About Rewire & Tim Richardson
Independent writing, benchmark evaluations, and reproducible research on machine learning in genomics, proteins, and molecular discovery.
Independent Research & Publication
Rewire (rewirebio.io) is an independent publication and computational project written by Tim Richardson. It examines machine learning applications across sequence biology and molecular design, scrutinizing whether reported capabilities survive held-out test splits, robust baselines, and honest experimental verification.
Rewire is independently maintained. Analyses prioritize primary evidence, publicly inspectable evaluation code, and clear distinctions between statistical proxy scores and demonstrated biological mechanisms.
Core Research Areas
Work is focused across four primary domains:
- Genomics: Sequence foundation models, chromatin accessibility, noncoding regulatory variation, and the limitations of sequence likelihood as a surrogate for functional assay measurements.
- Proteins & Language Models: Residue-level versus pooled representations, zero-shot mutation effect predictions, and distinguishing statistical embeddings from mechanistic explanations.
- Molecular Design: Small-molecule representation learning, mass spectrometry identification baselines, and chemical property prediction.
- Benchmarks & Reproducibility: Rigorous holdout evaluations, leakage prevention, and controlling for confounding variables such as gene annotation releases.
Original Work & Benchmark Suites
Rewire maintains reproducible benchmark records and independent evaluations of widely used biological models:
- MFASS Splice-Variant Benchmark (v1 & v2): An empirical re-analysis of 27,733 variants from minigene reporter assays, evaluating SpliceAI, Pangolin, and supervised baselines across connected exon-and-gene holdout components.
- benchmarks.rewirebio.io ↗: A standalone platform and database for benchmark datasets, runs, and literature records, including the MFASS v2 corrected benchmark run.
- Genomic Foundation Models Evaluation Guide: A two-ledger framework comparing capability claims against held-out validity tests across DNA language models.
- A DNA Likelihood Is Not a Functional Assay: Auditing tokenization contracts, reverse-complement symmetry, and surrogate likelihood limits.
- A Protein Embedding Is Not an Explanation: Selecting, extracting, and testing protein language model representations without mistaking scores for biological explanations.
- A FASTA File Is Not a Specification: Defining complete execution specifications and confidence standards for protein structure prediction.
Public profiles and code
Public source code, profile identity, and project repositories:
- LinkedIn: Tim Richardson on LinkedIn ↗
- GitHub Organization: rewire-bio on GitHub ↗
- Benchmark Runner Repository: rewire-bio/rewire-benchmarks on GitHub ↗
- Benchmarks Platform: https://benchmarks.rewirebio.io ↗
- Direct Contact: [email protected]