Lab Newsletter — September 21, 2026: Will the Drug Work?

We’ve spent the week deep in molecules — structures, immune receptors, antibodies. Today we zoom out to a bedside question that all of it ultimately serves: will this drug work for this patient? Two people with the “same” cancer can respond completely differently, because their tumors differ underneath — in mutations, gene expression, copy number. Precision oncology’s dream is to read that molecular profile and predict the response before treating. Today’s digest is about the deep-learning models learning to do exactly that — a concrete, high-stakes instance of the virtual cell idea: given a cell’s state and a perturbation, predict the outcome.

🗺️ The foundation: a map from genotype to sensitivity

You can’t learn drug response without data linking molecular profiles to outcomes at scale. Iorio et al. (Cell, 2016) built that map with the Genomics of Drug Sensitivity in Cancer (GDSC) resource. They took “cancer-driven alterations identified in 11,289 tumors from 29 tissues (integrating somatic mutations, copy number alterations, DNA methylation, and gene expression)” and mapped them “onto 1,001 molecularly annotated human cancer cell lines and correlated with sensitivity to 265 drugs.” Crucially, they found that “cell lines faithfully recapitulate oncogenic alterations identified in tumors” — the assumption that makes the whole enterprise possible. This landscape, and its siblings, became the training ground for everything that followed.

🧮 The first move: predict effectiveness from a genomic signature

With a landscape in hand, could a network learn the genotype→response mapping directly? Chang et al. (Scientific Reports, 2018) showed it could with CDRscan, “a novel deep learning model that predicts anticancer drug responsiveness based on a large-scale drug screening assay data encompassing genomic profiles of 787 human cancer cell lines and structural profiles of 244 drugs.” Its “two-step convolution architecture” learned to read a tumor’s mutation profile as a signature of vulnerability. The premise is precision medicine in one line: “cancer therapy can be tailored to an individual patient based on the genomic profile of a tumour.”

🌉 The honest problem: from cell lines to real tumors

Cell lines are not patients, and pretending otherwise is where these models break. Chiu et al. (BMC Medical Genomics, 2019) confronted that head-on with DeepDR, predicting “drug response of tumors from integrated genomic profiles by deep neural networks.” They name the crux plainly: “due to essential differences between cell lines and tumors, to date the translation into predicting drug response in tumors remains challenging.” Their answer was to pretrain on the abundant cell-line and tumor genomic data, then transfer — an early recognition that in this field the generalization gap, not raw accuracy on cell lines, is the real target.

🧬 More omics, integrated: expression, mutation, copy number

A tumor is more than its mutations. Sharifi-Noghabi et al. (Bioinformatics, 2019) asked how to combine data types with MOLI, “multi-omics late integration with deep neural networks for drug response prediction.” Their starting point is a well-earned fact — “gene expression has been shown to be the most informative data for drug response prediction” — but “recent evidence suggests that integrating additional omics can improve the prediction accuracy, which raises the question of how to integrate.” MOLI learns separate representations for each omic and fuses them late, reflecting the lab’s own conviction that multi-omics integration done well beats any single view of the cell.

💊 Bring in the drug: molecules as graphs

Predicting response for a new drug means the model must understand chemistry, not just biology. Liu et al. (Bioinformatics, 2020) closed that loop with DeepCDR, “a hybrid graph convolutional network for predicting cancer drug response.” It “integrates multi-omics profiles of cancer cells and explores intrinsic chemical structures of drugs,” representing each compound as a molecular graph. Modeling both sides — the cell’s state and the drug’s structure — is what lets a predictor generalize across the vast grid of (cell, drug) pairs, and it connects this work to the same molecular-graph representations behind drug design and synthesis planning.

🔍 Make it interpretable: a visible model of the cell

Accuracy alone won’t reach the clinic — oncologists need to know why. Kuenzi et al. (Cancer Cell, 2020) built for that with DrugCell, motivated by a sobering fact: “most drugs entering clinical trials fail, often related to an incomplete understanding of the mechanisms governing drug response,” and most ML models “have not reached clinical practice due to their lack of interpretability and their focus on monotherapies.” DrugCell is “an interpretable deep learning model of human cancer cells trained on the responses of 1,235 tumor cell lines to 684 drugs,” in which “tumor genotypes induce states in cellular subsystems that are integrated with drug structure to predict response.” Its network is structured to mirror real cell biology, so a prediction comes with a mechanistic trace — and it predicts drug synergy, pointing toward rational combinations.

🧫 Why it’s our kind of problem

Read across the six and it’s the virtual cell problem in miniature: encode a cell’s molecular state, apply a perturbation, predict the phenotype. DrugCell’s “visible” network — built to match cellular subsystems rather than as a black box — is precisely the mechanistic-plus-learned hybrid the lab believes cell modeling needs, and its interpretability is the same value the lab prizes across imaging and omics. Two lab themes recur loudly. First, multi-omics integration (MOLI, DeepCDR) — the conviction that no single readout captures a cell — is core to the lab’s single-cell work. Second, the field’s real bottleneck is generalization and data (the cell-line→patient gap DeepDR names), which is exactly why open datasets, honest benchmarks, and shared infrastructure — the BioEngine ethos — matter more than any single architecture. Getting from “we designed a molecule” to “it will help this patient” is the whole game; these models are early, imperfect, but pointed straight at it.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (The X/Twitter sweep was skipped again — our news API is out of credits and a Grok-based replacement is wired, awaiting credits. Anchors were verified via NCBI E-utilities.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.