Predicting how a cell's transcriptome shifts when a gene is silenced — a solo entry to the Virtual Cell Challenge 2026, built in the open.
Created by Saber Hafezqorani
What it is. Sidechain is a multi-agentic research workflow that tries to decode one of biology's hardest problems – understanding how a cell's biology changes when a gene is silenced. The idea is to climb a ladder of models from the simplest baseline up and keep only what the metric pays for.
How it's shared. Code, configs and tests land here, and I'll share the journey with the broader scPerturbSeq research community through LinkedIn and blog posts. Let's catch up along the way.
A side chain is the part of an amino acid that makes it different from the other canonical nineteen.
Our models are named the same way: each series is an amino acid whose side chain matches
the model's character, and a number counts entries within it — GLY (no side chain: nulls and
baselines), ALA (a single statistical shift), SER (transfers a group: cross-line delta
transfer), CYS (bridges chains: context-aware models), HIS/LYS/ARG (long-range charged:
graph and prior heads), PHE/TYR/TRP (the aromatic heavyweights: deep generative models),
PRO (bends the backbone: fusion). A letter suffix marks a one-knob variant — SER-1p is
SER-1 with Poisson cells. Every model that has reached the board, with what each name means
and what it borrows, is on the models page. For
the chemistry behind the pun, Compound Interest's
20 common amino acids poster.
Every submission, as Arc scored it — the ones that worked and the ones that did not. The
live board shows only a team's latest entry
and re-ranks as others submit, so the rank shown below is the one each entry held when it was
scored — its first appearance in our board snapshots. An entry marked calibration is one
sent purely to test a new method and its performance on the official board
(what that means). The table is
auto-generated by scripts/standings.py after each submission.
| date (UTC) | submission | overall | rank when scored |
|---|---|---|---|
| 2026-08-21 | sidechain r1-delta-even v1 (SER-1) |
0.0788 | #2 of 42 |
| 2026-08-21 | Sidechain SER-1p |
0.0730 | #10 of 95 |
| 2026-08-22 | Sidechain SER-1n |
0.0822 | #9 of 97 |
| 2026-08-24 | Sidechain SER-2 |
0.0937 | #25 of 216 |
| 2026-08-27 | Sidechain SER-3n |
0.0963 | #54 of 301 |
| 2026-08-27 | Sidechain SER-3fn |
0.1034 | #58 of 362 |
| 2026-08-30 | Sidechain SER-3afgn |
0.0992 | #101 of 476 |
| 2026-08-30 | Sidechain SER-3afn |
0.1071 | #86 of 476 |
| 2026-08-31 | Sidechain SER-4afn |
0.1078 | #85 of 481 |
| 2026-09-01 | Sidechain SER-5acfnt |
0.1075 | #114 of 533 |
| 2026-09-07 | Sidechain PHE-2 · calibration |
-0.9807 | #746 |
| 2026-09-12 | Sidechain SER-6aefn |
0.1091 | #245 of 903 |
| 2026-09-17 | Sidechain SER-7abefn |
0.1131 | #314 of 1022 |
| 2026-09-29 | Sidechain SER-8abefkn · calibration |
0.1132 | #486 of 1225 |
| 2026-09-29 | Sidechain SER-9abefkn |
0.1144 | #478 of 1224 |
| 2026-09-30 | Sidechain SER-10abefnw |
0.1353 | #403 of 1236 |
| 2026-09-30 | Sidechain SER-11abefknw |
0.1360 | #415 of 1255 |
| 2026-10-02 | Sidechain SER-12aefkw |
0.1438 | #405 of 1292 |
| 2026-10-02 | Sidechain SER-13aefknw |
0.1360 | #443 of 1292 |
| 2026-10-03 | Sidechain SER-14aefksw |
0.1626 | #283 of 1307 |
| 2026-10-07 | Sidechain SER-15aefkrsw · calibration |
-0.0039 | #917 of 1377 |
| 2026-10-08 | Sidechain SER-16aefhkrsw · calibration |
0.1423 | #471 of 1403 |
| 2026-10-08 | Sidechain SER-17aefhkrsw |
0.1657 | #344 of 1408 |
| 2026-10-16 | Sidechain SER-18aefksvw · calibration |
0.1612 | #453 of 1571 |
| 2026-10-16 | Sidechain SER-19aefhkrsw |
0.1684 | #426 of 1573 |
| 2026-10-19 | Sidechain SER-20aefhkrsw · calibration |
0.1673 | #430 of 1589 |
| 2026-10-19 | Sidechain SER-21aefhkrswx |
0.1757 | #382 of 1595 |
| 2026-10-20 | Sidechain SER-27abefhkrswx |
0.1875 | #339 of 1594 |
| 2026-10-20 | Sidechain SER-28abefhkrswx |
0.1875 | #340 of 1597 |
| 2026-10-21 | Sidechain SER-29abefhkrswx |
0.2226 | #221 of 1638 |
What each model is, and what its name means: saberhq.com/sidechain/models.