InsilicoCell

A screen that wants one compound to kill a tumour cell, bind its intended target, and push a disease transcriptome in the corrective direction has had to run three unrelated models to say so. Each was fit to its own task and confined to the compounds and cell lines in its own training table, so a molecule scored by one was frequently invisible to the next and the predictions could not be combined into a single ranking.

InsilicoCell collapses that into one pretrained model. Every entity is first embedded in its own modality — chemical fingerprints for compounds, a node2vec walk over the Gene Ontology graph for genes, a ProtTrans encoder for protein sequences, an autoencoder over transcriptomes for cellular context — and tokenization networks project all of them into a shared 768-dimensional token space. A prediction is then a short sequence of those tokens read by shared transformer encoder blocks, with a task head on the pooled CLS output. Coverage lives in the representation rather than in a lookup of entities seen during training, which lets the same weights score a compound, gene, protein or cell line the model never encountered.

It comes from the Bin Chen lab at Michigan State University, with collaborators at the University of Michigan, the University of Hawaii, Henry Ford Health and Corewell Health.

#Key Features

• Seven profiles from one checkpoint: drug sensitivity, drug-induced expression change, drug–protein binding affinity, TF–gene association, gene effect score, mutation status and copy number variation all run from the same released weights, with no refitting step in the prediction workflow.
• Entity-level holdout evaluation: benchmarks withhold whole compounds, genes, proteins and cell lines from training rather than sampling rows out of a shared cohort, so the reported generalization is new-entity transfer.
• Multi-objective virtual screening: candidates are ranked simultaneously on target binding, cell-line sensitivity and predicted transcriptional response, joining target-based and phenotypic screening that are normally separate stages.
• Multi-task training measurably helps: against the identical architecture trained one task at a time, joint training improved drug-sensitivity RMSE by 24.8% and raised prediction-versus-truth correlation in HeLa from 0.68 to 0.85.
• Extension past cell lines: patient, single-cell and spatial settings are reached by zero-shot inference or fine-tuning, with active sampling to pick informative samples when labelled data is scarce.
• Conversational access: a hosted Model Context Protocol server exposes over 20 APIs, so screens can be requested in ordinary language rather than through the Python scripts.

#Technical Details

InsilicoCell is a supervised multi-task transformer of just over 46 million parameters: six encoder blocks with 16 attention heads over 768-dimensional tokens, plus seven regression and classification heads. Pretraining used over 88 million labelled measurements from LINCS, BindingDB, ChEA, CTRP and DepMap, covering 816,230 compounds, 18,545 genes, 6,550 proteins and TFs, 1,204 cell lines, five treatment durations and six doses, and ran in two stages — joint multi-task training with dynamic loss weighting, then task-specific continuation from the stage-1 weights — for 100 epochs on three NVIDIA H100 GPUs over three days. Pretraining is on bulk, cell-line-scale measurements; single-cell, spatial and patient data are transfer settings, not part of that corpus. On drug–protein binding the model improved RMSE by 13.7% over the averaged performance of DTIAM, DeepConv, ConPlex and EnzPred on an external dataset, and on drug sensitivity in cell lines unseen by both models it improved correlation over TransCell by 139.7%. Substituting the transformer for a size-matched MLP or random forest degraded all seven tasks, and swapping the autoencoder cell representation for scGPT, Geneformer or scVI embeddings did not improve it.

#Applications

The intended use is virtual screening at library scale against several objectives at once. The authors screened Enamine libraries of 1.8 and 4.7 million compounds across three campaigns. All five hepatocellular carcinoma candidates reduced c-Myc protein in HepG2 cells, two of them inhibiting growth with IC50 values of 21.6 nM and 282 nM. In idiopathic pulmonary fibrosis, Z8413727732 suppressed COL1A1, α-SMA and FN1 in TGF-β-induced myofibroblasts with an IC50 of 3.96 μM while sparing normal fibroblasts, and a screen in mesenchymal stem cells raised OCT4, SOX2 and NANOG expression without cytotoxicity.

#Impact

InsilicoCell argues that compounds, genes, proteins and cells are better carried in one shared representation than in seven task-specific tables, and that the reward is optimizing several pharmacological objectives inside a single ranking. The wet-lab campaigns give that claim more weight than in-silico benchmarks alone, though they are early-stage hit-finding rather than efficacy results, and a provisional patent covers the disclosed compounds. The work remains a preprint awaiting peer review, and while code, weights and the zero-shot benchmark test sets are released under Apache 2.0, the training-set splits and fine-tuning code are stated as pending publication — so the model is reproducible at inference but not yet retrainable.