explains why the field made each architectural choice, and when older approaches still win.

De-facto standard for bio-ML research. Most modern papers use PyTorch. The official tutorials (60-minute blitz) are the fastest way to get productive.

Parsing FASTA/FASTQ, accessing NCBI, working with sequence records. The plumbing that sits between raw biological data and your ML pipeline.

Before you reach for PyTorch, fit a gradient-boosted tree. If the simple model performs comparably, the complex model isn't justified.

How protein language models are trained, what they encode, and how to use pre-trained embeddings as features. A conceptual foundation before you touch ESM or ProtTrans code.

A worked example of ML applied to a specific bioinformatics problem: correcting sequencing errors in long reads. Shows the full ML workflow from problem formulation to evaluation.

Nikolay Oskolkov · 2022

Scalability challenges in bio-ML, when your data doesn't fit in RAM and your model doesn't converge on a single GPU.

Biological data is high-dimensional, noisy, and structured. Understanding this shapes every modelling decision, especially what you should and shouldn't expect your model to learn.

One-hot, k-mer features, embeddings from language models: the representation is the model's view of the world. This course maps the options and their trade-offs.

The progression from DeepBind and Basenji to attention-based models. Architecture choices have biological consequences, this review traces why the field made each transition.

Genomic splits are not i.i.d. Data leakage ruins papers. How to design evaluations that actually reflect real-world generalisation.

Pre-trained on evolutionary sequences, embeddings that encode function without labels. The starting point for any protein property prediction task.

ColabFold makes AF2 accessible without a GPU cluster. Run a structure prediction before designing any model that takes sequence as input, it changes what you'll try.

Essential for molecular property prediction, protein-protein interaction modelling, and any task where your data is naturally a graph.

At researcher level, the baseline is not optional. It's the falsification check. Your deep learning model should meaningfully outperform a tuned gradient-boosted tree or it's not ready.

How protein language models encode evolutionary and functional information. Watch this before implementing any ESM-based approach.

CNN applied to medical imaging, from raw images to clinical-grade classifiers. A clean end-to-end case study for how ML decisions translate to biological impact.

ML-based integration of multi-omics data to predict gene essentiality. A worked example of how to combine data sources meaningfully.

Read it to understand the architecture and what AF2 actually predicts: the confidence metrics, the training objective, and where the model still fails.

Protein language models at scale. The conceptual leap from NLP to protein function prediction, and the empirical results that made the field take notice.

A critical, balanced review of where deep learning helps in genomics and where it doesn't. Required before starting any bio-ML project, read the 'obstacles' section twice.

ESM, ProtTrans, Ankh: a comparative analysis of what these models actually encode. The benchmark numbers matter less than understanding why a model generalises to a new task.

Equivariance, SE(3)-transformers, and why 3D structure requires a different inductive bias than sequences. The mathematical framing that unifies most modern structure-based models.

Sequence, structure, and function in a single model. The current frontier for protein representation. The API is accessible even without large compute.

Diffusion-based backbone generation followed by sequence design. The practical toolkit for de novo protein design: if you're doing generative work, start here.

Treats docking as a diffusion process over ligand poses. Outperforms AutoDock Vina on blind benchmarks. The new baseline for structure-based drug design.

At specialist level, GNNs are often the right architecture for molecular property prediction. PyG has the widest ecosystem of molecular datasets and pre-built layers.

Multi-omics integration for essentiality prediction. Relevant if you work on target identification, CRISPR screen analysis, or multi-modal data fusion.

The paper that unified sequence, structure, and function in a single generative model. Read for the training objective and the generation experiments, not just the benchmarks.

AF3 extends AlphaFold to protein-DNA, protein-RNA, and protein-ligand complexes. The diffusion-based architecture is a significant departure from AF2.

Treating molecular docking as a diffusion process over the product space of ligand poses. The paper that replaced AutoDock Vina as the benchmark to beat.