what we're learning, building, and thinking about. no marketing fluff — just updates on the science and the platform.
the log
23 articles · 7 topics
latest article · Protein Science
A very-high-confidence predicted model is usually within about half an ångström of the crystal structure, and about one residue in ten is still off by more than 2 Å. How to read pLDDT and PAE against resolution, B-factors and cryo-EM local resolution, and how to decide, task by task, whether a predicted model is good enough.
· 20 min read
The best open-source tools for computational drug discovery in 2026, by pipeline stage: protein structure prediction, docking, ADMET, generative chemistry, protein design and engineering, retrosynthesis and single-cell analysis — each with its licence, its primary paper and its latest release.
Scanpy and Seurat run the same canonical scRNA-seq pipeline and still disagree, because almost every step in it has a default and the two packages chose different ones. A practical comparison of the data structures, the pipelines, the places results genuinely diverge, and how to choose.
DiffDock finds where a ligand binds on a whole protein; GNINA and AutoDock Vina are strongest when you already know the pocket. Here is how the published benchmarks compare, why pose accuracy is not physical validity, and why no docking score is a binding affinity.
Rule Set 2, Rule Set 3 and DeepSpCas9 are three on-target sgRNA models trained on three different measurements. Here is what each score means, how it relates to measured indel frequency and screen depletion, and why an on-target score is not an off-target claim.
An in silico ADMET panel ranks compounds; in vitro assays measure them and in vivo studies measure the animal. Here is what each layer actually resolves, where predictions fail, and how to decide when a computational flag justifies an experiment.
A computational developability screen is a triage instrument, not a substitute for the panel. Sequence-motif liability flags have published grading against forced degradation; composite developability scores correlate with the measured assays at |rho| of about 0.3 or less. Here is what each half of the comparison actually supports.
A DeepPurpose drug–target interaction score is a sequence-only estimate of the label its checkpoint was trained on: pKd on a kinase panel, pIC50 from BindingDB, or a KIBA score. A binding assay measures the interaction itself. What each number is, how far apart they can be, and how to use the score without mistaking it for a Kd.
ProteinMPNN recovers 52.4% of native residues against Rosetta's 32.9%. That figure is a benchmark on solved backbones, not a success rate — here is what it does predict, what it cannot, and what to measure instead.
Docking places a ligand into a receptor structure you already have. Co-folding predicts the whole complex from sequence. Here is what DiffDock, Chai-1 and Boltz-2 each need as input, what their confidence scores mean, what Boltz-2's affinity head is and is not, and when to reach for each.
GenMol samples molecules as fragments with masked discrete diffusion; REINVENT 4 steers an RNN or transformer prior with reinforcement learning and curriculum scoring. Here is what each one is built for, what GuacaMol and MOSES do and do not measure, and how to run them.
ThermoMPNN, FoldX and an ESM zero-shot scan answer three different questions about a point mutation, and only two of them are in kcal/mol. A benchmark-anchored comparison of what each method predicts, where each one fails, and what the wet-lab comparator actually measures.
ADMET-AI, SwissADME and pkCSM are all free, and they are not the same kind of instrument. One returns learned probabilities and regression values at library scale, one returns rules and physicochemical readouts, one returns regression and classification from graph-based signatures. Here is how to pick.
P2Rank scores points on a protein's surface with a random forest; fpocket clusters alpha spheres found by Voronoi tessellation. P2Rank ranks the true site higher, fpocket finds more sites somewhere in a longer list, and the best published configuration is fpocket's pockets re-ranked by P2Rank's own scorer. How the two methods work, what the benchmarks actually measured, and which to reach for.
RDKit's MolLogP is the Wildman–Crippen atom-contribution model, and TPSA is a fragment sum that stands in for 3D polar surface area. Both are fast, deterministic and good enough to rank and filter compounds. Neither is a measurement. A practical comparison of calculated logP, TPSA and the rule-of-five descriptors against shake-flask logP and measured aqueous solubility, with the published error bars.
A predicted paratope and an experimentally mapped epitope are not two measurements of the same thing: one names antibody residues, the other names antigen residues, and each experiment defines a binding residue differently. What Paragraph actually predicts, how alanine scanning, HDX-MS and crystallography each define the interface, and how to compare them without fooling yourself.
Plain BLAST settles most homology questions, and it stops working in the twilight zone below about 30% identity. A practical comparison of pairwise BLAST, iterated PSI-BLAST and profile-HMM search with HMMER: what each detects, what each gets wrong, and how to choose.
Pocket predictors such as P2Rank work on AlphaFold and ESMFold models, but a predicted pocket is only as trustworthy as the residues lining it. Here is what changes between predicted and experimental structures, which P2Rank model to use, how to read pocket scores, and how the fold → pocket → dock pipeline runs on .bloom*.
Germline identity, a gene-specific CNN classifier and a 9-mer repertoire search all answer "how human is this antibody?" — about different objects, on different scales, with different blind spots. What each score measures, where they disagree, and what none of them predicts.
Learn how to predict absorption, distribution, metabolism, excretion, and toxicity properties of drug candidates using AI models — before spending resources on synthesis.
A comprehensive guide to the best open-source tools for computational drug discovery in 2025 — covering protein folding, docking, ADMET, molecular generation, retrosynthesis, and genomics.
A practical comparison of ESMFold and AlphaFold2 for drug discovery — covering accuracy, speed, when to use each, and how to run both on the Ivy Biosciences platform.
Exploring how machine learning is revolutionizing the pharmaceutical industry and accelerating therapeutic development from years to months.
newsletter
