Immunogenic cell death (ICD) is a form of tumor cell death that, instead of passing quietly, turns the dying cell into an in-situ vaccine: the cell broadcasts chemical danger signals — calreticulin on its surface, ATP and HMGB1 released outward — that recruit and activate dendritic cells and cytotoxic T cells. Finding drugs that trigger this mode of death, rather than the immunologically silent default of apoptosis, is one of the main routes to making "cold" tumors responsive to checkpoint immunotherapy. But the assays that detect ICD are destructive and slow: you stain or lyse the cells at a fixed endpoint, typically 24 or 48 hours after adding the compound, and you get one number per well. Nobody decides when to look; everyone waits out the full clock, or images exhaustively, which caps how many compounds a microscope can screen.
Meanwhile, machine translation solved a structurally identical problem years ago. A simultaneous translator must emit words while the speaker is still talking, so its policy continuously chooses between READ (wait for more input, gaining accuracy at the cost of delay), WRITE (commit an output now), and PREDICT (hallucinate the missing input and commit early anyway). This proposal transplants that decision structure into high-content live-cell screening. The core claim: screening compounds for ICD induction should be run as a simultaneous process in which a learned policy watches a stream of time-lapse microscopy frames and, at every frame, decides whether to acquire the next frame, to commit a prediction of the compound's eventual ICD outcome, or to roll out a predicted future of the cell's morphology and DAMP kinetics before committing. The trade-off is real in both domains — and in biology it is doubled, because the biology itself has a latency-quality curve: the poroptosis work showed that acutely lysed cells are poorly immunogenic, while cells dying slowly through ~4–8 nm membrane nanopores stay metabolically alive long enough to secrete sustained inflammatory signals and drive T-cell-dependent tumor rejection Poroptosis: A form of cell death depending on plasma membrane nanopores formation. Fast death and slow observation both cost you; the question is where to stop.
Core Claim
No existing system treats ICD screening as a sequential, time-budgeted decision problem. The closest works are static on both sides of the fence: PhenoSeq translates a single fixed-timepoint Cell Painting image into a predicted single-cell transcriptomic embedding via conditional diffusion, but never models time or makes any decision Cell Painting Generates Single-Cell Transcriptomics via Conditional Diffusion; BFReg-NN screens drug pairs for pyroptosis against a fixed nine-gene omics signature, with no notion of an observation budget AI-powered omics-based drug pair discovery for pyroptosis therapy targeting triple-negative breast cancer. On the policy side, the READ/WRITE/PREDICT machinery — static read-write schedules Incremental Decoding and Training Methods for Simultaneous Translation in Neural Machine Translation, adaptive segmentation into "meaningful units" Learning Adaptive Segmentation Policy for Simultaneous Translation, and RL agents with a PREDICT action that hallucinates missing source tokens Prediction Improves Simultaneous Neural Machine Translation — has only ever been applied to linguistic streams. We claim that (i) a streaming time-lapse of a perturbed cell population is a well-posed instance of the same formal problem, (ii) a PREDICT action implemented as a temporal rollout model — PhenoSeq's diffusion idea extended from one image to a predicted future trajectory of morphology and DAMP proxies — lets the policy commit at budgets where current-morphology confidence is still uninformative, and (iii) prefix-matched training, the analogue of chunk training that restored up to 12 BLEU points in incremental translation Incremental Decoding and Training Methods for Simultaneous Translation in Neural Machine Translation, is necessary to close the gap between early commitments and the full-timeline ceiling.
Research Program
We will build and evaluate a system, provisionally called SimulICD, with four components. First, a task formalization: each well is a stream of frames acquired at a fixed cadence after compound addition, with ground-truth labels taken from the destructive endpoint assays the field already runs — surface calreticulin, extracellular ATP and HMGB1, LDH leakage kinetics, and the cytokine panel (TNF-α, IL-6, IFN-β) used in the poroptosis study Poroptosis: A form of cell death depending on plasma membrane nanopores formation. A compound is ICD-positive if the endpoint cascade crosses validated thresholds, ICD-negative otherwise — including the crucial negative class of acute lysis, which releases DAMPs fast but is immunologically inferior, exactly the trap a fixed-early-timepoint classifier would fall into. Second, a streaming encoder: a per-frame vision transformer over a temporal encoder that consumes incremental frame prefixes, pretrained on the Cell Painting Gallery's large endpoint collections (JUMP's ~136,000 perturbations and cpg0004-lincs' 1,571 compounds across doses) Cell Painting Gallery: an open resource for image-based profiling, then fine-tuned on time-lapse data. Third, the PREDICT module: a conditional diffusion forecaster that, given the observed prefix, rolls out predicted future morphology and a 512-dimensional transcriptomic-state proxy in the spirit of PhenoSeq Cell Painting Generates Single-Cell Transcriptomics via Conditional Diffusion — but generating the next hours of the trajectory, not a static molecular snapshot. Fourth, the policy itself, trained to optimize a latency-regularized objective:
$$J(\pi) \;=\; \mathbb{E}\big[\text{BalAcc}(\hat{y}, y)\big] \;-\; \lambda\,\mathbb{E}\big[T_{\text{obs}}/T_{\text{total}}\big]$$
where $\pi$ maps each observed frame prefix to an action in $\{\text{READ}, \text{WRITE}, \text{PREDICT}\}$, $\hat{y}$ is the committed ICD prediction, and $T_{\text{obs}}/T_{\text{total}}$ is the fraction of the full (say, 48-hour, 96-frame) timeline acquired before commitment — a direct analogue of Average Proportion in simultaneous translation. Following the computation-aware latency critiques from speech translation CA*: Addressing Evaluation Pitfalls in Computation-Aware Latency for Simultaneous Speech Translation, Going Beyond Your Expectations in Latency Metrics for Simultaneous Speech Translation, we will also report wall-clock decision lag including model inference time, not just the algorithmic frame count. Training proceeds in two stages: supervised prefix-matched training (each training trajectory is randomly truncated so the model sees the distribution of prefixes it will actually face, the chunk-training analogue), then policy-gradient refinement of the READ/WRITE/PREDICT choice against $J(\pi)$, mirroring the RL agent design of the PREDICT-action paper Prediction Improves Simultaneous Neural Machine Translation. One design constraint comes from a genuine biological contradiction in the literature: RIPK1 activation drives ICD via necroptosis and CD80 upregulation in one study Targeted Induction of Cancer Cell Necroptosis Potentiates Anti-PD-1 Immunotherapy via CD80 Activation, while the anticonvulsant phensuximide selectively suppresses RIPK1-dependent ICD FDA-approved phensuximide inhibits RIPK1-dependent immunogenic cell death. Mechanism annotations are therefore unreliable labels; the policy must predict measured DAMP outcomes from observed trajectories, not from pathway tags.
Baselines
The naive reference point is the field's current default: image every well at the full fixed endpoint (48 hours), classify from the complete observation, and never stop early. This sets the quality ceiling and the throughput floor — it is the "do nothing" against which any early commitment must justify itself, since it uses the most information available. The strongest schedule-based competitor is the wait-$k$ analogue: acquire exactly $k$ frames and then commit, sweeping $k$ to trace a quality-latency curve, exactly as fixed read-write schedules do in simultaneous translation Incremental Decoding and Training Methods for Simultaneous Translation in Neural Machine Translation, Efficient Wait-k Models for Simultaneous Machine Translation. The strongest existing learned method is PhenoSeq-style per-frame classification Cell Painting Generates Single-Cell Transcriptomics via Conditional Diffusion: the same streaming frame encoder, applied at a fixed frame budget, with the diffusion-translated transcriptomic embedding as an auxiliary feature. To make this comparison honest we will use the same pretrained frame encoder, the same fine-tuning data, and the same tuning budget across all methods; the reported numbers to beat on the scGeneScope-style setting are PhenoSeq's single-profile balanced accuracy of 0.293 for synthetic transcriptomic embeddings versus 0.270 for raw imaging features, and its multi-profile test at $k = 32$ cells where accuracy more than doubled from 0.157 (imaging-only) to 0.271 (synthetic fusion). Because the honest threat is not PhenoSeq itself but a cheap adaptive rule on top of any classifier, we add a fourth baseline: confidence-thresholded early exit, where the same streaming encoder commits as soon as its predictive confidence crosses a threshold, with the threshold swept to trace its own curve. Finally, two ablations isolate our key ingredients: the full policy with the PREDICT action removed (READ/WRITE only, so the forecaster's contribution is separated from the policy's), and the full policy with the forecaster's rollouts replaced by replaying the last observed frame (so the gain comes from genuine anticipation, not from extra parameters). SimulICD counts as winning only if its quality-latency curve dominates the wait-$k$ and confidence-threshold curves at matched observation budgets and matched compute, and closes a stated fraction of the gap to the full-timeline ceiling at budgets below half the timeline.
Why not just…?
Why not just train one big classifier on the full time-lapse movies? Because full movies per well are precisely what screening cannot afford — the entire motivation is microscope and incubator bandwidth — and because a classifier trained on full trajectories degrades badly on the truncated prefixes an early-committing system actually faces. The translation literature quantifies this mismatch directly: chunk training that matched the incremental evaluation structure restored up to 12 BLEU points over sentence-trained baselines Incremental Decoding and Training Methods for Simultaneous Translation in Neural Machine Translation. We expect the same effect here, and our prefix-matched-training ablation tests it directly.
Why not just use a confidence threshold for early stopping? This is the most serious alternative, and we treat it as a first-class baseline rather than arguing around it. Our hypothesis for why it falls short: early ICD-relevant morphology is genuinely ambiguous — a cell showing membrane permeabilization at hour 6 could be on a subacute, highly immunogenic poroptosis-like trajectory or heading for immunologically poor acute lysis Poroptosis: A form of cell death depending on plasma membrane nanopores formation — so softmax confidence is miscalibrated exactly where early commitment matters. A policy with a PREDICT action can act on anticipated future rather than current certainty, which is what gave the translation PREDICT agent its simultaneous quality and latency gains (English→German BLEU from 16.75 to 17.54 while Average Proportion dropped from 0.79 to 0.74) Prediction Improves Simultaneous Neural Machine Translation. If the threshold baseline matches the full policy across the budget range, that is a finding we will report: it would mean the forecaster and RL machinery are not earning their complexity, and the useful contribution reduces to prefix-matched training plus calibrated early exit.
Why not just run the static omics screen and skip imaging altogether? BFReg-NN predicts synergistic pyroptosis-inducing drug pairs from complete multi-omic profiles with a reported C-index of 0.90 on the FUSCC triple-negative breast cancer cohort AI-powered omics-based drug pair discovery for pyroptosis therapy targeting triple-negative breast cancer, and it answers a complementary question: which compounds to prioritize. It cannot answer ours — how much live observation any candidate requires before its outcome is predictable — because it consumes bulk and single-cell sequencing data that are themselves destructive, expensive, and endpoint-fixed. The two are complements: BFReg-NN nominates candidates; SimulICD decides how to triage them cheaply, and a validation experiment will feed BFReg-NN-nominated pairs (e.g., the mitoxantrone–gambogic acid pair) into the streaming pipeline as high-priority streams.
Proposed Experiments and Decisive Checks
The evaluation proceeds in three stages. Stage one builds the missing benchmark, ICD-Stream: time-lapse imaging (six-channel Cell Painting-compatible staining adapted for live longitudinal acquisition) of the cell lines and perturbations already validated in the file — ppM1 for poroptosis, doxorubicin as in the nanovesicle work Genetically Engineered Biomimetic Nanovesicles Co-Deliveing a Checkpoint Inhibitor and Doxorubicin for Enhanced Cancer Chemo-Immunotherapy, the mitoxantrone–gambogic acid pair AI-powered omics-based drug pair discovery for pyroptosis therapy targeting triple-negative breast cancer, CL-387785 Targeted Induction of Cancer Cell Necroptosis Potentiates Anti-PD-1 Immunotherapy via CD80 Activation, plus freeze-thaw and detergent lysis as acute-death negatives — across MC38, 4T1, CT26, B16F10, A375, and A549 lines, with endpoint DAMP assays as labels. The frame encoder is pretrained on the Cell Painting Gallery Cell Painting Gallery: an open resource for image-based profiling. Stage two evaluates all baselines and ablations on quality (balanced accuracy and AUC for predicting endpoint ICD labels) against latency (fraction of timeline observed; wall-clock decision lag including inference time), reporting full quality-latency curves rather than single points. Stage three converts curves into the screening-relevant currency: compounds correctly triaged per microscope-day, and, for the top early-called inducers, an in vivo check in the 4T1 and B16F10 syngeneic models measuring CD8$^+$ infiltration, benchmarked against the reported rise in Granzyme-B-positive CD8$^+$ T cells from roughly 3% to 18.5% under combination necroptosis induction Targeted Induction of Cancer Cell Necroptosis Potentiates Anti-PD-1 Immunotherapy via CD80 Activation. Three checks are decisive. The policy must dominate the wait-$k$ and confidence-threshold curves at matched budgets, with the margin concentrated at low budgets where PREDICT should matter most, mirroring the translation result that anticipation improved both quality and latency simultaneously Prediction Improves Simultaneous Neural Machine Translation. Prefix-matched training must recover a measurable share of the gap between early-commit and full-timeline accuracy, the analogue of the 12-BLEU chunk-training restoration Incremental Decoding and Training Methods for Simultaneous Translation in Neural Machine Translation. And the policy must specifically separate subacute immunogenic death from acute lysis at timepoints where endpoint assays cannot — the biological latency-quality trade-off that motivated the whole undertaking Poroptosis: A form of cell death depending on plasma membrane nanopores formation. No reported result found on Lacuna for any existing method on this streaming ICD task, which is itself the gap this proposal fills.
References
- Poroptosis: A form of cell death depending on plasma membrane nanopores formation
- Cell Painting Generates Single-Cell Transcriptomics via Conditional Diffusion
- AI-powered omics-based drug pair discovery for pyroptosis therapy targeting triple-negative breast cancer
- Incremental Decoding and Training Methods for Simultaneous Translation in Neural Machine Translation
- Learning Adaptive Segmentation Policy for Simultaneous Translation
- Prediction Improves Simultaneous Neural Machine Translation
- Cell Painting Gallery: an open resource for image-based profiling
- CA*: Addressing Evaluation Pitfalls in Computation-Aware Latency for Simultaneous Speech Translation
- Going Beyond Your Expectations in Latency Metrics for Simultaneous Speech Translation
- Targeted Induction of Cancer Cell Necroptosis Potentiates Anti-PD-1 Immunotherapy via CD80 Activation
- FDA-approved phensuximide inhibits RIPK1-dependent immunogenic cell death
- Efficient Wait-k Models for Simultaneous Machine Translation
- Genetically Engineered Biomimetic Nanovesicles Co-Deliveing a Checkpoint Inhibitor and Doxorubicin for Enhanced Cancer Chemo-Immunotherapy